What you can achieve
Start from the outcome, not the tool. Most goals are served by more than one tool that does roughly the same thing — here's the capability and which tool fits.
Pick the outcome you need; the note under it names the tool that fits. Most outcomes have more than one tool, differing in speed, coverage, or the input they take.
When you describe a goal in chat, the orchestrator makes this choice for you. This page is the map behind that choice.
Structure
Predict a protein's 3D structure
Get atomic coordinates for a protein you can't look up.
- AlphaFold — fetches a pre-computed structure by UniProt ID (seconds, no GPU). Try this first.
- ESMFold — predicts from sequence. Use it for novel, engineered, or mutated sequences that aren't in a database.
- PDB — downloads an experimental structure (X-ray / cryo-EM / NMR) when one has been deposited — the highest-confidence option.
- ESM3 — design and embeddings, not just folding.
Pick: in a database by UniProt ID → AlphaFold or PDB. Novel / mutant / engineered → ESMFold.
Find where a drug could bind
- P2Rank — detects and ranks druggable pockets on a structure. This is the entry step before docking.
Binding and docking
Docking is a three-step capability, not three interchangeable tools — run as much of the chain as your question needs.
- P2Rank — where is the pocket, and is it druggable?
- DiffDock — how does the ligand pose in it? (generates candidate poses)
- MM-GBSA — which pose is real? (physics-based binding-energy rescore)
Pick: pockets only → stop at P2Rank. Poses → DiffDock. Confident ranking → add MM-GBSA. Ask for the whole chain in one request and the orchestrator runs the steps in order.
ADMET and safety
Predict ADMET (absorption, distribution, metabolism, excretion, toxicity)
- ADMET-AI — a broad panel, fast. Best for early screening and library triage.
- DeepPurpose — narrower: predicts drug–target interaction scores for the targets you supply — reach for it for selectivity and off-target questions.
- RDKit — descriptors and drug-likeness filters only (Lipinski, PAINS) — fast, deterministic, no ML.
Pick: broad screen → ADMET-AI. Off-target panel → DeepPurpose. Just drug-likeness flags → RDKit.
Predict solubility
Predict toxicity
- ADMET-AI covers DILI, hERG, AMES, ClinTox, Tox21 and CYP inhibition;
safety_classifycovers structural alerts. Acute-toxicity class and organ toxicity are no longer predicted by any tool.
Chemistry
Search for compounds
- ChEMBL — bioactivity and medicinal-chemistry data; start here when you know the target.
- PubChem — search by structure, similarity, or property.
- ZINC — purchasable compounds and their vendor catalog codes; reach for it when procurement matters.
Pick: know the target → ChEMBL. Search by structure → PubChem. Must be buyable → ZINC.
Design new molecules
- GenMol — the preferred small-molecule generator.
- REINVENT4 — reinforcement-learning generation toward a target profile.
Plan a synthesis and source building blocks
- AiZynthFinder — retrosynthetic routes; then ZINC to check which starting materials are purchasable.
Genomics and omics
Analyze single-cell data
- Scanpy — classical, fast, interpretable clustering and differential expression.
- scvi-tools — deep generative models; reach for it for batch correction and unknown cell types.
Pick: speed and interpretability → Scanpy. Batch effects / novel cell types → scvi-tools.
Go from disease to target
- Genomics and causal genomics form the expression → differential → causal → druggable-target pipeline.
Literature
Mine papers and entities
- PubMed and BioBERT — search literature and extract compound–target–disease relationships. For a packaged profile, generate a Literature passport.
Two shortcuts
For a quick goal-to-tool lookup, use Which tool?. For a subject profiled across many of these capabilities at once, generate a passport.
Was this page helpful?