An LLM-powered autonomous framework for transcriptome analysis with decision-aware workflow orchestration.
📄 Preprint: Kang B. (2026). ARIA: Adaptive Reasoning for Integrated Analysis. Research Square. doi:10.21203/rs.3.rs-9500973/v1
ARIA goes beyond traditional fixed pipelines by using a Large Language Model as a reasoning engine that evaluates intermediate results and adaptively selects the next analysis step — mimicking how an experienced bioinformatician thinks.
- Adaptive Decision Making: Automatically adjusts analysis strategy based on intermediate results (e.g., switches to GSEA when DEGs are insufficient)
- Multi-method Validation: Cross-validates results using DESeq2, edgeR, and limma-voom
- Comprehensive Reporting: Generates publication-ready HTML reports with embedded visualizations
- Reproducible: Docker-based execution ensures full reproducibility
- Extensible: Modular architecture allows adding new analysis tools
┌─────────────────────────────────────────┐
│ Layer 1: Reasoning Engine (LLM) │
│ - QC evaluation & decision making │
│ - Strategy adaptation │
│ - Literature-based interpretation │
├─────────────────────────────────────────┤
│ Layer 2: Decision Protocol │
│ - 8 Decision Points (DP1-DP8) │
│ - Rule-based + LLM-reasoning hybrid │
├─────────────────────────────────────────┤
│ Layer 3: Execution Engine │
│ - Docker container management │
│ - R/Python script generation & run │
│ - External API calls │
├─────────────────────────────────────────┤
│ Layer 4: Analysis Modules │
│ - DESeq2, edgeR, limma-voom │
│ - fgsea, WGCNA, STRING DB │
│ - Cell type deconvolution │
└─────────────────────────────────────────┘
| DP | Trigger | Action |
|---|---|---|
| DP1 | QC metrics available | Pass/Fail/Warning assessment |
| DP2 | DEG count < threshold | Switch to GSEA, adjust LFC cutoff |
| DP3 | Multi-condition design | Choose per-group or multi-factor model |
| DP4 | Unexpected signature | Add cell type deconvolution |
| DP5 | Method disagreement | Cross-validate with multiple DE tools |
| DP6 | Pathway results | Literature search + hypothesis generation |
| DP7 | Parameter uncertainty | Sensitivity analysis |
| DP8 | Analysis complete | Generate audience-appropriate report |
# Install
pip install aria-omics
# Run with default settings
aria analyze --input samplesheet.csv --genome GRCm39 --output results/
# Run with specific LLM backend
aria analyze --input samplesheet.csv --llm claude --output results/docker pull shoo99/aria:latest
docker run -v $(pwd):/data aria:latest analyze --input /data/samplesheet.csvARIA has been evaluated on 4 public RNA-seq datasets spanning different difficulty levels:
| Dataset | Accession | Difficulty | DEGs | Use Case |
|---|---|---|---|---|
| SEQC | GSE49712 | Easy | ~8,000 | Accuracy (qRT-PCR ground truth) |
| Bottomly | GSE26024 | Moderate | ~300 | Reproducibility (mouse brain) |
| Airway | GSE52778 | Complex | ~3,000 | Design recognition (paired) |
| Fmr1 KO | GSE84989 | Hard | ~100 | Adaptive strategy switching |
If you use ARIA in your research, please cite our preprint:
Kang B. (2026). ARIA: Adaptive Reasoning for Integrated Analysis — An LLM-Powered Framework for Autonomous Transcriptome Analysis with Decision-Aware Workflow Orchestration. Research Square. https://doi.org/10.21203/rs.3.rs-9500973/v1
BibTeX:
@article{kang2026aria,
title = {ARIA: Adaptive Reasoning for Integrated Analysis --- An LLM-Powered Framework for Autonomous Transcriptome Analysis with Decision-Aware Workflow Orchestration},
author = {Kang, Byeongsoo},
journal = {Research Square},
year = {2026},
month = {April},
doi = {10.21203/rs.3.rs-9500973/v1},
url = {https://doi.org/10.21203/rs.3.rs-9500973/v1},
note = {Preprint}
}MIT License. See LICENSE for details.