Liquid Agent is a chemoinformatics-first research agent prototype designed to explore patent space for medicinal chemistry applications. This project is built to test collaboration between cheminformatics experts and AI tool builders, with a goal of evaluating synergies before scaling the team.
We are building an agent that can:
- Accept flexible natural language input (e.g., compound, target, property of interest) via an LLM interface to support broader and more intuitive search.
- Retrieve structurally similar compounds from the SureChEMBL database.
- Identify related patents from the USPTO dataset.
- Extract and summarize experimental data (e.g., potency, assays, solubility) from patent documents.
- Analyze structure-activity relationships (SAR) by comparing analog series within patents.
- Identify "hidden" or cryptic compounds disclosed in patents.
- Provide overviews of compound landscapes for a given biological target.
- Python 3.10+
rdkitfor molecular similarity and fingerprinting.- [
langchain,langgraph] (planned) for agentic workflow orchestration. - [
polars] for fast dataframe operations. surechemblfor chemical-patent mappings.USPTOpublic patent data with API access via:- Sample key:
gJNDuYDHqyCCBNWo54eOktlaLkinpGEF - Reference: Amazon Bedrock Healthcare/Life Sciences USPTO Agent
- Sample key:
- https://github.com/microsoft/magentic-ui Framework 1
- https://github.com/microsoft/autogen Framework 2
- Or without framework ...
Early experiments live in the notebooks/ directory:
01_explore_langchain.py— Using LangChain with Ollama backend02_explore_litellm.py— Streaming output from local model via LiteLLM
A CLI tool is scaffolded in src/liquidagent/cli.py (entry point: biomimir) and will be extended to handle:
- Compound input:
biomimir search --smiles "<SMILES>" - Target input:
biomimir search --target "<gene/protein>"
- SureChEMBL paper: SciFinder to SureChEMBL: Interfacing Chemical Search Tools with Public Patent Data
- USPTO integration reference: AWS Bedrock Agent example 14-USPTO-search
This project is a prototype to test and co-develop agentic workflows for chemistry-aware patent search. The work spans:
- Integrating SureChEMBL compound similarity search.
- Building compound-to-patent mappings using public USPTO datasets.
- Parsing and extracting chemical and experimental data from patent documents.
- Performing SAR (structure-activity relationship) extraction and reasoning.
- Accepting natural language input via an LLM interface to drive agent workflows.
- Creating CLI commands and modular tools that can later compose into a LangGraph-style agent.
This is an opportunity to explore autonomy, creativity, and design in the context of building a deep, domain-specific research agent for early-stage drug discovery.
Paired exmaple of extraction from patents and paired exmaples. (from users's end)