What is an AI drug discovery tool?
An AI drug discovery tool is an open-source agent that automates one or more steps of the drug development pipeline — target identification, virtual screening, ADMET prediction, lead optimisation, molecular dynamics, or experimental design — using LLMs to reason over chemistry, biology, and clinical data. Twelve open-source projects now cover the major steps. This post compares them and tells you which one to pick for which job.
All twelve are MIT or Apache-licensed, actively maintained or paper-anchored. Star counts and stable flags verified against GitHub on 2026-05-25.
The 12 projects at a glance
| Project | Stars | License | Best for | Source |
|---|---|---|---|---|
| ChemCrow | 914 | MIT (stable) | General chemistry reasoning | ur-whitelab (Univ. Rochester) |
| Virtual Lab | 683 | MIT (stable) | Protein / nanobody design | Stanford Zou Group |
| MDCrow | 240 | MIT (stable) | Molecular dynamics | ur-whitelab |
| OriGene | 212 | MIT | Target discovery + disease mechanism | GENTEL Lab |
| DrugClaw (QSong) | 134 | MIT | DTI / ADR / DDI / pharmacogenomics | Q. Song lab |
| ChemGraph | 113 | MIT | Computational chemistry pipelines | Argonne National Lab |
| BloClaw | 105 | MIT (arXiv) | Multi-modal cheminformatics | Beijing 1st Biotech + PLA Gen. Hospital |
| AIAgents4Pharma | 86 | MIT | Full pharma pipeline | VirtualPatientEngine |
| DrugClaw (org) | 50 | Apache-2.0 | Rust-powered ADMET + screening | DrugClaw org |
| ChemClaw | 48 | MIT | Modular chemistry skills | AI4Chem |
| TCM-Agent | 18 | MIT | Traditional Chinese Medicine network pharmacology | AITCM |
| PyMolClaw | 3 | MIT | PyMOL natural-language visualisation | junior1p |
Star count alone is misleading — ChemCrow's 914 reflects historical first-mover status (it's been around since 2023 and is now stable/paper-locked), while ChemGraph's 113 reflects steady current development by a national lab.
Which AI drug discovery tool should I use?
Decision tree:
- General-purpose chemistry reasoning → ChemCrow
- Molecular dynamics simulations from natural language → MDCrow
- Multi-omics target identification → OriGene
- Drug-target interactions, ADR, drug-drug interactions → DrugClaw (QSong)
- Protein / nanobody design with multi-agent coordination → Virtual Lab
- Computational chemistry on HPC clusters → ChemGraph
- Full pharma R&D pipeline orchestration → AIAgents4Pharma
- High-performance ADMET screening in Rust → DrugClaw (DrugClaw org)
- Lightweight modular chemistry skills → ChemClaw
- Traditional Chinese Medicine pharmacology → TCM-Agent
- PyMOL figures for publications → PyMolClaw
- Multi-modal (RDKit + ESMFold + docking) → BloClaw
Now in detail.
ChemCrow — the pioneer (2023)
ChemCrow is the original chemistry AI agent from ur-whitelab (LangChain-based, published 2023). 914 stars. We've marked it stable — the codebase is the reference implementation, not actively developed.
Most modern chemistry agents either fork ChemCrow's design or react against it. The skill list (compound synthesis planning, retrosynthesis, property prediction, reaction mechanism) became the de facto template for the category.
Best for: anyone starting from scratch who wants the reference architecture. Fork ChemCrow, swap LangChain for your framework of choice, adapt the tool list.
Virtual Lab — Stanford protein design
Virtual Lab from Stanford Zou Group is a multi-agent system for nanobody design, with a principal investigator agent coordinating specialised research agents. 683 stars, marked stable (paper-anchored).
The architectural lesson is the "PI agent" pattern — one agent acts as the human investigator, prompting and reviewing the others. This decomposition reduces context bloat compared to a single monolithic agent.
Best for: protein / nanobody engineering specifically, or as a reference implementation for multi-agent science workflows you want to build yourself.
MDCrow — molecular dynamics through chat
MDCrow automates molecular dynamics simulation setup, execution, and analysis using natural language — from ur-whitelab. 240 stars, stable. Conversational interface over OpenMM-style pipelines.
The user value is the simulation-setup pain. MD setup files are notoriously error-prone (force field choice, water model, ionisation state); MDCrow turns "simulate the binding pocket of PDB 6LU7 with explicit water at 300K" into a working trajectory.
Best for: structural biology and biophysics labs that want to run MD without writing input scripts manually. Reuses ChemCrow's design philosophy.
OriGene — self-evolving target discovery
OriGene is a self-evolving virtual disease biologist for mechanism-guided therapeutic target discovery from GENTEL Lab. 212 stars. Automates disease mechanism analysis and drug target identification, with a self-improvement loop for the discovery prompts themselves.
The differentiator is the closed loop — most target ID tools are one-shot. OriGene iterates: propose hypothesis → test against literature → refine.
Best for: target identification stage, especially for diseases with unclear mechanisms (rare diseases, complex polygenic conditions).
DrugClaw (QSong) — the knowledge graph powerhouse
DrugClaw by Q. Song lab is an agentic RAG system with 57 skills across 15 task categories — DTI (drug-target interaction), ADR (adverse drug reaction), DDI (drug-drug interaction), pharmacogenomics, repurposing. 134 stars. Built on LangGraph.
The strength is breadth and integration. Rather than excelling at one task, DrugClaw integrates a wide drug-intelligence knowledge graph and lets you query across DTI / ADR / DDI / repurposing in one session.
Best for: pharmacovigilance, drug repurposing, or any work needing integrated drug-knowledge queries.
ChemGraph — Argonne National Lab quality
ChemGraph is an agentic framework for computational chemistry and materials science workflows from Argonne (ALCF). 113 stars. Orchestrates chemistry simulation pipelines on HPC systems.
The institutional credibility matters here. ChemGraph is the only project on this list that comes from a US national lab, which means the test coverage and HPC integration are real (Polaris, Aurora benchmarked).
Best for: computational chemistry on HPC clusters, materials science, anyone with access to DOE compute resources.
BloClaw — multi-modal cheminformatics
BloClaw is an omniscient multi-modal agentic workspace for scientific discovery from Beijing 1st Biotech + PLA General Hospital. 105 stars (plateaued since 2026-04). Integrates cheminformatics (RDKit), protein folding (ESMFold), molecular docking, and autonomous RAG. XML-Regex dual-track routing, runtime sandbox for data viz. arXiv:2604.00550.
The unique value is the multi-modal stack in one agent — most tools pick one modality (chemistry OR structure OR docking), BloClaw runs the full pipeline.
Best for: workflows that genuinely need all three (cheminformatics → folding → docking) without orchestrating separate tools.
AIAgents4Pharma — the full pipeline orchestrator
AIAgents4Pharma is a multi-agent platform spanning drug discovery, development, and clinical pipeline automation from VirtualPatientEngine. 86 stars.
The positioning is end-to-end. Where other tools focus on one stage (discovery OR development OR clinical), AIAgents4Pharma tries to cover all three with specialised agents for each.
Best for: teams wanting a single platform for the full pharma R&D loop. Trade-off: less depth in any one stage than specialists like OriGene (discovery) or MDCrow (simulation).
DrugClaw (org) — Rust-powered ADMET
DrugClaw by the DrugClaw org is an AI research assistant built with Rust for high performance — covers compound screening, ADMET prediction, and lead optimisation. 50 stars, Apache-2.0.
Different team from the QSong DrugClaw (the name collision is real — see our DrugClaw disambiguation page if you've confused them). The Rust runtime makes it materially faster for high-throughput virtual screening.
Best for: large-scale screening campaigns where Python's overhead per molecule actually matters (think millions of compounds, not thousands).
ChemClaw — modular chemistry skills
ChemClaw is a curated skills package for chemistry-focused AI workflows from AI4Chem. 48 stars. Installable modularly via npx with stable and experimental skill tiers.
The differentiator is composability. Rather than a full agent, ChemClaw ships skill modules you can wire into Claude Code or any compatible runtime. Pick stable skills for production, experimental skills for exploration.
Best for: teams already running Claude Code who want chemistry skills without adopting another full agent platform.
TCM-Agent — Traditional Chinese Medicine
TCM-Agent is an LLM-powered multi-agent system for Traditional Chinese Medicine network pharmacology from AITCM. 18 stars. Compound queries against PubChem, target analysis via TTD, molecular similarity via Morgan fingerprints.
The vertical focus is rare in this space — TCM network pharmacology has specific data sources (Chinese herbal databases, multi-target compound profiles) that general drug-discovery agents don't handle natively.
Best for: TCM research, multi-target compound analysis, herbal medicine ingredient profiling.
PyMolClaw — natural-language molecular visualisation
PyMolClaw turns natural language into publication-quality PyMOL figures from junior1p. 3 stars. 13 scripts covering structure alignment, binding sites, Goodsell-style illustrations.
The user pain it solves: PyMOL's command syntax is famously hostile to newcomers. PyMolClaw lets you say "show the binding pocket coloured by hydrophobicity with the ligand as sticks" and get a working .pml + rendered figure.
Best for: structural biologists who want figures without the PyMOL learning curve. Also useful in combination with MDCrow (run MD → visualise with PyMolClaw).
Which tool for which step in the pipeline?
| Stage | Recommended tool | Backup |
|---|---|---|
| Target identification | OriGene | DrugClaw (QSong) |
| Virtual screening (small) | ChemCrow + custom | ChemGraph |
| Virtual screening (large, millions) | DrugClaw (org, Rust) | — |
| ADMET prediction | DrugClaw (org) | AIAgents4Pharma |
| Drug-target interactions | DrugClaw (QSong) | OriGene |
| Adverse drug reactions / DDI | DrugClaw (QSong) | — |
| Drug repurposing | DrugClaw (QSong) | OriGene |
| Molecular dynamics | MDCrow | ChemGraph |
| Protein / nanobody design | Virtual Lab | BloClaw |
| Multi-modal (chem + structure + docking) | BloClaw | — |
| Publication figures | PyMolClaw | (manual PyMOL) |
| TCM / herbal pharmacology | TCM-Agent | — |
| Full pipeline orchestration | AIAgents4Pharma | — |
| Modular skills for Claude Code | ChemClaw | — |
FAQ
Are these AI drug discovery tools free?
Yes. All twelve are open-source (MIT or Apache 2.0). Free for academic and commercial use including modification.
Do I need a GPU?
Depends on the LLM and the task. The agents themselves run wherever Python runs. Cloud LLMs (Claude, GPT, Gemini) need no GPU. Local LLMs (Ollama, vLLM) need GPU memory matching model size. Molecular dynamics (MDCrow) needs GPU for non-trivial systems.
Which has the most peer-reviewed credibility?
ChemCrow (2023 paper) and Virtual Lab (Stanford Zou Group, paper-anchored) are the most cited. BloClaw has an arXiv preprint. OriGene and MDCrow have papers in domain journals.
Can I chain multiple tools in one workflow?
Yes. Common pipelines: ChemCrow plus MDCrow (chemistry reasoning then MD validation), OriGene plus DrugClaw (target ID then DTI/ADR screening), Virtual Lab plus PyMolClaw (design then figure). The OpenClaw skill model makes inter-tool composition straightforward.
Which one is best for a pharmacology PhD student?
Start with ChemCrow for general chemistry reasoning, add DrugClaw (QSong) for DTI / ADR / repurposing queries. If your project involves MD, add MDCrow. This three-tool combination covers ~80% of pharma-grad workflows.
Are any of these used in industry?
The Rust-based DrugClaw (org) targets high-throughput screening which is industry-relevant. AIAgents4Pharma's positioning (full pipeline) is also industry-flavoured. ChemCrow is referenced in pharma research papers. None has a public list of commercial deployments — open-source adoption is hard to track.
What's the difference between the two DrugClaws?
Different teams. QSong-github/DrugClaw = 134⭐, LangGraph plus 57 skills, drug-intelligence knowledge graph. DrugClaw/DrugClaw = 50⭐, Apache-2.0, Rust runtime focused on ADMET plus screening. We list both because they're independently useful for different scales.
Try it
- General chemistry → start with ChemCrow
- Target ID → OriGene
- DTI / ADR → DrugClaw (QSong)
- Molecular dynamics → MDCrow
- Protein design → Virtual Lab
- Browse all → Claw4Science drug-molecular group
Building a drug discovery agent we haven't listed? Submit it — we feature serious chemistry/pharma work regardless of star count, as long as the use case and methodology are clear.
Originally published 2026-03-29 covering 8 tools. Last updated 2026-05-25 — added ChemClaw, TCM-Agent, PyMolClaw, BloClaw; refreshed all star counts; added stable markers for paper-locked projects (ChemCrow, Virtual Lab, MDCrow); restructured for AI-Overview citability.