WO2025022019A1 - Method for the predictive diagnosis of a pathological condition or a pathological state - Google Patents

Abstract

The present application relates to an in vitro method for the predictive diagnosis of a pathological condition from at least one biological sample taken from a subject and comprising microorganisms, the method comprising the identification and the relative abundance of said microorganisms present in the sample, the diagnosis being carried out using a pre-trained artificial intelligence model on the basis of a training set wherein the labelled data set comprises training subject profiles, each training subject profile comprising the identity and the relative abundance of all the microorganisms identified in at least one sample from said training subject without any preselection, wherein each training subject profile is labelled with the phenotype of the training subject from which it is derived.

Description

Description

Predictive diagnostic process for a pathology or pathological condition

[1]The present application relates to a method, in particular in vitro, for diagnosing or predictively diagnosing a pathology or a pathological condition from a biological sample taken from a subject. According to a particular aspect, the invention relates to a method, in particular in vitro, for predictively diagnosing a pathology of the digestive system or an extra-digestive pathology of a subject from the analysis of the microbiota present in a biological sample taken from the digestive system, and/or outside the digestive system such as the vagina and/or in the stools of a subject. Even more particularly, the invention relates to a method for diagnosing necrotizing ulcerative enterocolitis (NUE) in premature newborns from a biological sample taken from their stools. According to another particular aspect, the subject of the present invention is the predictive diagnosis of premature delivery from a biological sample taken from the vagina of a pregnant woman. The present method is therefore in the field of diagnosis, predictive diagnosis, particularly in vitro, and personalized medicine.

[2]Preterm birth is a major cause of morbidity and mortality in newborns. A proportion of spontaneous preterm births appear to result from an inflammatory reaction following an infection of the genital tract. However, a large proportion of preterm births remain without an identified cause, without clinical signs. Despite various studies relating to the vaginal microbiota and the occurrence of preterm birth, there is currently a need for a reliable clinical method for predicting the occurrence of preterm birth. Unfortunately, clinicians currently have no reliable tool to predict the risk of preterm birth.

[3] Patent EP 3161167 describes a method for assessing a risk of premature delivery based on the detection, in a vaginal or cervical sample obtained by swabbing from a pregnant woman, of the quantity of the following bacteria: Vimonas micra, Ureaplasma urealyticum or Ureaplasma parvum, Atopobium vaginae, Peptoniphilus lacrimalis, Megasphaera cerevisiae and Parvibacter caecicola, compared to a reference level. The quantification of the bacteria is carried out by amplification of a small region of ribosomal DNA (16S rDNA) by quantitative polymerase chain reaction (qPCR).

[4] Patent EP 2 972 308 B9 describes a serum or plasma peptide biomarker, produced by human cells, and not by the microbiota, the detection of which is used in a method for assessing a risk of premature birth. [5]International application WO 2020/227053 describes a method for determining the risk of premature birth comprising determining the abundance of Saccharibacteria TM7-H1 and optionally BVAB1, Sneathia amnii and Prevotella in a vaginal sample from a pregnant woman, from the nucleotide sequence of a small portion of the 16S rDNA of the microorganisms.

[6]These examples illustrate the possibility of a relationship between the nature of the microbiota and the physiological or pathological state of a subject. But it is also known that the complexity of microbiotas makes it difficult to determine specific and predictive microbial signatures characteristic of a pathological state. This situation is made all the more complex by the very strong interindividual variations. To date, several techniques for analyzing the microbiota exist. However, current approaches do not allow for precise characterization of microbiotas.

[7]Genes expressing the small subunit of ribosomal RNA (rRNA), i.e. genes called 16S ribosomal DNA “16S rDNA” for prokaryotic microorganisms, such as bacteria and archaea, and “18S rDNA” for eukaryotes, including yeasts, are used to enable the description of the structure of the microbiota (Chakoory et al., 2022).

[8]The publications of Park et al in 2021 and 2022 describe a method for predicting the probability of preterm birth from the detection of a restricted number of microorganisms present in the vaginal microbiome.

[9]In the 2021 publication by Park et al., the method involves the simultaneous quantification by qPCR amplification of specific small fragments of DNA from each of the following 10 microorganisms: Lactobacillus crispatus, Lactobacillus iners, Weissella koreensis, Bacteroides fragilis, Prevotella bivia, Prevotella amnii, Prevotella salivae, Ureaplasma urealyticum, Ureaplasma parvum, Gardnerella vaginalis.

[10]In the 2022 publication by Park et al, based on a sequencing approach of a small V3-V4 region of the 16S rDNA gene and literature studies, predictions of the probability of preterm birth are established on the basis of 10 bacteria (Lactobacillus crispatus, Lactobacillus fornicalis, Lactobacillus gasseri, Lactobacillus iners, Lactobacillus jensenii, Gardnerella vaginalis, Ureaplasma parvum, Atopobium vaginae, Prevotella timonensis and Peptoniphilus grossensis) as well as 7 additional bacteria based on previous work by other authors (Bifidobacterium breve, Dialister proprionicifaciens, Lactobacillus paracasei, Mobiluncus curtisii, Prevotella disiens, Staphylococcus aureus, Streptococcus anginosus). But the joint exploitation of these data did not allow a documented clinical use according to the state of the art. [11]Therefore, there is a need to develop a method to predict the probability of preterm birth more reliably, taking into account inter-individual variability and poorly represented species. Monitoring pregnant women would make it possible to identify women at risk and anticipate the care of newborns.

[12] Furthermore, necrotizing enterocolitis (NEC) is the most common life-threatening gastrointestinal emergency encountered by preterm infants in neonatal intensive care units. It is defined as ulcerative inflammation of the intestinal wall. Current clinical practice for diagnosing NEC is based on clinical, radiological and haematological findings constituting the Bell criteria, according to a recent review (D'Angelo et al., 2018). The clinical signs of early NEC are often very subtle and may initially manifest as feeding intolerance and nonspecific symptoms (malaise, bradycardia) before gastrointestinal symptoms become evident. These include increased gastric residuals, bloody stools and abdominal distension; These may progress to generalized hypotonia, lethargy, and cardiorespiratory failure, which may also be present in other neonatal conditions, including sepsis and viral intestinal infections. If the disease is not diagnosed and treated early, it can lead to severe sepsis, intestinal perforation, and significant morbidity (gastrointestinal necrosis, chronic intestinal failure) and mortality (up to 40% for severe forms).

[13]To date, clinicians have no reliable diagnostic tool for predicting NEC. The pathophysiology of NEC remains poorly understood and effective methods for its early detection have yet to be established. Therefore, current efforts to understand and predict NEC focus on the study of its risk factors. Preterm birth represents the most important risk factor for the development of NEC. In very low birth weight neonates (<1.5 kg at birth), the incidence of NEC ranges from 5% to 13%. In addition, prolonged administration of antibiotics during the first week of life and substitution of breast milk with formula or infant formula are frequently associated with the later onset of NEC.

[14] Colonization of the gut microbiota has been widely considered to play a role in the development of NEC in preterm infants, but as in the case of the probability of preterm birth assessed from the vaginal microbiome, the complexity of microbiotas makes it difficult to determine specific microbial signatures that are predictive of a physiological or pathological state, not allowing the identification of a single opportunistic pathogen or microbial community. pathogen as a cause of ECUN. This failure is mainly due to the early and very dynamic establishment of the neonatal intestinal microbiota, influenced by many factors, including environment, sex, gestational age, mode of delivery, feeding mode and antibiotic treatments.

[15]Therefore, there is also a very important need for a reliable and reproducible method for predictive diagnosis of pathologies affecting newborns, particularly premature newborns. Such a predictive diagnosis would make it possible to identify newborns at risk and anticipate the management of pathologies likely to seriously affect their lives.

[16]To give a third example of the possibility of a relationship between the nature of the microbiota and the physiological or pathological state of a subject, type I diabetes (T1D) is an autoimmune disease that results from the destruction of the beta cells of the pancreas by the patient's lymphocytes. This destruction results in the patient's inability to secrete insulin, which leads to the inability to use glucose as an energy resource, thus to hyperglycemia at the same time as an intracellular energy deficiency. Excess sugar in the blood is found in the urine.

[17] T1D affects children and young adults. In the short term, it is responsible for a significant deterioration in quality of life since affected subjects must constantly adapt their insulin intake (subcutaneously) to blood sugar, food intake and energy expenditure. In the medium and long term, chronic hyperglycemia leads to multi-organ alterations, particularly nervous and vascular.

[18]The incidence of T1D has been increasing continuously since at least 1988. In France, it is 18 per 100,000 in those under 15 years of age, over the period 2013-2015, i.e. a prevalence of around 1.3 per 1,000. The incidence of diabetes in young people is increasing by 3 to 4% per year, at the same time as the age of onset is decreasing (Gale E 2002).

[19] Immune activation is multifactorial and depends partly on the HLA system and postnatal infectious events. There is thus a familial aggregation of cases, an association with other autoimmune diseases, and a possible link with certain viral agents, notably group B coxsackieviruses.

[20]After destruction of pancreatic beta cells, the only treatment is lifelong insulin replacement therapy. To date, the only curative treatment is allogeneic beta cell transplantation, which is a complicated treatment, requiring prolonged immunosuppression, with average results. [21] The diagnosis of type I diabetes is based on the demonstration of hyperglycemia, glycosuria, and activation of the immune system directed against beta cells, as evidenced by the presence of anti-GAD, anti-Zn T8, and anti-insulin antibodies. This immune activation precedes the disease by several months, and a new strategy is emerging which consists of detecting children at high risk of developing type I diabetes in the siblings of a child already affected, and offering them immunomodulatory treatment. This identification of high-risk children is currently based exclusively on the presence or absence of autoantibodies. However, not all children who have autoantibodies develop type I diabetes.

[22] Indeed, the activation of an immune response depends on a balance between populations that activate and inhibit the immune reaction, this balance being likely to be largely influenced by exogenous agents, particularly viral and bacterial. In this context, the hypothesis that digestive dysbiosis can lead to immune activation is a promising avenue. An international cohort of children at risk of type I diabetes (Vatanen T, Nature. 2018 Oct;562(7728):589-594) made it possible to study the digestive microbiota of these children in comparison with that of children who had not developed the pathology, without however being able to identify, with the methods used, microbial taxa characteristic of one or the other of the situations (pathological and healthy). Therefore, there is also a very important need for a reliable and reproducible method for predictive diagnosis of type I diabetes based on the analysis of the microbiota taken from the stools of children at risk. Such a predictive diagnosis would make it possible to identify children likely to develop the disease and anticipate the management of this chronic pathology affecting quality of life and which can lead to serious after-effects or even death without appropriate management.

[23]The possibility of early identification of children at high risk of developing autoimmunity and then diabetes would allow a therapeutic revolution towards personalized preventive medicine for this extremely disabling disease. Indeed, recent immunomodulatory preventive treatments currently available have proven their effectiveness in preventing diabetic disease in children at very high risk. However, these treatments are not without adverse effects, and must be used in a targeted manner.

[24]To give a fourth example of the possibility of a relationship between the nature of the microbiota and the physiological or pathological state of a subject, neonatal sepsis is a disease due to the presence in the blood of an infectious agent, most often of a bacterial nature. This situation is potentially very serious due to two threats: hemodynamic failure due to the disseminated inflammatory reaction (septic shock), and bacterial dissemination in vital sites, particularly the meninges (purulent meningitis). It therefore requires urgent diagnosis and treatment, which is based on the administration of intravenous antibiotics. These initially target the germs most frequently involved (probabilistic antibiotic therapy); once the bacteria have been identified, the antibiotic therapy is adapted in order to limit as much as possible the selection of strains resistant to antibiotics.

[25]Neonatal sepsis affects approximately 1 in 1000 full-term newborns. In the setting of normal pregnancy and birth, prevention is based on maternal history and detection of vaginal carriage of streptococcus B. If carriage occurs, antibiotic therapy is given to the mother during labor, so that the newborn is protected even if streptococcus is transmitted during birth.

[26]On the other hand, in cases of prematurity, neonatal sepsis is much more frequent, affecting more than one in four children. This increased frequency is due to the fragility of premature children, the presence of invasive equipment (catheters, probes) and prolonged hospitalization (hospital germs, multiple daily manipulations by many caregivers). The germs responsible for sepsis are most often found in the digestive tract of children, and sometimes on the skin, particularly in the case of an indwelling catheter.

[27]The diagnosis of sepsis is currently based on the combination of non-specific symptoms (fever, malaise, tachycardia, vomiting, etc.), blood markers of the inflammatory response (neutrophilic polynucleosis, elevated CRP) and sometimes the detection of bacteria in the blood (by blood culture). This last test must be carried out before any antibiotic therapy (which would mask the result), and requires a considerable blood volume (at least 1 ml, or 2% of the total blood volume of a 500 gram premature baby). The identification of a germ generally takes 1 to 2 days, and the characterization of its sensitivity to antibiotics can take up to a week.

[28]The adaptation of the treatment is therefore late, exposing the newborn to an unnecessarily broad spectrum antibiotic therapy (with the consequence of an imbalance of the digestive microbiota and the selection of resistant strains).

[29]Therefore, there is also a very important need for a reliable and reproducible method for predictive diagnosis of sepsis based on the analysis of the microbiota collected from the stools of children at risk. Such a predictive diagnosis would make it possible to identify children at risk and anticipate the management of this chronic pathology that can seriously affect their lives.

[30]Prediction of neonatal sepsis would allow increased monitoring of at-risk newborns, and would allow earlier treatment in case of symptoms. In addition, characterization has a priori of the germs probably responsible, carried in particular in the digestive tract of the newborn, would allow to prescribe from the outset a treatment more adapted to the profile of these bacteria.

[31]Several methods of analyzing the microbiota for the purpose of diagnosis or predictive diagnosis of a pathology are known from the prior art.

[32]A first method called "metabarcoding" makes it possible to determine the taxa present in a sample thanks to their genetic signature, unique for each taxa. The idea is to have a DNA fragment present in all the taxa to be analyzed and which constitutes a genetic marker. This marker is a DNA fragment framed by highly conserved regions and therefore the most "universal" possible, and which, once sequenced, shows variations in genetic sequences between different taxa. In the context of the microbiota, this method often includes the amplification of fragments of a size between 300 and 470 base pairs of the V3 and/or V4 regions of the gene expressing the 16S rRNA. However, this method has several limitations: biases are likely to be generated during the amplification step carried out by PCR and can alter the vision of the real diversity of the microbiota. Indeed, it is known that the primers used that cannot be "universal" to amplify nucleotide sequences will favor the amplification of the sequences of certain microorganisms to the detriment of others, resulting in a possibly erroneous abundance of microorganisms or even the non-detection of certain microorganisms. In addition, the short length of the sequenced DNA fragments provides only a low taxonomic resolution, not allowing the description of microbial communities at the species level.

[33]Another method comprising a direct metagenomic sequencing step (in English "shotgun") followed by an assembly step to generate complete genomes (in English: Metagenome Assembled Genomes or MAGs) and a step of affiliation of the MAGs leads to an identification restricted to the dominant species.

[34]Another method involves a direct metagenomic sequencing step followed by affiliation of unassembled raw reads smaller than 300 base pairs from a portion of the 16S rRNA-expressing gene. Affiliation of these small sequences leads to low resolution of microbial identification and overestimation of diversity, particularly through detection of false positives.

[35] There is therefore a need to obtain more precise, reliable, reproducible and relatively quick to implement diagnostics and predictive diagnostics, so that they can be used by clinicians in their decision-making. Description of the invention

[36]The inventors have succeeded in developing a unique method for addressing the various issues mentioned above. This method advantageously comprises the use of all the microorganisms identified in the microbiota of a subject by an artificial intelligence model to establish a diagnosis or a predictive diagnosis of a pathology or pathological condition.

[37]The present invention thus has as its first subject a method, in particular in vitro, for diagnosis or predictive diagnosis of a pathology or pathological condition in a subject, from at least one biological sample taken from the subject and containing microorganisms, said method comprising the following steps: a) sequencing, from the nucleic acid isolated from the subject's sample, the nucleotide sequences corresponding to at least one sequence of interest selected from the group consisting of: a fragment of a gene expressing 16S ribosomal RNA (rRNA), a fragment of a gene expressing 18S rRNA, a fragment of 16S rRNA, a fragment of 18S rRNA, b) from the sequencing of step a), determination of the identity and relative abundance of the microorganisms present in said sample without any preselection, c) determination of the predictive diagnosis of said pathology or pathological condition by a model artificial intelligence model from at least the abundances of the identities obtained in step b), said artificial intelligence model having previously been trained on the basis of a labeled data set, where the labeled data set comprises profiles of training subjects, each training subject profile comprising the identity and relative abundance of all the microorganisms identified in at least one sample of said training subject, where each training subject profile is labeled with the phenotype of the training subject from which it originates, and where only the abundances of the identities of the microorganisms that were not present in the labeled data set are excluded from the data from step b).

[38]The phenotype label assigned to each training subject depends on the purpose of the method according to the invention and the type of data used for training. The labeled data set includes at least two different states for the phenotypes and in particular antinomic states: a positive phenotype associated with a diagnosis/diagnosis positive predictive and a negative phenotype associated with a diagnosis/negative predictive diagnosis. Thus, for a diagnosis, the training subject phenotype can be classified as "not affected" or "affected" by the pathology or pathological condition or "healthy" and "sick", these types of classification being synonymous. For a predictive diagnosis, the training subject phenotype can be classified as "having developed" or "not having developed" the pathology or pathological condition or "with appearance" or "without appearance" of the pathology or pathological condition, these types of classification being synonymous.

[39]The invention has the advantage of training the artificial intelligence model more efficiently by using the identity of all the microorganisms identified in the labeled data set. The absence of a step of pre-selection of the identity of microorganisms in the labeled data set for training the artificial intelligence model makes it possible to preserve all the diversity and individual variability of the microbiotas and all the associated microbial interactions in the context of a specific pathology or pathological state.

[40]In addition, the method according to the invention has the advantage of restricting to a minimum (or even of not applying any restriction) the exclusion of the identities of the microorganisms from the data from step b) transmitted to the artificial intelligence model during step c), making it possible to preserve as much as possible the microbial diversity present in the subject's sample. Indeed, the selection of the identities sent to the artificial intelligence model is in no way done on the basis of a relative abundance that is too low in the subject's sample or their absence of known involvement in the pathology or pathological condition, but only on the basis of their presence in the training data set. Thus, if the data set is sufficiently large and exhaustive, no identity of microorganisms is excluded from the data transmitted to the artificial intelligence model to carry out step c).

[41] It was not clear that using the identity of all microorganisms without prior selection during training could give relevant results. This is even contrary to what was expected. Indeed, it is traditionally considered that complex high-dimensional data, used as input to an artificial intelligence model, can contain noise and irrelevant information that can harm learning and therefore the performance of the model (Botteghi, N., Guo, M. & Brune, C. Deep kernel learning of dynamical models from high-dimensional noisy data. Sci Rep 12, 21530 (2022)). The search for microbial signatures for the diagnosis and predictive diagnosis of pathologies and pathological conditions is particularly complex due to the very strong inter-individual variations in the microbiota. The microbiota of each individual is indeed influenced by many factors including the latter's lifestyle, diet and environment. This is also why, although to date, several Although microbiota analysis techniques exist, they do not allow for a precise characterization between microbiotas and pathologies, the risk of developing said pathologies, or the evolution of the latter. Thus, the most likely result would have been to obtain a large number of false positive or false negative diagnoses.

[42]This is why, while the state of the art showed that the complexity of microbiotas made it difficult to determine specific and predictive microbial signatures characteristic of a pathological state or pathology, a situation made all the more complex by the very strong inter-individual variations, all together these complex aspects transmitted to the artificial intelligence model previously trained according to the invention made it possible, against all expectations, to obtain predictive and diagnostic results of great precision, reliability, reproducibility and relatively quick to implement. The method of the invention thus meets a previously unmet clinical need and provides simple and quality information to a clinician.

[43]This is therefore a major advance that allows us to reveal links between these communities of microorganisms and pathologies and pathological conditions, whether these are already present in the subject, whether they are evolving or whether they develop or occur a posteriori. The establishment of predictive diagnoses advantageously allows us to anticipate the subject's care, or even to carry out preventive treatments.

[44]The method of the invention takes into account as the identity of each microorganism the classification by taxonomic rank, this rank preferably being the species of the microorganism. No preselection is carried out during the identification, in particular on the basis of their relative abundance and/or their known involvement in the diagnosis or predictive diagnosis.

[45]According to one embodiment, the microorganisms of the labeled data set as well as those of step b) are identified at the level of the same taxonomic rank. This rank is notably chosen from the phylum to the species, and is preferably the species.

[46]Alternatively, when training the artificial intelligence model and during step b), the identity of each microorganism corresponds to the most confident taxonomic rank, which can be a species, a genus, a family, an order, a class or a phylum. Thus in this case, whether for the labeled dataset or the identification of step b), the identities of the microorganisms will not all have the same rank. This aspect advantageously allows to preserve the maximum exhaustiveness of the labeled dataset when training the intelligence model. In the case where it is not possible to assign a species to a nucleotide sequence or to a set of sequences, it/they will be assigned the most confident taxonomic level, which can be a genus, a family, an order, a class or phylum, (and potentially followed by the term “unclassified”), as well as its/their abundance.

[47] “Most confident taxonomic rank” means the most precise taxonomic rank obtainable from the nucleotide sequence or set of nucleotide sequences used to identify a microorganism. Obtaining the most confident rank depends on several factors, described in detail below.

[48]The diversity of microbiotas given to the artificial intelligence model during its training can be ensured by using data from training subjects of multinational origins, in particular multi-continental, in particular from all continents. Thus, the training subjects are divided into different groups of geographical origin. In particular, the distribution of subjects in the different groups is as representative as possible of geographical diversity.

[49]According to one embodiment of the invention, the labeled data set comprises at least one determined clinical data item, where each training subject profile comprises a value for the or each determined clinical data item, and where step c) comprises providing the artificial intelligence model with the corresponding value of the subject for the or each determined clinical data item.

[50]According to one embodiment of the invention, the method according to the invention thus has the advantage, from a simple sample of vaginal microbiota during pregnancy, in the ^1st trimester and/or in the ^2nd trimester and/or ^3rd trimester, and its sequencing, of predicting with high certainty the occurrence of a premature birth or a full-term birth.

[51]In particular, the method of the invention allows the predictive diagnosis of the occurrence of premature birth, the accuracy of which can reach 88%. Such a degree of reliability is unmatched among the methods for diagnosing premature birth to date.

[52]According to another embodiment, the method according to the invention also has the advantage, from a simple sample of microbiota in the stools of a subject, and its sequencing, of determining with high certainty the development of a disease of the digestive system or an extra-digestive disease. This approach can advantageously be used in the context of personalized medicine to evaluate the relevance of more precise clinical monitoring and/or the use of therapeutic treatment.

[53]Thus, the method of the invention allows a reliable prediction of ulcerative necrotizing enterocolitis with an accuracy of up to 94.9%. Such a degree of reliability is very useful for identifying premature newborns at risk, strengthening monitoring and allowing rapid therapeutic responses avoiding possible serious health problems. To this end, the method of the invention allows early and very effective diagnosis of ECUN and equally effective distinction of unaffected infants.

[54]According to one embodiment of the invention, the method is intended for the predictive diagnosis of type I diabetes in a child. The method according to the invention, in a similar manner, also makes it possible to reliably predict the occurrence of type I diabetes (T1D), with an accuracy of up to 73.6% in particular. The method of the invention thus makes it possible to identify early on children at high risk of developing autoimmunity and then diabetes, which would allow a therapeutic revolution towards personalized preventive medicine to avoid the disabling consequences of the pathology.

[55]According to one embodiment of the invention, the method aims at a predictive diagnosis of neonatal sepsis in an infant. The method according to the invention also makes it possible to reliably predict the occurrence of sepsis, with an accuracy of up to 92.3%. The method of the invention thus makes it possible to identify premature newborns at risk, to strengthen monitoring and to adapt treatment to the profile of these bacteria involved in the pathology.

[56]The invention also relates to a method for training an artificial intelligence model intended to obtain a diagnosis or a predictive diagnosis, said method using a labeled data set comprising profiles of training subjects, where each training subject profile comprises the identity and relative abundance of all the microorganisms identified in at least one sample of said training subject without any preselection, and where each profile is labeled with the phenotype of the training subject from which it originates.

[57]The characteristics described above and below in relation to the labeled dataset and in general to the artificial intelligence model and its training apply mutatis mutandis to the present object.

[58]The training method according to the invention makes it possible to obtain a more reliable and more precise artificial intelligence model in these predictions, for the aforementioned reasons.

[59]This training method has in particular made it possible to identify microorganisms that would be key players in various pathologies, pathological states and absences of the latter. Microorganisms can thus be identified as being able to play the role of probiotics or for the development of new treatments, or even new diagnostics and predictive diagnostics. In this context, thanks to the method according to the invention, the inventors were able to observe the association of several species of microorganisms with the presence of a given pathology, on the one hand, and observe the association of several species of microorganisms with the absence of a given pathology, on the other hand.

[60]Notably, the inventors found that several Lactobacillus species were associated with non-ECUN cases, while several other bacterial species such as: unclassified Enterobacter, unclassified Enterobacteriaceae, Enterococcus faecalis, unclassified Klebsiella, Haemophilus parainfluenzae, Enterococcus durans, and Enterobacter cancerogenus were associated with ECUN cases. These findings suggest that the subject's diagnosis is a function of both dominant, subdominant, and even rare taxa, emphasizing that no individual species or taxonomic group of species is exclusively responsible for an increased risk of ECUN. Instead, without being bound by any theory, the inventors suggest it is likely that various microbial consortia may cause inflammatory cascades leading to the onset of ECUN.

[61]The data obtained using the training process therefore also make it possible to have a precise mapping of the microorganisms associated with the presence of a state which could lead to a pathology or a pathological state, and of the microorganisms associated with the absence of a state leading to a pathology or a pathological state, on the other hand.

[62]According to a particular aspect, the method according to the invention also has the advantage of not increasing the number of obstetric examinations on pregnant women carried out during pregnancy, since the vaginal sample can be recovered during an examination already scheduled.

[63]According to another particular aspect, the method according to the invention advantageously makes it possible to carry out early therapeutic interventions in order to prevent the development or the worst complications of an extra-digestive pathology based on the analysis of the intestinal or fecal microbiome of a subject.

[64]The present invention also relates to a computer program product comprising executable instructions, which when executed on a computer allow the implementation of step c) of determining the diagnosis/predictive diagnosis of the method according to the invention. The characteristics previously and subsequently described in relation to the artificial intelligence model apply mutatis mutandis to the present subject.

[65]According to one embodiment of the invention, the computer program product comprises instructions enabling the predictive diagnosis of premature delivery in a subject. [66]According to one embodiment, the computer program product comprises instructions enabling the predictive diagnosis of ECUN in a subject.

[67]According to one embodiment, the computer program product comprises instructions enabling the predictive diagnosis of type I diabetes in a subject.

[68]According to one embodiment, the computer program product comprises instructions enabling the predictive diagnosis of sepsis in a subject.

[69]The invention also relates to the use of a computer program product according to the invention for the diagnosis/predictive diagnosis of a pathology or a pathological condition. The characteristics previously and subsequently described in relation to the diagnostic/predictive diagnosis method according to the invention apply mutatis mutandis to the present subject.

[70]The invention finally relates to the management or treatment of a subject whose diagnosis or positive diagnosis of a pathology or pathological condition has been determined as positive using the diagnostic/predictive diagnostic method of the invention. Said treatment may be a curative treatment or a prophylactic treatment depending on the situation. The management may be enhanced clinical monitoring, particularly in the context of the predictive diagnosis of premature birth.

Detailed description of the invention

[71]The present invention thus has as its first object a method, in particular in vitro, for diagnosis or predictive diagnosis of a pathology or a pathological state in a subject, from at least one biological sample taken from the subject and containing microorganisms.

[72]By “diagnosis” is meant in the invention the determination of the presence or absence of a pathology or pathological condition in a subject. A positive diagnosis is understood in the invention as corresponding to the determination of the presence of the pathology or pathological condition in the subject. A negative diagnosis is understood as corresponding to the determination of the absence of the pathology or pathological condition in the subject.

[73]By "predictive diagnosis" is meant in the invention the determination of the risk of developing/occurring/appearing a pathology or the occurrence of a pathological condition in a subject not presenting any symptoms. The positive predictive diagnosis is understood in the present invention as a high risk of appearing the pathology or the pathological condition. Conversely, a negative predictive diagnosis is understood in the present invention as a low risk of appearing the pathology or the pathological condition. [74]A positive diagnosis/predictive diagnosis may be considered determined when the associated certainty is greater than 50%, preferably a certainty greater than or equal to 55%, 60%, 65%, 70%, 75%, 80%, 85%, 90%, 91%, 92%, 93%, 94%, 95%, 96%, 97%, 98%, 99% or equal to 100%. Similarly, a negative diagnosis/predictive diagnosis may be considered determined when the associated certainty is greater than 50%, preferably a certainty greater than or equal to 55%, 60%, 65%, 70%, 75%, 80%, 85%, 90%, 91%, 92%, 93%, 94%, 95%, 96%, 97%, 98%, 99% or equal to 100%.

[75] “Pathology” means a disease, a biological imbalance or discomfort. The pathology corresponds in particular to a digestive pathology, an extra-digestive pathology or a pathology of the newborn, in particular enterocolitis of the type, more particularly ulcerative necrotizing enterocolitis (ECIIN). “Ulcerative necrotizing enterocolitis” means a disease characterized by inflammation and necrosis of the intestinal mucosa. Even more particularly, among said digestive pathologies, we can cite: digestive cancers, that is to say affecting at least one of the organs of the digestive system, chronic inflammatory diseases, such as in particular Crohn’s disease, ulcerative colitis, irritable bowel syndrome and celiac disease.

[76]The pathology is advantageously either a pathology of the organ from which the biological sample is taken, or a pathology of another organ in the environment from which the sample is taken.

[77] “Extra-digestive pathology” means a condition or pathology that does not directly affect an organ of the digestive system but one of the consequences of which is likely to directly or indirectly affect the microbiota of the digestive system and vice versa. Among the extra-digestive or non-digestive conditions and pathologies for which a predictive diagnosis can be carried out by a method according to the invention, we can cite: diabetes, sepsis, obesity, cardiovascular diseases, metabolic diseases, liver diseases, kidney diseases, urogenital diseases, pulmonary diseases, joint diseases, muscle diseases, inflammatory diseases, asthma, allergies, arthritis, neurodegenerative diseases (Parkinson's, Alzheimer's, etc.), psychiatric diseases, behavioral diseases, all types of cancer for all types of organs.

[78]A "pathological condition" means a state of alteration of the functions, morphology or health of an organ or organism, the cause of which is known or unknown, and which is characterized by the presence or absence of one or more signs. A pathological condition includes, in particular, premature delivery. [79]By "condition or pathology of the digestive system" is meant a condition or pathology affecting at least one organ selected from: the mouth, the salivary glands, the pharynx, the esophagus, the stomach, the pancreas, the liver, the gallbladder, the bile duct, the small intestine and the large intestine. The large intestine includes the ascending colon, the transverse colon, the sigmoid colon and the rectum. According to a particular aspect of the method of the invention, said pathology is an intestinal pathology.

[80]"Premature delivery" means delivery occurring before the start of the ^37th week of amenorrhea.

[81]According to a particular aspect of the method of the invention, said pathology is a digestive pathology of a subject chosen from: children, infants (children beyond their first month of life and up to the age of 24 or 30 months) and newborns (children under 28 days according to the definition of the World Health Organization), said newborns being born at term, i.e. between the ^37th week and the end of the ^40th week of amenorrhea, or premature, i.e. born before the ^37th week of amenorrhea.

[82]The term “subject” means an animal or a human being, the animal being in particular a mammal. According to a particular embodiment of the invention, the stage of development of the subject is chosen from: adult (from 18 years), adolescent (12 - 17 years), child (2 - 11 years), infant (28 days - 23 months), newborn (0 - 27 days) and premature newborn (< 37 weeks of amenorrhea). According to a particular aspect of the method of the invention, the subject is a pregnant woman, a newborn, an infant or a human child.

[83]The term “biological sample” means any sample from the subject containing microorganisms. In particular, said biological sample is chosen from: a sample from the digestive system, a sample of excretions, in particular a sample of stool from the subject, a vaginal sample, a cervical sample, a skin sample, and any other biological sample containing microorganisms.

[84]The sample collection is carried out in particular in a conventional and well-known manner by a specialist. A given biological sample comprises a community of microorganisms designated by the term “microbiota”.

[85]According to one embodiment, the sample may correspond to the grouping of several samples taken from various areas of a sampling region in the subject, in order to attempt to obtain the maximum diversity of microorganisms. [86]“Microorganism” means any unicellular or multicellular microorganism such as, but not limited to, bacteria, archaea, viruses, unicellular eukaryotes such as yeasts, etc.

[87]Among the microbiota hosted by a human subject, we can distinguish the skin microbiota, the mucosal microbiota, the pulmonary microbiota, the oral microbiota, the vaginal microbiota, the urinary microbiota, and the microbiotas of the digestive system (oral or salivary microbiota, stomach microbiota, small intestine microbiota, colonic microbiota, anal microbiota). The microbiota present in the stools, or fecal microbiota, corresponds to all the microorganisms found in the stools following transit through the digestive system of a subject, which may reflect the intestinal microbiota in the broad sense with a closer proximity to the colonic microbiota. Transient microorganisms can also be found in this microbiota. The term "microbiome" refers to all the genomes carrying the genes hosted by the microorganisms constituting the microbiota. The microbiome can also be considered as the set of microorganisms including their genomes in a particular biological environment such as the colon.

[88] "Digestive system" means the set of organs of multicellular animals that receives food, digests it to extract nutrients, and excretes waste in the form of fecal matter. The organs of the human digestive system include: the mouth, salivary glands, pharynx, esophagus, stomach, pancreas, liver, gallbladder, bile duct, small intestine, and large intestine. The large intestine includes the ascending colon, transverse colon, sigmoid colon, and rectum. "Excretion" means unusable or toxic waste that is excreted by the subject such as urine, feces, or stool, or secretion products such as bile or saliva.

Step a)

[89]Step a) corresponds to the sequencing of the nucleic acid of the microorganisms present in the biological sample(s), said nucleic acid having been previously isolated from the sample.

Extraction of nucleic acid from the sample

[90]The term “nucleic acid” means all nucleic acid molecules present in the biological sample, in particular deoxyribonucleic acid (DNA) and ribonucleic acid (RNA), including respectively the genes expressing 16S ribosomal RNA (rRNA) and/or those expressing 18S rRNA, in particular rRNA and even more particularly 16S rRNA and 18S rRNA. [91] "16S rRNA-expressing gene" means the DNA nucleotide sequence comprising the nucleotide sequence encoding the 16S rRNA. A gene expressing a 16S rRNA is also referred to as "16S rDNA".

[92]A “gene expressing 18S rRNA” means the DNA nucleotide sequence comprising the DNA nucleotide sequence encoding 18S rRNA. A gene expressing 18S rRNA is also referred to as “18S rDNA”.

[93]Genes expressing the small subunit of rRNA, i.e. genes called “16S rDNA” for prokaryotic microorganisms, such as bacteria and archaea, and “18S rDNA” for eukaryotes, including yeasts, are used to enable the description of the structure of the microbiota (Chakoory et al., 2022).

[94]In order to isolate the nucleic acid from the sample, any commercial nucleic acid extraction kit can be used. It should be noted that the yield (quantity of nucleic acids) of the kits as well as the quality of the nucleic acids can vary depending on the type of sample. It is generally necessary to compare the efficiency of the kits to select the most efficient one. The extraction can be carried out manually or using an automaton. In addition to commercial kits, there are extraction processes for which the reagents are produced directly in the laboratory. There are also extraction protocol standards aimed at homogenizing nucleic acid extraction procedures worldwide. In particular, in the context of ECUN, the H protocol published by the IHMS (International Human Microbiome Standards) can be used for DNA extraction from newborn stools: (see IHMS (human-microbiome.org)).

[95]According to one embodiment of the invention, the method comprises the isolation of the nucleic acid from a plurality of microorganisms present in said biological sample, in particular from all of the microorganisms.

Nucleic acid sequencing

[96]The isolated nucleic acid is then sequenced in order to obtain the nucleotide sequences corresponding to at least one sequence of interest chosen from the group consisting of: a fragment of a gene expressing 16S rRNA, a fragment of a gene expressing 18S rRNA, a fragment of 16S rRNA and a fragment of 18S rRNA (hereinafter referred to as "sequences of interest"). Indeed, 16S rDNA, 18S rDNA, 16S rRNA and 18S rRNA are highly conserved in all microorganisms, but also include discriminating variations between taxa which thus makes it possible to analyze the sequences belonging to the microorganisms and also to distinguish them. Thus, the aim of the sequencing step is to recover all of the sequences corresponding to at least one sequence of interest. Of course, by "set of sequences" means the set of sequences that the sequencing method can obtain. The key point here is that there is no discrimination of certain sequences of interest among those found in the sample, no preselection is carried out. The analysis uses the entire sequencing data.

[97]According to a preferred embodiment, the nucleotide sequences corresponding to at least one sequence of interest chosen from the group consisting of: a fragment of a gene expressing 16S rRNA and a fragment of a gene expressing 18S rRNA are obtained.

[98] “Sequencing” means any known method for determining the nucleotide sequence of a nucleic acid. Among these methods, direct metagenomic sequencing known as “shotgun” is preferred, and is notably described in the document Quince C, et al. Shotgun metagenomics, from sampling to analysis. Nat Biotechnol. 2017 Sep 12;35(9):833-844. Briefly, this type of sequencing involves the fragmentation of the isolated nucleic acid into fragments whose size varies depending on the sequencing platform used (typically from 200 to 550 bp on average for the Illumina® platform and from a few dozen bases to > 100,000 bp for the Nanopore® platform), which are subsequently linked to adapters (here also specific to the platform used) for the preparation of the sequencing library. The libraries obtained are then sequenced using a high-throughput sequencing platform (typically Illumina® or Nanopore®). The sequences obtained are then filtered to remove poor quality sequences and sequences corresponding to the subject's genome, according to well-established principles in the technical field. The filtered sequences are then organized for identification, as seen in detail below.

[99]The use of Illumina® sequencing data from gene capture approaches by hybridization is also preferred and notably described in the document Comtet-Marre, Sophie & Chakoory, Oshma & Peyret, Pierre, (2022), Targeted 16S rRNA Gene Capture by Hybridization and Bioinformatic Analysis. Briefly, the isolated nucleic acid is fragmented and linked to sequencing adapters in a manner similar to the "shotgun" method. In parallel, oligonucleotide probes, in particular biotinylated, complementary to the sequences of interest are synthesized and then hybridized with the sequencing libraries. The complexes formed are captured, in particular using magnetic beads coated with streptavidin, and amplified by PCR using primers complementary to the adapters. The captured and amplified fragments are sequenced with a high-throughput sequencing platform, then filtered, as previously described. The filtered sequences are then organized. Thus, in this context, according to a particular embodiment of the invention, said method comprises a preliminary step of specific isolation of the nucleic acid from a plurality of microorganisms present in said biological sample. [100]Sequencing can also be of the "amplicon sequencing" or "metabarcoding" type, notably described in the document Durazzi, F., Sala, C., Castellani, G. et al. Comparison between 16S rRNA and shotgun sequencing data for the taxonomy characterization of the gut microbiota. Sci Rep 11 , 3030 (2021). However, this type of sequencing is less favored since it involves a preliminary PCR amplification of portions of the 16S rRNA or 18S rDNA using primers, in particular using so-called universal primers which can lead to a biased overrepresentation of certain microorganisms or to the exclusion of certain microorganisms. The use of primers specific to microbial taxonomic groups can also lead to the exclusion of some of the microorganisms present in the sample analyzed. The amplified sequences are ligated to specific adapters to produce sequencing libraries and sequenced using a high-throughput sequencing platform, similar to what is described above.

[101]A “fragment” of a nucleotide sequence means a fragment of at least 20% of the length of that sequence. A “fragment of at least 20%” means a fragment of at least 20%, at least 25%, at least 30%, at least 35%, at least 40%, at least 45%, at least 50%, at least 55%, at least 60%, at least 65%, at least 70%, at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 97%, at least 98%, at least 99% or 100% of the nucleotide sequence in question.

[102]The sequenced 16S rDNA and/or 16S rRNA fragment of microorganisms belongs in particular to prokaryotes. Additionally, the 18S rDNA and/or 18S rRNA fragment also belongs to eukaryotes and micro-eukaryotes.

Step b)

[103]The purpose of step b) is to identify all the microorganisms present in the sample from the sequencing of step a) as well as their relative abundance, and to provide relevant input data to the artificial intelligence model for determining the diagnosis. Here again, “set of microorganisms” means all the microorganisms identifiable according to the sequencing method used. Identifying all the microorganisms present in the sample and providing this set (without the identities absent from the training set) to the artificial intelligence model makes it possible to preserve the maximum individual variability of the subject as well as the associated microbial interactions in the context of a specific pathology or pathological condition and to ensure a personalized diagnosis/predictive diagnosis.

[104]For this purpose, according to one embodiment of the invention, the method comprises organizing the sequenced sequences to reconstruct the nucleotide sequence of at least one part of a gene expressing 16S rRNA and/or of a gene expressing 18S rRNA. In particular, step b) comprises in particular firstly a step of organizing the sequences obtained in step a) by aligning them with known sequences of microorganisms present in a database. Said known sequences comprise at least said sequence of interest selected for the greatest number of known microorganisms, in order to determine direct correspondences or to reconstruct sequences of new microorganisms and/or to obtain longer sequences in order to increase the reliability of the identity of the microorganisms present in the biological sample of the subject. In the context of metabarcoding sequencing, the organization is done in particular by direct correspondence. In the context of the “shotgun” method or of gene capture by hybridization, the organization can be done by direct correspondence and/or reconstruction.

[105]The determined set of microorganisms is in particular selected from those available in online databases, in particular public ones. Among these public databases, the SILVA database (https://arb-silva.de). Another example of databases is the “Greengenes” database (https://greengenes.secondgenome.com/). The person skilled in the art can thus easily determine whether a given nucleotide sequence comes from a known or unknown microorganism, or from a human or animal subject.

[106] Thus, according to a particular embodiment, the method according to the invention comprises a step of reconstructing at least part of the sequence of the gene expressing the 16S rRNA and/or the sequence of the gene expressing the 18S rRNA of the microorganisms present in the biological sample. Of course, the reconstructable length depends on the sequenced length of the fragment of the sequence of interest and the sequencing effort, i.e. the number of readings generated during sequencing (sequencing depth).

[107]More particularly, in a particular embodiment, during the step of reconstructing at least one nucleotide sequence, at least 70% of the length of the gene expressing the 16S rRNA and/or at least 70% of the length of the 16S rRNA is reconstructed. An increase in the size of the reconstructed part allows for greater precision in determining the identity of the microorganism, making it possible to go as far as the taxonomic rank of the species. The length of a 16S rDNA gene being approximately 1500 base pairs on average, a nucleotide sequence of at least 70% of the length of the gene comprises approximately 1050 base pairs, on average.

[108]According to one embodiment of the invention, it uses all of the metagenomic data of the microbiota which then allow the reconstruction of complete sequences of interest and a precise affiliation of the microorganisms of the microbial community at the genus or species level, or even the identification of new microorganisms. [109]The organization step is notably followed by a classification step by taxonomic ranks of the correspondences and/or reconstructions making it possible to determine the identity of the microorganisms present in the subject's biological sample.

[110]Identification can in particular be supplemented by phylogenetic analyses in order to situate the new microorganisms in relation to the closest known microorganisms.

[111]By “identity determination” we mean the identification of microorganisms, following a nomenclature, organized into hierarchical categories (classification by taxonomic ranks), in other words in taxonomic ranks, these categories consist of belonging to the domain of life (least precise rank) to the definition of the species (most precise rank). The taxonomic ranks of interest extend from the phylum to the species. The taxonomic classification is carried out by comparing each reconstructed sequence of interest or whose correspondence is attributed with 16S rDNA sequences and/or 18S rDNA sequences contained in databases. Among the public databases that can be used, we can notably cite again the SILVA database. The most confident taxonomic rank that can be identified depends on several parameters including the type of sequencing, the sequencing parameters, the determined set of microorganisms used for the alignment (see below), etc. The invention thus has the advantage of taking into account each determined identification. There is thus no preselection carried out, making it possible to preserve all the diversity of the subject's sample. This exhaustiveness contributes to obtaining a diagnosis/predictive diagnosis of higher quality than with the methods of the prior art. According to one embodiment, the same taxonomic rank among the taxonomic ranks of interest is preserved for all the sequences. According to a preferred embodiment, the most precise taxonomic rank among the taxonomic ranks of interest for each sequence is determined. This second aspect allows better identification of the microbial diversity of the sample, and ensures a more reliable diagnosis.

[112]By “determination of the relative abundance” is meant the determination for each of the microorganisms considered for the method according to the invention, of the abundance of the microorganism relative to the total abundance of the microorganisms considered for the method according to the invention. The determination of the abundance depends on the sequencing method used, and is well known to those skilled in the art.

Step c)

[113]In this step, an artificial intelligence model previously trained on the basis of a labeled data set determines the diagnosis/predictive diagnosis on the basis of the data obtained in step b). The artificial intelligence model can also take as input at least one clinical data of the subject, as will be seen in detail later.

[114]The artificial intelligence model thus presents an internal structure reflecting the relationship between on the one hand (1) the relative abundance of microorganisms within the sample, as well as optionally at least one clinical data of the subject, and on the other hand (2) the diagnosis/predictive diagnosis of the pathology or pathological state.

[115]The artificial intelligence model is a supervised learning model and corresponds in particular to a classification model, a deep learning model, a neural network (NN), a deep neural network, a decision tree, a K-nearest neighbors model (KNN), a random forest (RF), a naive Bayesian classification (NB), an “Extreme gradient boosting” algorithm (XGBoost), a logistic regression or a support vector machine (SVM). In particular, the artificial intelligence model is a deep neural network with an input layer composed of neurons equivalent to the number of features in the training data, followed by one or more hidden layers and an output layer which gives the result of the diagnosis/predictive diagnosis.

Training

[116]By “previously trained” we mean a process allowing the artificial intelligence model to learn from a set of labeled training data to associate in a weighted manner the identity and abundance of microorganisms present in samples of subjects, and optionally at least one clinical data of these subjects, with the corresponding diagnosis/predictive diagnosis.

[117]The invention thus also relates to a method for training an artificial intelligence model intended to obtain a diagnosis or a predictive diagnosis, said method using a labeled data set.

[118]The labeled data set or training set comprises profiles of training subjects. The training subjects belong to the same species as the subject whose sample(s) are analyzed in the method of the invention. In order to strengthen the training, the training subjects advantageously come from various nations, and in particular from various continents. Parity between the sex types of the subjects in the training set is also advantageous, depending of course on the pathology or of the pathological state considered. These different aspects make it possible to obtain a better representativeness of the subject's microbiotas. Indeed, unlike the prior art which focuses on a restriction of the microorganisms analyzed, the principle of the invention is to preserve all the diversity of the microbiota of each of the training subjects, so that the artificial intelligence model can determine all the possible relationships, independently of any bias introduced by the knowledge at a given time. Contrary to what could be expected with such complex input data, the results obtained following the training give an excellent accuracy of prediction of diagnosis/predictive diagnosis of the physiological or pathological state for which the artificial intelligence model was trained. The inventors were thus able to show that microorganisms with a very low relative abundance, generally excluded from the training for this reason, proved to be very relevant for determining the predictive diagnosis of pathologies and pathological states. What could previously be considered noise is demonstrated here as a discriminating point.

[119]The training subjects may in particular be specifically recruited for this purpose, or may come from one or more databases, in particular public ones, and more particularly from the most exhaustive and diversified subject cohort databases available. These databases include in particular raw sequencing data from one or more samples from each subject, and optionally at least one clinical data from each subject.

[120]The training subjects are notably divided into two groups, namely a training group and a test group. The training group is used to train the artificial intelligence model, and the test group is used to qualify its performance. Typically, the training group represents 80% of all training subjects, and the test group 20%.

[121]The training subject profiles each comprise the identity and relative abundance of the identified microorganisms present in at least one sample of the training subject, as well as optionally at least one clinical data of the training subject. The relative abundances are notably obtained by implementing steps a) and b) described above on samples of subjects, or the single step b) on sequencing data of samples of subjects. The identities of the microorganisms (and therefore their abundance) can be restricted for training to the same given taxonomic rank so that all of the microorganisms are identified at the level of the same rank, starting from the phylum and up to the species. However, no preselection is carried out on the identified microorganisms, notably on the basis of their relative abundance and/or their known involvement in the diagnosis or the predictive diagnosis. According to one embodiment preferred, no restriction on taxonomic rank is made, and the most confident taxonomic rank is retained for all identities.

[122]When the method of the invention is intended for the diagnosis of early delivery of the pregnant woman, the sample(s) of each training subject are in particular taken during the same trimester, and typically during the ^1st , ^2nd or ^3rd trimester, or even the same month.

[123]Learning supervision is achieved by labeling the profiles of training subjects with their phenotype. The subjects are classified into at least two phenotypes, and preferably into two opposing phenotypes. In the context of a diagnosis, the phenotypes of the subjects are notably affected/not affected by the pathology/pathological state. Concerning the predictive diagnosis, the phenotypes of the subjects are notably with appearance/without appearance of the pathology or pathological state. Advantageously, the training set includes a balanced number of each phenotype, or a greater proportion of positive phenotype.

[124]The data of the training subjects are notably normalized. This normalization is in particular of the min-max type on the entire training set. This type of normalization corresponds to a linear transformation of the features in a uniform range, while preserving all the distance ratios of the original data. This is done to prevent the numerical values of the larger features (abundances of microorganisms) from surpassing those of the smaller numerical features, thus minimizing the bias in the discrimination of pathological states. The main objective is to ensure the comparability of the data across microbial samples or groups of samples, such as those classified as diseased or healthy. Indeed, the large variability of the sizes of the databases and the sequencing depth induces strong dependencies among the abundances of the different taxa. Thus, data normalization ensures that all features (taxa) in the data contribute equally to the learning process, although not all features are equally important for the classification decision.

[125]Where at least one clinical datum is used in the input data in addition to the data relating to microorganisms, it is of course relevant to the pathology or pathological condition for which the diagnosis/predictive diagnosis is carried out. By “at least one clinical datum” is meant one, two, three, four, five, six, seven, eight, nine, ten or more than ten clinical data characteristic of the subject.

[126]In particular, in the case of a newborn pathology, the clinical data may belong to the subject himself or to his mother. In this context, it may be used in particular: at least one of the following data:

- the actual age of the subject at which the sample was taken, in number of days of life

- the subject's birth weight,

- the gestational age of the child at birth,

- the subject's mode of birth (vaginal or cesarean section),

- the gender of the subject (masculine, feminine),

- the dosage of blood components or markers of the subject or the mother,

- the dosage of fecal components or markers of the subject or the mother,

- the presence of at least one other pathology in the subject or the mother,

- the administration of medical treatment to the subject or to the mother,

- the mother's ethnicity/nationality,

- feeding the mother and/or newborn,

- the mother’s lifestyle (physical activity, consumption of alcohol, tobacco, drugs, etc.).

[127]By “ethnicity” is meant a group of people who are brought together by a certain number of characteristics. In a method according to the invention, the characteristic “ethnicity” is notably chosen from the group consisting of: “African-American”, “American-Indian”, “Black”, “White”, “Caucasian”, “Hispanic”, “Asian”, “Multi-ethnicity”.

[128]When clinical data are used, they are encoded in particular in the following manner: categorical data (such as gender and mode of birth in the case of newborns) are converted into vectors using "one-hot encoding", i.e. all elements of the vector are converted to 0 except the categorical variable which is converted to 1. Continuous data (actual age, birth weight and gestational age in the case of newborns) are transformed into a discrete variable by creating a set of contiguous intervals ("bins" in English) that cover the range of values of the variable. The clinical data "day of life" is discretized into intervals with an increasing step of 9 (from 0 to 99 days) and 99 (100 to 499 days). A time step of 1 could also be considered over the first 3 weeks of life when the pathology most frequently appears. The clinical data "weight" is discretized into intervals with an increasing step of 99 (from 500 to 2899 grams). The weight of the children can also be followed if necessary by interval of 9 throughout the first 3 weeks of life until the possible appearance of the pathology. Gestational age at birth can be converted into factors due to the limited number of values.

[129]In the case of the diagnosis of premature birth, said clinical data is chosen in particular from:

- the gestation period, - the age of the pregnant woman,

- the ethnicity of the pregnant woman, and

- a combination of these clinical data.

The duration of gestation may in particular be expressed in number of weeks of gestation or designated by the period at which the biological sample is taken.

[130]This period is notably chosen from: the first trimester of pregnancy, the second trimester of pregnancy, the third trimester of pregnancy.

[131]The age of the pregnant woman, in a method according to the invention, can be defined in number of years or by her belonging to an age group. More particularly, the age of the pregnant woman can be attributed to one of the following two groups: “less than 35 years” and “equal to or greater than 35 years”.

[132]Prior to the learning, all of the microorganisms present in each training subject profile are compiled, so as to determine the number of microorganism identity abundance entries of the artificial intelligence model. According to one embodiment, the artificial intelligence model comprises at least 500 microorganism identity abundance entries, in particular at least 600 entries, in particular at least 700 entries, in particular at least 1000 entries, particularly at least 1300 entries.

[133]According to one embodiment, the artificial intelligence model comprises at least 10 determined clinical data inputs, in particular at least 20, particularly at least 30, in particular at least 40.

[134]According to one embodiment of the invention, the method being intended for the predictive diagnosis of early delivery in a pregnant woman, the artificial intelligence model comprises at least 600 entries of abundance of identities of microorganisms and optionally at least 10, in particular at least 15, entries of determined clinical data.

[135]According to one embodiment of the invention, the method being intended for the predictive diagnosis of ECUN, the artificial intelligence model comprises at least 1000, in particular at least 1300, entries of abundance of identities of microorganisms and optionally at least 40, in particular at least 45, entries of determined clinical data.

[136]According to one embodiment of the invention, the method being intended for the predictive diagnosis of type I diabetes, the artificial intelligence model comprises at least 1000, in particular at least 1300, entries of abundance of identities of microorganisms and optionally at least 40 entries of determined clinical data. [137]According to one embodiment of the invention, the method being intended for the predictive diagnosis of sepsis, the artificial intelligence model comprises at least 600, in particular at least 1300, microorganism abundance entries and optionally at least 40 determined clinical data entries.

Microorganism signatures from training

[138]The training method according to the invention makes it possible to highlight different signatures of microorganisms characteristic of a positive diagnosis/predictive diagnosis (hereinafter “first signatures”) or negative diagnosis/predictive diagnosis (hereinafter “second signatures”). By “signature” is meant a set of identities of microorganisms. This method also allows the discovery of new microorganisms.

[139]According to this aspect of the invention, a first signature of microorganisms associated with a diagnosis of the appearance and/or development of ECUN, in particular obtained by a method according to the invention, is characterized in particular by the presence of microorganisms of the species:

- Unclassified Enterobacter,

- Unclassified Enterobacteriaceae,

- Enterococcus faecalis,

- Unclassified Klebsiella,

- Haemophilus parainfluenzae,

- Enterococcus durans and

- Enterobacter cancerogenus.

[140]These microorganisms were in fact found, notably in greater quantities, in biological samples statistically associated with the diagnosis of the presence of ECUN (i.e. with a probability of more than 50%).

[141]A first signature associated with a high probability of premature delivery, notably obtained by a method according to the invention, is characterized in particular by the presence of microorganisms of the genus:

- Anaerococcus,

- Peptoniphilus,

- Prevotella, in particular Prevotella bivia,

- Gardnerella in particular Gardnerella vaginalis,

- Sneathia in particular S neat hi a amnii.

[142]Indeed, these microorganisms were found to be present or present in greater quantity in biological samples statistically associated with a high probability of premature delivery (more than 50%). [143]According to this aspect, a second signature associated with a plurality of microorganisms statistically associated with a diagnosis of absence of ECUN, in particular obtained by a method according to the invention, is characterized in particular by the presence of microorganisms of several species of Lactobacillus associated with non-ECUN cases. Indeed, these microorganisms were discovered as present or present in greater quantity in the biological samples statistically associated with a prediction of absence of ECUN. The second signature associated with a diagnosis of absence of ECUN may comprise other microorganisms, such as: the genera Bifidobacterium, Bacteroides, the species Bifidobacterium longum, Bacteroides fragilis, Lactobacillus casei.

[144]A second signature associated with a high probability of delivery at term (more than 70%), notably obtained by a method according to the invention, is characterized in particular by the presence of microorganisms of the Christensenellaceae family and of the genus:

- Bacteroides, or

- Lactobacillus, in particular Lactobacillus crispatus.

[145]Indeed, these microorganisms were found to be present or present in greater quantities in biological samples statistically associated with a high probability of full-term delivery.

Diagnosis and Predictive Diagnosis

[146]The diagnosis/predictive diagnosis is determined from the identities and abundances of microorganisms determined during step b). Of these data obtained in step b), only those of the microorganisms absent from the training set are purified. In this sense, the larger the training set, the more likely it is that it will be exhaustive, and that no purification will be carried out in the data obtained in step b). However, in the event that a sample from a subject was discovered to include a microorganism identity that was not present in the training set, it is possible a posteriori to re-train the artificial intelligence model with this new input. It is thus possible to obtain a continuous enrichment of the artificial intelligence model, and therefore a continually improved accuracy of the predictions.

[147]The data retained from step b) following the exclusion of microorganisms absent from the training data set are in particular normalized. This normalization is in particular of the min-max type on the basis of the training set.

[148]The diagnosis/diagnosis obtained in step c) may in particular be associated with a certainty/confidence index, typically ranging from 0 to 1, reflecting the probability of correspondence. Thus, the artificial intelligence model can determine a positive diagnosis of a pathology with a confidence index of 0.8, indicating that there is an 80% chance that the analyzed microbiota is associated with this pathology. Conversely, the artificial intelligence model can determine a negative diagnosis with a confidence index of 0.8, indicating that there is an 80% chance that the analyzed microbiota is not associated with the pathology and therefore a 20% chance that it is.

Step d)

[149]The method according to the invention may comprise a step d) of compiling several diagnoses/predictive diagnoses for a final determination of the diagnosis/predictive diagnosis.

[150]According to a particular embodiment of the invention, at least two biological samples from the subject are used, in particular at least three. By “at least two biological samples” is meant two, three, four, five, six, seven, eight, nine, ten or more than ten biological samples from the same subject. The samples can be taken at the same time, or at different times.

[151]According to one embodiment of the invention, when several biological samples are used for the same subject, steps a) to c) are carried out on each sample, such that step d) comprises the compilation of the diagnosis/predictive diagnosis obtained in step c) for each sample and the final determination of the diagnosis/predictive diagnosis. Thus, the diagnosis/predictive diagnosis can be considered positive/negative if more than 50% of the result of steps c) correspond to this state.

[152]According to one embodiment of the invention, when the samples are taken at the same time, step d) makes it possible to reinforce a first diagnosis determined in the first step c), in particular to overcome a potential undesired selection of microorganisms by the choice of the sampling area in a sampling region. Thus, the samples are in particular taken in different areas of the same sampling region, in order to ensure the exhaustiveness of the representation of the microorganisms in the subject's sampling region.

[153]According to one embodiment of the invention, when the samples are taken at different times, step d) makes it possible to monitor the changes in the subject's microbiota and in particular the change in their phenotype (from sick to healthy following treatment, or from healthy to sick), allowing a clinician to confirm a curative effect or to take the necessary measures in the event of the appearance of a pathology or pathological state. Legend of figures

[154]The present invention is further explained by the following figures and examples.

[155]Figure 1 shows an overview of the steps followed for an embodiment of the method for diagnosing a pathology according to the invention from the identification of microorganisms and their abundance in the sample of a subject, followed by a step of predicting the diagnosis/predictive diagnosis using the trained and optimized DNN model.

[156]Figure 2 illustrates the steps of an example of training a deep neural network model according to the invention and the adjustment of its hyperparameters allowing the optimization of the prediction of the diagnosis/predictive diagnosis.

[157]Figure 3 illustrates the prediction performance obtained by the deep neural network model based on the input data provided. The input data are the data from direct metagenomic sequencing (Fettweis cohort) processed by RiboTaxa or by MetaPhlAn3. MetaPhlAn3 uses the high-quality reads from direct metagenomic sequencing to compare them to a reference genome database of microorganisms available at: segatalab.cibio.unitn.it/data/Pasolli_et_al.html and determine the taxonomic composition of the analyzed microbiota (from domain to species) and the relative abundances of the identified microorganisms (TSV file).

[158]Figure 4 shows the performance of deep neural network models trained on data from direct metagenomic sequencing and genus-level metabarcoding.

[159]Figure 5 represents the final structure of an artificial intelligence model (trained deep neural network model) according to the invention optimized to predict ECUN.

[160]Figure 6 shows the true positive rate (on the ordinate) as a function of the false positive rate (on the abscissa) in the context of predicting the occurrence of ECUN, where the AUC is equal to 0.987.

[161] Figure 7 shows the accuracy (ordinate) versus sensitivity (abscissa) in predicting the occurrence of ECUN, where the AUC is equal to 0.992.

[162]Figure 8 shows the 20 input features of the trained deep neural network model contributing most to the prediction of ECU N or non-ECU N phenotypes summarized by the SHAP explainer. [163]Figure 9 illustrates the analysis of the longitudinal follow-up of samples following the prediction of the deep neural network model trained in the context of predicting the occurrence of NEC. The unlabeled circle on the left represents the actual phenotype of the infant. Samples from infants without pathology are indicated in dark gray and samples from NEC infants in light gray. Each labeled circle represents a sample collected from each of the infants and the numbers inside the circles correspond to the day of collection (in days of life). The color of these circles represents the phenotype predicted by the neural network according to the same color code as the unlabeled circles. The single square represents the samples that were reclassified into the “control” group and the double square represents the samples that were reclassified into the “NEC” group.

[164]Figures 10 and 11 represent examples of SHAP plots illustrating the most important features (microorganisms) that influence the prediction towards the control phenotype in the CORTECs cohort. For each feature, the negative values associated with the arrows correspond to SHAP values associated with a contribution towards the prediction of the control phenotype (f(x)=0). The label next to each feature (microorganism) represents its abundance in the sample.

[165]Figures 12 and 13 represent examples of SHAP plots illustrating the most important features (microorganisms) that influence the prediction towards ECUN in the CORTECs cohort. For each feature, the positive values associated with the arrows correspond to SHAP values associated with a contribution towards the prediction of the ECUN phenotype (f(x)=1). The label next to each feature (microorganism) represents its abundance in the sample.

[166]Figure 14 shows the 20 input features of the trained deep neural network model contributing most to the prediction of T1D or non-T1D phenotypes summarized by the SHAP explainer.

[167]Figure 15 represents the longitudinal analysis approach of the predictions made on the set of samples of children who had at least 3 samples in the “sepsis” test set. The final phenotype of the child is determined by the phenotypic group having the greatest number of samples of the same condition.

[168]Figure 16 shows the 20 input features of the trained deep neural network model contributing the most to the prediction of sepsis phenotypes summarized by the SHAP explainer. Examples

Collecting training data

[169]The inventors collected raw microbiota sequencing data and associated clinical data from patient cohorts established as part of studies of different pathologies and pathological conditions: preterm birth (PB), necrotizing enterocolitis (NCE), sepsis and type 1 diabetes (T1D).

[170]The first step consisted in selecting relevant scientific publications that had made these data available. A search by precise keywords was carried out in the PubMed and Google Scholar publication databases. The microbiota sequencing data had to have been obtained by direct metagenomic sequencing, known as “shotgun”. Only prospective studies with samples taken before the onset of the pathology or pathological state, allowing a predictive diagnosis, were retained. In addition, the inclusion of control subjects was required.

Bioinformatics processing of shotgun metagenomic sequencing data

[171] The shotgun metagenomic data were processed with the RiboTaxa bioinformatics chain (Chakoory et al., 2022) to obtain the taxonomic profiles of the microbiota (identification of microorganisms at all taxonomic ranks and associated relative abundances). The RiboTaxa approach consists of reconstructing 16S and 18S rDNA sequences using reference databases, here, the SILVA SSU 138.1 NR99 database (Quast et al., 2013), then allowing identification of microorganisms down to the species level. RiboTaxa performs quality control of the raw reads, reconstruction of the 16 and 18S rDNA sequences, determination of their relative abundance and the identity of the microorganisms.

[172]For each sample, raw reads were provided as input to RiboTaxa. Reads were processed to remove Illumina adapters, known Illumina artifacts, and to trim the ends of reads when the base quality score was below Q20. Resulting reads containing more than one “N”, or with quality scores below 20 averaged over the read, or a length less than 60 bp, were rejected.

[173]High-quality reads were then assembled into full-length to near-full-length 16S and 18S rDNA sequences using two assemblers included in RiboTaxa. MetaRib (Xue et al., 2020) takes all high-quality reads as input while EMIRGE (Miller et al., 2011) uses only reads corresponding to 16S and 18S rDNA. filtered with SortMeRNA (Kopylova et al., 2012). The dual reconstruction approach (EMIRGE and MetaRib) maximizes the reconstruction of genes expressing 16S/18S rRNA and describes the structure of microbiotas as accurately as possible. Although both assemblers (EMIRGE and MetaRib) require a reference database (here SILVA, which is the most complete and of high quality), it is possible to reconstruct sequences very distant from the reference sequences, which thus makes it possible to identify new microorganisms that would not be identified by other approaches (quantitative PCR, classic analyses of metagenomic data, PCR amplification of a portion of the gene expressing 16S rRNA then sequencing).

[174]For the reconstruction of the 16S/18S rRNA-expressing gene, the default parameters were used, except for the parameters that depend exclusively on the sequencing length of the input data:

- parameter A "max_read_length" represents the longest read size of the input dataset,

- parameter B “insert_mean” represents the average size of inserts of paired reads and

- the C parameter “insert_stddev” represents the standard deviation of the size distribution of inserts of paired-end reads.

Parameters B and C were estimated using the script “mean_size.py”, available at: gist.github.com/timoast/af73c0e9fac00187ee49.

[175]The reconstructed 16S and 18S rDNA sequences were then clustered with a 97% identity threshold and classified into different taxonomic ranks, from domain to species, using the SILVA database. After removing human 18S rDNA as a contaminant, relative abundances were calculated by RiboTaxa.

[176]All the obtained taxonomy tables were grouped into a single table containing all profiles at the phylum, class, order, family, genus and species level using RiboTaxa's RiboTaxa_group_taxonomy.sh script.

[177]For training the artificial intelligence model below, all microorganisms identified in all samples were retained, instead of applying selection before training, in order to preserve microbial diversity and inter-individual microbial interactions. Artificial Intelligence Model

[178]For the predictive diagnosis of each pathology/pathological state presented in examples below, a fully connected deep neural network model, corresponding to the previously described “computer program product”, was implemented and optimized on the same strategy, using the Python programming language and dedicated libraries such as scikit-learn, Tensorflow (https://tensorflow.org), Keras (https://github.com/keras-team/keras-tuner) and Adam (Kingma and Ba, 2017).

[179]The architecture of the deep neural network consists of an input layer whose number of neurons depends on the number of input characteristics (number of microorganisms identified and number and nature of clinical data), hidden layers whose number and the number of corresponding neurons are determined during training and optimization of the model, and an output layer containing 2 neurons, one for a “pathology/pathological state” output, the other for a “no pathology/pathological state” output.

[180]In order to obtain the most efficient model possible, different mathematical functions were selected and the values of the hyperparameters of the deep neural network were optimized according to the training data obtained for each of the pathologies.

[181] The rectified linear unit activation function (ReLLI) was used for all hidden layers. Activation functions play an important role in training neural networks by providing the nonlinearity needed for the model to learn complex representations. The neuron dropout technique on each hidden layer was also employed to mitigate overfitting of the neural network, which leads to poor generalization of the model and reduced performance on new data. Neuron dropout is a learning method that involves randomly removing neurons during model training, with the removed nodes being excluded from subsequent steps. The output layer activation function uses the Softmax function to assign a value based on a probability between 0 and 1 to each class (pathology/disease state, no pathology/disease state). This value allows the model to make a ‘risk of pathology’ or ‘no risk of pathology’ decision.

[182]Different values of other hyperparameters were tested. The number of epochs (number of times the entire dataset is propagated through the neural network) was varied from 1 to 40. The cross-entropy loss between the target value and the predicted value was optimized over the epochs with learning rates, ranging from 0.0001 to 0.01. The number of hidden layers was varied from 1 to 3 and the number of neurons in the first hidden layer from 32 to 512 with an increasing step size of 32. To facilitate model convergence, the number of neurons in the hidden layers was set to half that of the previous layer. These optimizations were implemented using Keras (https://github.com/keras-team/keras-tuner).

[183]To define the best combination of hyperparameters, the training dataset was split with a ratio of 8:2 to obtain 80% training data and 20% test data. A K-fold cross-validation was applied with the training data (Figure 2). These were divided into K subsets of almost equal size; K-1 subsets being used for training the model and the remaining subset for validating the produced model. In this way K models were built, with each time a redistribution of the K subsets and the definition of new hyperparameters. The best combination of hyperparameters for each model was selected by averaging the accuracy metric of the K models. The optimized model was then trained into a final classification model using the training dataset and tested on the test data.

[184]The performance of the optimized deep neural network model was estimated on the test data (20% of the entire data set) by comparing the phenotype predicted by the model and the phenotype observed in the subject. For example, if the model correctly classifies a sample from a subject with a pathology or disease state, it is considered a true positive (TP), otherwise it is a false negative (FN). On the other hand, if the model correctly classifies a sample from a subject without a pathology or disease state, it is considered a true negative (TN), otherwise it is a false positive (FP). Due to class imbalance (samples from subjects with the pathology or disease state are generally less abundant in the datasets), model performance was measured using several metrics: accuracy (total number of correct predictions over the total number of subjects), sensitivity (rate of subjects with the pathology correctly predicted by the model or true positive rate), specificity (rate of subjects without the pathology correctly predicted by the model or true negative rate), area under the curve (AUC) of the receiver operating characteristic (ROC)/AUROC, and precision-recall AUC (PR-AUC). [185]Accuracy is calculated as follows:

TP + TN

Accuracy =

TP + FP + TN + FN

[186]The sensitivity is calculated as follows:

Sensitivity

[187]Specificity is calculated as follows: Specificity

[188]Finally, AUROC corresponds to the area under the ROC curve which shows the sensitivity (rate of true positives) as a function of the specificity (rate of true negatives). The PR-AIIC measures the sensitivity over the precision (ratio of TPs to the total number of TPs and FPs). AUCs were calculated using the scikit-learn package (Pedregosa et al., 2011) and plotted using matplotlib (Hunter, 2007) (v3.1). The 95% confidence intervals (CIs) of the AUCs were estimated using the bootstrap method (Efron and Tibshirani, 1994) with 1,000 iterations. ROC curves and the Sankey plot were generated using matplotlib and plotly (v5.15.0), respectively.

[189]A SHAP (SHapley Additive exPlanations) approach has been used to explain the output of any machine learning model. Models can be interpreted by calculating the importance of input data related to the classification performance of the model. The importance of input elements (metadata, microorganisms) was calculated using SHAP. SHAP's DeepExplainer function is a method for decomposing the output of a deep neural network (prediction) by assigning contribution values to each data of the neural network input. This function allows highlighting the input data with the most weight in predicting a phenotype.

Normalization and vectorization of training data

[190]Relative abundances were then normalized to avoid the influence of highly abundant taxa via the transformation below, called min-max normalization:

[191]

X — (x - Xmin) / (Xmax ~ Xmin) [192]where: x is the original data, x' is the normalized data. x _min and x _ma x are respectively the minimum and maximum values of the original value (abundance). The above equation is a linear transformation that preserves all abundance ratios of the original data after normalization.

[193]In addition, one or more clinical data were used depending on the pathology or pathological condition for which a diagnosis or predictive diagnosis was made.

[194]The clinical data were either discrete or continuous variables. To better handle the data, continuous variables were transformed into discrete values through a discretization step. This process involves transforming a continuous-valued variable into a discrete variable by creating a set of contiguous intervals (or bins) that span the range of values of the variable. Grouping numerical features into interval-based groups is beneficial for classification and can significantly improve model performance.

[195]The next step was to apply a one-hot encoding technique on all discrete data using LabelEncoder from the scikit-learn library (Pedregosa et al., 2011). Thus, the discrete values were vectorized, i.e. all elements of the vector were converted to 0 except the categorical variable, which was converted to 1.

[196]For each dataset, a dataset comprising vectorized clinical data and normalized microbial abundances served as input data for model training.

Longitudinal analysis of predictions carried out on all samples from the same subject.

[197]For the ECUN, sepsis and DT1 datasets, stool samples were collected serially for the same subject allowing a longitudinal analysis of the predictions made for the same child. This approach made it possible to measure the ability of a model to perform from the first sample despite the dynamics of the microbiotas. In this approach, a subject was considered correctly classified when all of its samples were correctly classified. Subjects for whom at least one sample was misclassified were considered misclassified. The inventors also took advantage of the longitudinal sampling of subjects to explore the evolution of the microbiota over time and to redetermine the final phenotype of each subject misclassified by the deep neural network. Subjects for whom the phenotype prediction was unequal across samples and who had at least 3 samples in the test dataset were identified. The number of samples in each phenotypic group was calculated and the final phenotype of the subject was determined by the phenotypic group with the largest number of samples. The phenotype thus determined was compared with the observed phenotype (affected by a pathology or a pathological state, not affected). Finally, a lollipop plot was generated to visualize this longitudinal follow-up analysis approach using the ggpubr package (v0.4.0).

Example 1: Predictive diagnosis of preterm birth using a deep neural network trained with vaginal microbiota data.

Collecting data from the training game

[198]The inventors selected five studies that looked at the vaginal microbiota in relation to preterm birth using the English keywords: “vaginal microbiome”, “shotgun metagenomics” and “premature birth”: Feehily et al., 2020; Fettweis et al, 2019; Goltsman et al, 2018; Pace et al, 2021; Tortelli et al. 2021.

[199]Raw data and associated metadata were obtained for each cohort under the accession numbers listed in Table 1 or upon request. ENA represents European National Archive, NIH represents National Institute of Health, SRA represents Sequence Read Archive.

Table 1

[200]For each cohort, the following sample metadata were retained:

- term birth (TB) or preterm birth (PTB) phenotype,

- time of sample collection: 1st trimester of pregnancy i.e. 1-13 weeks gestation, 2nd trimester of pregnancy i.e. 14-26 weeks of gestation, 3rd trimester of pregnancy i.e. >= 27 weeks of gestation,

- age of participants (less than 35 years old, greater than or equal to 35 years old),

- ethnic group (African-American, American-Indian, Asian, Black, Caucasian, Hispanic, Multi-ethnic, White) and

- the participant's identifier (ID).

[201]A total of 1290 samples were retrieved. Only samples collected during pregnancy were used. Table 2 represents the general properties of the individual studies included for training the deep neural network. These present the number of samples or the number of participants, TB represents a term birth and PTB a preterm birth.

Table 2

Data preprocessing

[202]When preprocessing using RiboTaxa, for the reconstruction of 16S/18S rDNA genes, the parameters A, B and C described in the following Table 3 were used. Table 3

[204]Species-level taxonomic profiles together with clinical data containing information on participant ethnicity, age, phenotype, and time of sample collection were used to train a deep neural network.

Comparison of the optimized deep neural network with other learning models

[205] The performance of the optimized deep neural network was compared with three state-of-the-art classification algorithms: k-nearest neighbors (KNN), logistic regression (LR), and support vector machine (SVM). All these models were implemented in Python (version 3.9.10). The scikit-learn library (vO.24.2) was used. Each model was trained on the same data set, i.e., the 1290 samples. The best hyperparameters and configurations were identified using the grid-search cross-validation (GSCV) method of scikit-learn. The GSCV method identifies the best combination of hyperparameters during the 10-fold cross-validation process to achieve the optimal performance of the models.

[2061Comparison of deep neural network trained with microbial diversity data obtained with RiboTaxa and MetaPhlAn3

MetaPhlAn 3 (Beghini et al. 2021) uses clade-specific marker genes to identify the presence and relative abundance of microorganisms from metagenomic data. MetaPhlAn3 was used to process shotgun metagenomic data from the Fettweis cohort with default parameters and using the CHOCOPhlAnSGB database (Jan21 release). Species-level microbial diversity profiles were used as inputs for training a deep neural network. The performance of the resulting model was compared to a deep neural network model trained with diversity data obtained by pre-processing the same sequencing data with RiboTaxa. Results

[207]The processing by the RiboTaxa bioinformatics chain of the metagenomic sequencing data of the five studies made it possible to obtain complete to almost complete 16S or 18S rDNA sequences with a minimum length of 1045 bases. A precise description of the vaginal microbiota was thus obtained for each sample, this description includes an identification at the species level and the relative abundance of each species. The approach of reconstructing genes expressing 16S and/or 18S rRNA makes it possible to reconstruct sequences very distant from the reference sequences, which thus makes it possible to identify new microorganisms that would not be identified by other approaches (quantitative PCR, classic analyses of metagenomic data, PCR amplification of a portion of the gene expressing 16S rRNA then sequencing).

[208]Input data composed of vaginal microbiota profiles associated with four metadata (phenotype, ethnicity, age, time of sample collection) were used to perform deep neural network training to distinguish term from preterm deliveries. The training dataset included 17 categorical values (vectorized clinical data) and 636 numerical values (normalized microbial abundances).

[209]The following Table 4 brings together the characteristics of the deep neural network obtained.

Table 4

[210]The final model evaluation was performed on the test set consisting of 239 samples that were not used for building the artificial intelligence learning model. The diagnostic accuracy reached 84.10%, while the sensitivity and specificity reached 63.41% and 88.38% respectively. In repeated tests of the deep neural network, the inventors demonstrated an AUROC of 0.877 ± 0.11.

[211]On the same input dataset, the performance of the deep neural network (DNN) was superior compared to logistic regression (LR), K-nearest neighbors (KNN) and a support vector machine SVM models which demonstrate similar accuracy, yet quality (Table 5). Table 5

[212]The performance of predicting the risk of preterm delivery was improved by focusing the training of a model on data from samples collected during the second trimester only. The model then showed a sensitivity higher by 10% to 73.40% while maintaining very good accuracy and specificity at 82.58% and 85.61% respectively. This result shows that the relevant selection of input data is necessary to obtain the best performing results.

[213]The strategy of obtaining microbial diversity profiles by reconstruction of the gene expressing 16S and/or 18S rRNA allowed obtaining the best performances compared to the use of other marker genes (Figure 3).

Example 2: Comparison of the performance of models trained on metabarcoding data versus direct metagenomics data for the predictive diagnosis of preterm birth.

Data collection for the training game

[214]The Fettweis et al. study included 232 women whose vaginal samples were analyzed by both direct shotgun metagenomics and metabarcoding (sequencing of the 16S rDNA V3-V4 region). Raw shotgun metagenomics data (952 Gb) and metadata for the Fettweis et al. cohort were obtained after National Institute of Health data access approval. This dataset represented 173 women who delivered at term (667 vaginal samples, scored TB) and 55 women who delivered preterm (155 vaginal samples, scored PTB). Raw metabarcoding data (58 Gb) from 749 TB samples (173 women) and 205 PTB samples (55 women) were open access and were downloaded from HMP DACC (https://portal.hmpdacc.org).

Preprocessing of training data

[215]For data from shotgun sequencing, RiboTaxa chaining was used. For the reconstruction of the gene expressing 16S and/or 18S rRNA, the parameters A, B and C were as follows: --max_read_length = 301, --insert_mean = 120, --insert_stddev = 300. [216]Metabarcoding sequencing data were processed with DADA2 (R package 1.16). A first step of quality control filter of the reads was performed with standard parameters: maxN=0, truncQ=2, rm.phix=TRUE and maxEE=2. After learning the error rates with the “learnErrors” function, the reads were dereplicated to obtain unique sequences or ASVs (Amplicon Sequence Variants) with their abundance (number of reads corresponding to each unique sequence). The sample inference algorithm was then applied to correct the dereplicated sequences from the quality profiles of the raw sequences. The pairs of reads thus obtained were merged to obtain the complete amplicon sequences. Finally, chimeric sequences were identified and eliminated and the remaining ASVs were taxonomically classified with the “assignTaxonomy” function and the SILVA SSU 138.1 NR99 database (Quast et al., 2013, https://benjjneb.github.io/dada2/training.html) were used. Absolute abundances of ASVs within each sample were converted to relative abundances using the “transform_sample_counts” function of the R package phyloseq (2.10).

[217]Since the metabarcoding sequencing approach focuses on the analysis of a portion of the 16S rDNA, taxonomic analysis cannot be performed at the species level. Therefore, microorganism identifications were only performed at the genus level and the two sequencing approaches were compared with the taxonomic rank of the genus.

[218]Microbial taxonomic profiles at the genus level obtained from shotgun metagenomics and metabarcoding data, as well as clinical data (ethnicity, age, phenotype, and time of sample collection) were pre-transformed as previously described.

Deep Neural Network Training

[219]A deep neural network was implemented and trained for each of the direct metagenomics and metabarcoding data, and then the performance of the produced models was evaluated with the test dataset.

Results

[220]The model trained on the metabarcoding data (at the genus level) achieved an accuracy of 80.10% (on a total of 191 samples of the test data), a specificity of 86.84% (on 152 TB samples) and a lower sensitivity of 53.84% (on 39 PTB samples) (Figure 4). [221] Regarding the data from direct metagenomics, the model trained at the taxonomic rank of the genus allowed an improvement of almost 10% in sensitivity reaching 63.33% (on 33 PTB samples) for a specificity of 87.12% (on 132 TB samples).

[222]These results illustrate that the approach of reconstructing genes expressing 16S and/or 18S rDNA allows a better identification of microorganisms thanks to long rDNA sequences and thus leads to a more efficient model compared to metabarcoding which provides short rDNA sequences. The biases linked to the PCR amplification of amplicons, inherent to metabarcoding, can also impact the representativeness of microbial diversity and degrade the performance of the classification model.

Example 3: Predictive diagnosis of ECUN using a deep neural network trained with data from fecal microbiota.

Data collection for the training game

[223]The following keywords were used to identify studies that investigated NEC in preterm infants and included stool samples: “premature infants” AND (“stool microbiome” OR “intestinal microbiome”) AND “shotgun metagenomics” AND “necrotizing enterocolitis”. At the end of the selection process, two studies were retained: Masi et al. (2021) and Olm et al. (2019).

[224] Raw shotgun metagenomic sequencing data and metadata from Masi et al. (2021) were downloaded from ENA under BioProject PRJEB39610 (n = 524; 974.51 GB). In addition to their own cohort, Olm et al. (2019) also used sequencing data from different previously published datasets. All raw data and metadata used in the Olm cohort (n = 1038 in total) were downloaded from SRA under the BioProjects: PRJNA294605 (n = 141; 596.53 GB), PRJNA417343 (n = 184; 152.21 GB) PRJNA396794 (n = 295; 1.35 GB). Tb), PRJNA376566 (n = 358; 905.22 Gb) and SRA study SRP052967 (n = 60; 114.21 Gb).

[225]A total of 1,305 control samples (from 160 infants) and 257 N-EC samples (from 48 infants who developed NEC) were used for model training. No samples collected after the onset of NEC were analyzed. Five clinical data characteristics common to both studies were collected: phenotype (control, NEC), mode of birth (vaginal, cesarean), gender (boy, girl), gestational age at birth (in weeks), day of life (DOL, in days) and birth weight of the newborn (in grams), and infant identification. Only Masi et al. (2021) reported that children in their cohort received probiotics (Lactobacillus acidophilus, Bifidobacterium inf antis, and B. bifidum). [226]The clinical data of the subjects are presented in Table 6.

Table 6

Preprocessing of training data

[227]When preprocessing the sequencing data using RiboTaxa, for the reconstruction of the gene expressing 16S and/or 18S rRNA, the parameters A, B and C were as follows:

- Masi cohort: --max_read_length = 151, --insert_mean = 144, --insert_stddev = 100;

- Olm cohort: --max_read_length = 301, --insert_mean = 120, --insert_stddev = 100.

[228]For model training, as previously reported, microbial species abundance profiles were normalized and clinical data were discretized and vectorized. A dataset comprising 47 categorical values and 1,282 numerical values (normalized microbial abundances) for each of the samples was obtained.

Model evaluation on external data [229]To further evaluate the performance of the optimized model, 50 fecal samples from 17 preterm infants including 7 who developed ECIIN, from the CORTECs cohort followed by the inventors, were analyzed. In addition, 40 infants from two published cohorts (Ward et al. 2023 and Schwartz et al. 2023) were also included to test the model performance.

[230]The constitution of the CORTECs cohort was approved by the ethics committee of CPP-Sud-Est VI (protocol code 2021/CE 26, the approval date is May 4, 2021). The CORTECs cohort aims to address prenatal and postnatal risk factors for ECU N. All prematurely born children hospitalized in the neonatal intensive care unit (NICU) of the Clermont-Ferrand University Hospital (France) were proposed to enter the cohort. Written informed consent was obtained from the families of study participants before enrollment. Infant stools were collected daily during their NICU stay, between May 2021 and June 2022. Stools were collected in a diaper using a sterile loop and then dispensed into eNAT buffer (Copan) before being briefly held at 4°C. Samples were stored at -80°C until DNA extraction.

[231] Cases of ECUN were identified by physicians based on systemic and abdominal findings and radiographic features. They were stratified according to disease severity according to Bell stages. Cases of ECUN were matched to a control preterm infant (two to one case) who did not develop ECUN. Case-control matching was based on gestational age at delivery, mode of delivery, sex, birth weight, and pre- and postnatal antibiotics. For each ECUN infant, available samples were selected within a 1-week window before the onset of ECUN and samples from corresponding control cases were matched according to the age of the ECUN subject.

[232]Genomic DNA was extracted using the standard operating procedure for fecal samples (protocol H) recommended by the International Human Microbiome Standards (IHMS SOP 07 V1). DNA quality was assessed using the Nanodrop 2000 fluorometer (Thermo Scientific) and the Agilent 4150 TapeStation system with ScreenTape Genomic DNA (Agilent). DNA quantity was assessed using the Qubit 3 fluorometer (Invitrogen) with the Qubit dsDNA High Sensitivity Assay Kit (Invitrogen). Hybridization capture of the 16S rRNA-expressing gene and sequencing data processing: Capture probes were designed to target the 16S rRNA-expressing gene (Gasc et al., 2016). Sequencing libraries were produced for each sample using the Nextera XT library preparation kit. The gene capture experiment was performed according to the protocol described by Ribière et al. (2016) and Comtet-Marre et al. (2023). Briefly, biotinylated RNA capture probes were obtained by in vitro transcription. 500 ng of libraries were were mixed with 2.5 μg salmon sperm DNA and incubated with 500 ng biotinylated probes in hybridization buffer for 24 h at 65°C. Probe/target heteroduplexes were captured using 500 μg streptavidin-coated paramagnetic beads (Dynabeads M-280 Streptavidin, Invitrogen). Beads were collected using a magnetic stand (Ambion), washed once with 500 μL 1x SSC/0.1% SDS buffer, and then three times with 500 μL 0.1x SSC/0.1% SDS buffer preheated to 65°C. Captured DNA fragments were eluted with 50 μL of 0.1 M NaOH and transferred to a sterile tube containing 70 μL of 1 M Tris-HCl buffer pH 7.5. Captured DNA was amplified by PCR with 25 cycles using primers complementary to Illumina adapters. To increase enrichment efficiency, a second capture cycle was performed. Captured DNA was then sequenced on the Illumina MiSeq 2 x 300 bp platform.

[233]For the Ward et al. 2023 cohort, infants were recruited from two level III neonatal intensive care units (NICUs) in Cincinnati (USA) and one level III NICU in Birmingham (UK). Reported NICU cases were Bell stage II or III. A total of 115 direct metagenomic sequencing data were used, from 3 NICU neonates (9 samples) matched to a total of 35 control preterm neonates (106 samples). Stool samples were collected between days 3 and 22 of life. Raw data and metadata were downloaded from ENA (BioProject PRJNA63661).

[234]Schwartz et al. 2023 is a prospective US study to investigate factors associated with bloodstream infection and gut microbiome in neonatal intensive care units. In this cohort, two infants (8 samples) developed necrotizing enterocolitis (N-ECU) and were selected. Raw data and metadata were downloaded from the NCBI repository (BioProject PRJNA884103).

[235]For all 3 cohorts, clinical data included phenotypes (control, NEC), mode of birth (vaginal, cesarean), sex (male, female), gestational age (in weeks), day of life (in days), and newborn birth weight (g) as well as child identifier (Table 7). Table 7

[236]Raw sequencing data from the three cohorts were processed using the RiboTaxa pipeline and all input data were normalized or transformed as described previously. Species that were not present in the training samples were excluded as the model cannot account for them. For each sample, the relative abundance table of microorganisms at the species level concatenated with the subject's clinical data was used as input to the trained model. Each prediction was compared to the child's phenotype (control or ECIIN). SHAP plots were also generated. The final prediction of the children was also determined using longitudinal samples from the same infant using the same longitudinal follow-up analysis approach.

Results

[237]All sequencing data were analyzed with the RiboTaxa pipeline (Chakoory et al., 2022), allowing the reconstruction of complete to near-complete 16S rDNA genes to provide an accurate description of the gut microbiota down to the species level, including the identification of dominant (>1%), subdominant (<1%) and rare (<0.1%) microorganisms thus allowing the best representativeness of the microbiota. [238] It is frequently demonstrated that Enterobacteriaceae bacteria are more abundant in children who will develop ECIIN. Differential analysis of the diversity of fecal microbiota data intended for training also shows a significantly higher mean relative abundance of unclassified Enterobacter and unclassified Enterobacteriaceae in ECIIN samples compared to control preterm samples (p < 0.05, Welch's f-test). Despite these repeated observations in studies, they still do not represent a reliable microbial signature of ECIIN risk because they are not universally found and the notion of an associated relative abundance threshold is difficult to determine.

[239]To address this issue, a deep neural network was developed and trained using 1,402 features (1,355 microbial species identified in stool and 47 clinical data: 10 gestational age groups, 18 weight groups, 15 DOL, 2 birth modes, and 2 sex groups) (Figure 5). The final model contained 448 units (neurons) in the first hidden layer and a total of 3 hidden layers. The model was trained in less than 5 min on an i86linux32 computer, 4.0 GB RAM x 8 cores (32.8 GB total)

[240]The following Table 8 brings together the characteristics of the deep neural network obtained.

[241]Table 8

[242]The final model evaluation was performed on the test set consisting of 313 samples (from 140 infants). The model showed an excellent accuracy of 94.9%, a specificity of 95.8% (249 out of 260 control samples) and a very good sensitivity of 90.6% (48 out of 53 ECUN samples). In repeated tests of the deep neural network, the inventors demonstrated an AUROC of 0.987 ± 0.01 (Figure 6), suggesting a good balance between sensitivity and specificity and a PR-AUC value of 0.992±0.002 (Figure 7). Interestingly, Olm et al. applied gradient-enhanced classification to distinguish ECUN infants from controls using taxonomic data and obtained only 64% accuracy (Olm et al., 2019).

[243]In 92.8% of ECUN infants (26 out of 28) and 90.1% of control children (101 out of 112) the predictive diagnosis was correct for all samples from the same child, demonstrating the robustness of the diagnosis despite the dynamic colonization of the intestinal microbiota of newborns.

[244]Due to the seriousness of the consequences of the occurrence of ECUN in preterm newborns, the inventors sought to improve the performance of the model by using a majority voting strategy, determining a predictive diagnosis from the majority phenotype predicted for the different samples from the same child, when they were available.

[245] In this study, only 16 samples (from 16 infants) out of 313 tested samples were misclassified by the deep neural network. Among the 16 misclassified samples, 6 belonged to 6 infants (2 controls and 4 ECIIN) for whom more than three serial samples were present in the test dataset. Thus, 22 longitudinal samples belonging to the 6 infants were considered. This approach allowed to determine the correct phenotype of each child.

[246]The SHAP approach implemented in the deep neural network allows the identification of key species contributing to the model prediction, which can be similar to complex microbial signatures of pathology or healthy state. The 20 most important features contributing to the model prediction are presented in Figure 8.

[247]The four most important contributors were Lactobacillus spp. species and their high SHAP values were associated with control child samples. The 16S rDNA-based microbiota characterization approach shows its full power here with the identification by RiboTaxa of two unclassified bacteria, which could not be revealed with other metagenomic sequencing data analysis approaches. These bacteria, bacterium_129 and bacterium_ARbO3, contributed to the prediction of the control phenotype. Phylogenetic analysis revealed that bacterium_129 was potentially a novel Lactobacillus species sharing 96.32% identity with L. casei strain Dwan5, while bacterium_ARbO3 shared 99.71% identity with Bacillus cereus strain ADY07.

[248]In contrast, the affiliated species Enterobacter unclassified, Syntrophomonas uncultured, Streptomyces vanillaeus, Enterobacteriaceae unclassified and Enterococcus faecalis contributed the most to the ECUN classification.

[249] Interestingly, low abundant species such as Ruminococcus sp., uncultured Staphylococcus, Streptococcus parasanguinis and Proteus spp. were also observed to contribute to ECUN classification, while unclassified Bifidobacterium were associated with control preterm infant samples. [250]The model's predictive performance comes mainly from microbiome data. Excluding clinical features during training resulted in performance values (AUROC = 0.931, PR-AIIC = 0.956) that were not significantly different from those of the model including clinical data (p > 0.05, Mann-Whitney U test between ROC and accuracy-specificity curves with and without metadata).

[251]To evaluate the performance of the model on data external to those used for training the model, the inventors used data from 3 cohorts (France, USA, England).

[252]In the CORTECs cohort, species of the family Enterobacteriaceae (unclassified Klebsiella, unclassified Escherichia-Shigella, and Enterobacter spp.) that are commonly associated in the pathogenesis of NEC were present in both groups and varied in relative abundance across infants. Across all cohort samples, the optimized deep neural network followed by the longitudinal prediction analysis approach performed on all infant samples illustrated in Figure 9 demonstrated a sensitivity of 100% (7 NEC stage 1a infants, corresponding to 21 samples) and a specificity of 80% (8 out of 10 controls, corresponding to 23 samples).

[253]Similarly, in the Ward cohort, a sensitivity of 100% (3 infants with NEC representing 9 samples) and a specificity of 86% (30 control infants out of 35, corresponding to 90 samples) were achieved. In the Schwartz cohort, a sensitivity of 100% (2 infants with NEC, corresponding to 8 samples) was achieved.

[254]In summary, the prediction of the phenotype of samples from the 3 external training cohorts resulted in a sensitivity of 100% and a specificity of 84.4%. Thus, the inventors have achieved a very powerful model, capable of effectively classifying samples from different areas and practices of the NICU despite the heterogeneity of the microbiome between the cohorts.

[255]Among the features contributing to the different predictions (Figures 10 and 11), the prediction of control samples was mainly related to the presence of a higher abundance of Lactobacillus spp. including L. rhamnosus, L. casei and Lactobacillus sp. In contrast, an ECUN prediction was related to a higher abundance of Enterococcus faecalis, Veillonella ratti, unclassified Klebsiella, Enterococcus durans, Enterobacter cancerogenus, Clostridium neonate or C. perfringens. Low abundance species such as uncultured Staphylococcus, Haemophilus parainfluenzae and Staphylococcus epidermidis contributed to the prediction of ECUN in some samples, highlighting a trend towards co-variation between dominant and rare species suggesting the existence of a complex network of ecological interactions between these species. Interestingly, it was observed that depending on the children, the contribution profiles of the microorganisms varied, demonstrating the interest in considering the maximum number of microorganisms for training the model in order to effectively take into account all the inter-individual variability of the microbiotas. There are therefore several microbial signatures for the same pathology, reinforcing the interest in not selecting a restricted number of microorganisms for training predictive diagnostic models.

[256]Clinical data also contributed to the classification of samples into one of the two phenotypes. Birth weight <800 g and gestational age <30 weeks were the two factors often associated with ECUN, while vaginal delivery and gestational age >31 weeks were associated with samples from non-ECUN infants.

Example 4: Predictive diagnosis of type 1 diabetes in children using deep neural network

Data collection for the training game

[257]The following keywords were used to identify studies that investigated type 1 diabetes in children and collected stool samples before the identification of the pathology: “infants” AND (“stool microbiome” OR

"intestinal microbiome") AND "shotgun metagenomics" AND "Type 1 diabetes". This research resulted in the identification of an international study "The Environmental Determinants of Diabetes in the Young (TEDDY)" carried out in the United States (Colorado, Florida, Washington) and in Europe (Finland, Germany, Sweden) (TEDDY Study Group, 2008).

[258]The primary objectives of the prospective study were to identify environmental and genetic factors triggering or protecting against the development of islet cell antibodies or type 1 diabetes (Rewers et al., 2018). For this, 7013 children from the general population were recruited, with a predetermined risk of type 1 diabetes of 3% and 788 children with first-degree relatives with type 1 diabetes and with a predetermined risk of type 1 diabetes of 10%. Medical visits took place quarterly until the age of 4 years, then every 6 months until the age of 15 years. Participants were followed by blood sampling every three months for measurements of autoantibodies directed against islet cells and detection of diabetes. Stool samples were collected longitudinally between 3 and 72 months of life to characterize the gut microbiota by metabarcoding and direct metagenomic sequencing. Each child with diabetes was matched to one or two controls. [259]In this example, only direct metagenomic sequencing data were used and data from children with autoantibodies without type 1 diabetes were excluded. Data from controls of these children were also excluded. Thus, the inventors used a total of 6,955 metagenomic data corresponding respectively to 1,975 IA+DT1 samples (from 91 IA+DT1 children who tested positive for one or more autoantibodies and who were diagnosed with type 1 diabetes), 273 T1D samples (from 19 T1D children who tested negative for one or more autoantibodies but who were diagnosed with type 1 diabetes) and 4,707 control samples (from 468 control children of T1D and IA+DT1 children). Five clinical data were aggregated: phenotype (control, AI+T1D, T1D), sex (boy, girl), month of life at sampling (in months), child identification, and day of life of the child at diagnosis of T1D (in days). Child matching information was also recorded. Raw data (4.96 TB) and metadata were received after data access approval by the National Institute of Health.

[260]The clinical data of the subjects are presented in Table 9

Prétraitement des données d’entrainement Table 9

Preprocessing of training data

[261]When preprocessing using RiboTaxa, for 16S/18S rDNA gene reconstruction, the parameters A, B and C were as follows: --max_read_length = 102, --insert_mean = 200, --insert_stddev = 100.

[262]The IA+DT1 and T1D children were grouped into a single group of type 1 diabetic children for model training, subsequently designated as TD1. Thus, training was performed on the entire dataset (4707 samples from 144 control children and 110 T1D children) to produce a “no prior” model, and three subsets of the data were created based on the month in which T1D was diagnosed. For this, the child’s day of life at diagnosis was converted into months by dividing it by 30 days. The groups were then established as follows: T1D model “24-48 months” (2361 samples from 68 control children and 52 T1D children), T1D model “48-72 months” (1101 samples from 23 control children and 20 T1D children) and T1D model “24-72 months” (3193 samples from 83 control children and 66 T1D children). For each group, only control children matched with the included T1D children were kept. Only samples taken before T1D diagnosis were kept. The models were designated by an interval of children’s ages at the time of T1D diagnosis and included in the model. These intervals cover at most a period of 2 to 6 years (24-72 months), corresponding to the period when the majority of T1D cases were diagnosed.

Results

[263]Controlled and high-quality species-level relative abundance profiles together with three clinical data (phenotype, sex, months of life at sampling) were used to train 4 deep neural networks for predictive diagnosis of T1D risk.

[264]Each model had a different number of input features: DT 1 “no priors” model (1476 microbial species, 71 clinical data groups: 69 sample groups and 2 sex groups), DT 1 “24-48 months” model (1305 microbial species and 42 clinical data groups: 40 sample groups and 2 sex groups), DT 1 “48-72 months” model (1014 microbial species and 17 clinical data groups: 15 sample groups and 2 sex groups), and DT1 “24-72 months” model (1354 microbial species and 59 clinical data groups: 57 sample groups and 2 sex groups). For each model, all species detected in all samples were retained. [265]Hyperparameters were varied for each model (Table 10). Model training was performed on: i86linux32, 4.0 GB RAM x 8 cores (32.8 GB total), no GPU, and took 2 min or less.

[266]The following Table 10 brings together the main optimal hyperparameters of the deep neural network models thus obtained.

Table 10

[267]The performance of the models targeted on a “DT 1 onset age” window was overall the best compared to that of the “without prior” model taking all the data (Table 11), with in particular a sensitivity ranging from 70.8% to 76.5% for these models against 63% for the “without prior” model. This result illustrates once again the importance of selecting the data in a relevant manner.

Table 11

[268]The serial sampling of children carried out on the TEDDY cohort was used to apply a longitudinal analysis approach to the predictions made on the entire sample of each child as described previously. This approach made it possible to obtain for the models targeted on the age groups, the correct identification of 63.2% at 81.3% of children who later developed type 1 diabetes and 68.1% to 71% of children without type 1 diabetes.

[269]The prediction of type 1 diabetes pathology is weighted by a set of microorganisms as illustrated in Figure 14.

Example 5: Predictive diagnosis of sepsis using deep neural network models

Data collection for the training game

[270]The following keywords (in English) were used to identify studies that investigated sepsis in newborns: (“newborns” OR “premature infants”) AND (“stool microbiome” OR “intestinal microbiome”) AND “shotgun metagenomics” AND (“sepsis” OR “bloodstream infection”). The inventors also selected studies that included clinical information such as: mode of birth (vaginal or cesarean), gender (male-female), gestational age (in weeks), actual age (in days of life) and birth weight (in grams). Finally, two studies were retained.

[271] Raw metagenomic sequencing data and metadata from Heston et al., 2023 were downloaded from Sequence Read Archive (SRA) as part of BioProject PRJNA947616 (n = 622; 1.17 TB). Raw data from the Schwartz et al., 2023 cohort were downloaded from SRA as BioProject PRJNA884103 (n = 195, 234.7 GB) and metadata were received from the study authors.

[272]A total of 418 and 167 metagenomic data were extracted from Heston et al., 2023 and Schwartz et al., 2023, respectively. Children who developed other pathologies such as necrotizing enterocolitis and children born at term (>37 weeks of amenorrhea) were excluded. In addition, no samples collected after the onset of sepsis were analyzed. Five clinical metadata features were collected and reported in both studies, such as phenotypes (control, sepsis), mode of birth (vaginal, cesarean), gender (boy, girl), gestational age at birth (in weeks), day of life (DOL, in days), birth weight of the newborn (in grams), and infant identification.

Preprocessing of training data

[273]When preprocessed using RiboTaxa, for the reconstruction of the gene expressing 16S/18S rRNA, the parameters A, B and C were as follows:

- Heston cohort: --max_read_length = 152, --insert_mean = 144, --insert_stddev = 124; - Schwartz cohort: --max_read_length = 302, --insert_mean = 268, --insert_stddev = 144

[274]Microbial diversity data were normalized and clinical data were discretized and vectorized to obtain 44 categorical values (9 gestational age groups, 16 weight groups, 15 DOL groups, 2 birth modes, and 2 sex groups) and 637 numerical values (microbial abundances).

Results

[275] Direct metagenomic (“shotgun”) sequencing data were used to describe the microbiota at high resolution (species level). 585 metagenomic stool sample data (486 from 87 preterm infants and 99 from 29 preterm infants who subsequently developed sepsis) were analyzed using RiboTaxa (Chakoory et al., 2022), allowing the identification of a total of 637 unique species. This uniformity allows a single model to accommodate data from diverse study protocols. Controlled, high-quality species-level relative abundance profiles and 5 clinical data (gestational age, birth weight, day of life at sampling, mode of birth, and sex of the child) were used to train a deep neural network to predict the risk of sepsis before the onset of infection leading to pathology.

[276]The deep neural network model was trained and then trained using 681 different features (637 microbial species and 44 clinical data groups). All species detected in all samples were retained, instead of applying a selection before training to preserve inter-individual variations in microbiota between infants. A total of 42,882 trainable parameters were tested and the optimal hyperparameter setting for the final model had 64 units (neurons) in the 1st hidden layer and a total of 3 hidden layers (Table 12). The model training was performed on: i86linux32, 4.0 GB RAM x 8 cores (32.8 GB total), without GPU and took 2 min.

[277]The following Table 12 summarizes the main characteristics of the obtained deep neural network model. Table 12

[278]The evaluation of the trained deep neural network was performed on the test dataset consisting of 117 samples (from 60 control children and 14 children with sepsis). The model demonstrated an accuracy of 92.3%, a sensitivity of 72.2% and a specificity of 96.0%. In repeated tests, the inventors demonstrated an AUROC of 0.941 ± 0.013 and a PR-ALIC value of 0.942± 0.011 suggesting a good balance between sensitivity and specificity.

[279]Among the children who subsequently developed sepsis, 72.2% showed a risk of sepsis in all their tests. Conversely, 96% of the control children showed no risk of sepsis in any of their tests. The model therefore provides an excellent predictive diagnosis allowing the risk of sepsis to be identified from the first sample taken.

[280]Where possible, children who had samples predicted with the wrong phenotype, a longitudinal analysis of the predictions made on all their samples was performed on the principle described above. Two control newborns each had 3 samples allowing this analysis and the majority of their samples were predicted with the correct phenotype (Figure 15), thus allowing 96.7% of children who did not develop sepsis to be correctly identified.

[281] The decomposition of the contributions of the different input data showed that the contribution of clinical data was notable. They represented 11 of the 20 most important features contributing to the validation of the model (Figure 16). The features of gestational age of 25 and 28 weeks of amenorrhea and weight 500-599 grams were associated with the prediction “sepsis” while gestational ages of 29 and 30 weeks of amenorrhea were associated with the prediction “control”, reflecting the observed fragility of the most preterm infants. This list also included microorganisms such as Bifidobacterium species associated with the prediction of the control group and often correlated with diet, particularly with breastfeeding, while Streptococcus and Staphylococcus species were associated with the prediction of the sepsis group. [282]Thus, to assess the importance of clinical datasets in predicting sepsis in infants, the inventors excluded the microbial diversity data and the new model was trained only on the five clinical data which represented 44 input features for the deep neural network. The newly developed model showed a decrease in sensitivity, with 61.1% (versus 72.2%) but still with a very good specificity of 97.0%. This result confirms the important weight of clinical data in the predictive diagnosis of sepsis and the necessary contribution of fecal microbiota data to obtain greater sensitivity.

[283]The performance of the deep neural network model trained with the “microbiota combined with clinical data” datasets was also tested by processing metagenomic data at different taxonomic levels (Phylum, Class, Order, Family, Genus). The models were evaluated with the same test dataset of 117 samples from control children and 18 samples from children who developed sepsis.

[284]Table 13 summarizes the performance of the different models developed.

Table 13

[285]These results demonstrate that deep neural network models trained on microbial diversity data retained at the taxonomic rank of the order and class are more efficient than at the taxonomic rank of the species. Nevertheless, the classes/orders associated with the sepsis group remain very broad and do not allow precise identification of species potentially linked to a risk of sepsis. On the other hand, the model trained on the species is slightly less efficient but allows for a list of microorganisms involved in the pathology, which could allow clinicians to adapt the treatment according to the microorganisms identified in infants.

Claims

Claims 1 . In vitro method for predictive diagnosis of a pathology or pathological condition in a subject, from at least one biological sample taken from the subject and containing microorganisms, said method comprising the following steps: a) sequencing, from the nucleic acid isolated from said at least one biological sample, the nucleotide sequences corresponding to at least one sequence of interest selected from the group consisting of: a fragment of a gene expressing 16S ribosomal RNA (rRNA), a fragment of a gene expressing 18S rRNA, a fragment of 16S rRNA, a fragment of 18S rRNA, b) from the sequencing of step a), determining the identity and relative abundance of the microorganisms present in said sample, c) determining the predictive diagnosis of said pathology or pathological condition by an artificial intelligence model from at least the abundances of the identities obtained in step b), said model artificial intelligence having been previously trained on the basis of a labeled dataset, where the labeled dataset comprises profiles of training subjects, each training subject profile comprising the identity and relative abundance of all the microorganisms identified in at least one sample of said training subject without any preselection, where each training subject profile is labeled with the phenotype of the training subject from which it originates, said training subject phenotype being classified as without occurrence or with occurrence of the pathology or pathological condition, and where from the data of step b) only the abundances of the identities of the microorganisms that were not present in the labeled dataset are excluded. 2. Method according to claim 1, in which during the training of the artificial intelligence model and during step b), the identity of each microorganism corresponds to the most confident taxonomic rank. 3. A method according to claim 1 or 2, wherein the training subjects have multinational origins. 4. Method according to claim 1 or 2, in which the labeled data set comprises at least one determined clinical data, where each training subject profile includes a value for the or each determined clinical data item, and wherein step (c) includes providing the artificial intelligence model with the subject's corresponding value for the or each determined clinical data item. 5. Method according to one of claims 1 to 3, in which at least two biological samples are used, where steps a) to c) are carried out on each sample and where the method comprises a step d) of compiling the predictive diagnosis obtained for each sample and of final determination of the predictive diagnosis. 6. Method according to any one of claims 1 to 4, in which step b) comprises the organization of the nucleotide sequences obtained in step a) to reconstruct the sequence of at least 70% of the length of said at least one selected sequence of interest. 7. Method according to any one of claims 1 to 5, being intended for the predictive diagnosis of early delivery in a pregnant woman. 8. Method according to claims 3 and 6, wherein said at least one determined clinical data category is selected from the group consisting of: age, ethnicity, trimester of pregnancy and a combination thereof. 9. Method according to any one of claims 1 to 5, being intended for the predictive diagnosis of ulcerative necrotizing enterocolitis in an infant. 10. Method according to claims 3 and 8, wherein said at least one category of determined clinical data is selected from the group consisting of: age in number of days since birth, birth weight, gestational age, mode of birth, gender, diet of the infant's mother, the result of the dosage of blood components or markers, the administration of medical treatment, the presence of at least one other pathology and a combination of these. 11. Method according to any one of claims 1 to 5, being intended for the predictive diagnosis of type I diabetes in a child. 12. Method according to any one of claims 1 to 5, being intended for the predictive diagnosis of neonatal sepsis in an infant. 13. Computer program product comprising executable instructions, which when executed on a computer allow the implementation of step c) of the method according to any one of claims 1 to 12. 14. A method for training an artificial intelligence model intended to obtain a predictive diagnosis, said method using a labeled data set comprising profiles of training subjects, where each training subject profile comprises the identity and relative abundance of all the microorganisms identified in at least one sample of said training subject without any preselection, and where each training subject profile is labeled with the phenotype of the training subject from which it originates, said training subject phenotype being classified as having developed or not having developed the pathology or pathological condition.