Small molecules

Small-Molecule Design Tools

This is an agent skill for routing small-molecule design tasks to suitable open tools. It is meant to improve agent performance on complex workflows where the agent must choose methods, load focused references, and keep licensing or data constraints in view while it works.

Use the skill in two passes:

1. Identify the physical or chemical task.
2. Open the matching reference file before recommending tools or writing code.

For commercial or product-facing work, also open references/licensing-and-data.md. Code, weights, data, and base-model terms often differ for the same upstream project.

Reference Map

Start with references/tool-matrix.md when the user wants a broad comparison across categories.

Makeability and synthesis:

• references/synthesizable-generation.md: makeable molecule generation, analogs, make-on-demand libraries, and route co-generation.
• references/molecular-generation.md: general molecule generators where validity matters more than synthesis route guarantees.
• references/retrosynthesis-planning.md: multi-step route planners such as AiZynthFinder, ASKCOS, Syntheseus, and SynPlanner.
• references/singlestep-retrosynthesis.md: one-step precursor prediction with ReactionT5v2, RXNGraphormer, GDiffRetro, RetroDiT, ConRetroBert, TempRe, and related models.
• references/agentic-retrosynthesis.md: LLM and agent-assisted synthesis planners.
• references/forward-and-reaction-modeling.md: forward reaction and mechanism prediction.
• references/template-and-rule-infrastructure.md: reaction templates, SMARTS rules, RDChiral, rxnutils, and SynTemp.
• references/synthesizability-scoring.md: makeability scoring and route consistency checks.

Target-based design:

• references/protein-structure-prediction.md: receptor structure prediction and preparation.
• references/docking-and-cofolding.md: classical docking, learned docking, protein-ligand co-folding, and pose scoring.
• references/binding-affinity-and-fep.md: ML affinity, OpenFE, OpenMM, endpoint methods, and free-energy workflows.
• references/property-and-qsar-prediction.md: QSAR, activity, property, and active-learning screens.
• references/admet-prediction.md: ADMET, hERG, BBB, CYP, and deployable local predictors.
• references/target-and-selectivity-prediction.md: off-target, selectivity, shape, and pharmacophore tools.
• references/structure-based-generation.md: pocket-conditioned generation and pocket detection.

Cross-cutting:

• references/agentic-drug-design.md: agent frameworks that orchestrate generation, scoring, ADMET, and affinity tools.
• references/chemical-language-models.md: chemistry language models and multimodal chemistry models.
• references/licensing-and-data.md: license layers, Enamine terms, PDB/PDBBind notes, ChEMBL/TDC data terms, and commercial-use checklist.
• references/worked-example-kras-glue.md: public worked example on daraxonrasib, KRAS, and CypA using PDB 9BG6.
• references/watchlist.md: promising tools with missing code, missing weights, missing license clarity, or immature releases.

Routing Rules

Use the user's actual task to choose the first reference:

User task	Start with
Generate makeable analogs of a hit	synthesizable-generation.md
Project a molecule into synthesizable space	synthesizable-generation.md
Generate valid molecules for exploration	molecular-generation.md
Plan a route to a known target molecule	retrosynthesis-planning.md
Predict one reaction step	singlestep-retrosynthesis.md
Check whether a proposed reaction route is plausible	forward-and-reaction-modeling.md, then synthesizability-scoring.md
Prepare or predict a receptor structure	protein-structure-prediction.md
Dock a ligand into a known structure	docking-and-cofolding.md
Co-fold a protein-ligand complex from sequence and ligand	docking-and-cofolding.md
Estimate binding affinity	binding-affinity-and-fep.md
Build a QSAR or property model	property-and-qsar-prediction.md
Screen ADMET risk	admet-prediction.md
Screen off-target or selectivity risk	target-and-selectivity-prediction.md
Generate molecules into a pocket	structure-based-generation.md
Orchestrate a full agentic loop	agentic-drug-design.md, then the category references it calls
Check whether a tool is usable in commercial work	licensing-and-data.md, then the tool card

First Picks

Use these as starting points, then read the relevant reference file for details:

Need	First pick
Synthesizable-space projection	PrexSyn
Training-free analog generation	SynTwins, with license review before reuse
3D molecule and route co-generation	SynCoGen
Make-on-demand design over Enamine-style libraries	SyntheMol or APEX, depending on license needs
General de novo molecule generation	GenMol, with SmiSelf for validity repair
Multi-step route planning	AiZynthFinder first, ASKCOS when broader planning and conditions are needed
One-step retrosynthesis	ReactionT5v2
LLM-assisted route planning	DeepRetro or Synthelite, followed by round-trip checks
Reaction templates and data cleanup	RDChiral or rdchiral_plus with rxnutils
Commercially clean ML affinity starting point	Boltz-2
Protein-ligand co-folding	Boltz-1, Chai-1, or Umol after checking current weight terms
Learned docking into a known pocket	DiffDock-L or Uni-Mol Docking v2
Classical CPU docking	AutoDock Vina
Physics-based relative free energy	OpenFE
QSAR from ChEMBL-style data	Chemprop or DeepChem
Local ADMET triage	ADMET-AI
Broad off-target scan	ChEMBL multitask model plus ESP-Sim where shape overlap matters
Pocket-conditioned generation	REINVENT 4, DiffSBDD, PILOT, Pocket2Mol, or TargetDiff, followed by makeability checks

Operating Notes

• Read source licenses directly. README badges and paper text can disagree with the repository license, model-card license, or dataset terms.
• Check code, weights, data, and base-model terms separately. LLM fine-tunes add the original model terms to the project terms.
• Treat docking and co-folding confidence scores as pose-confidence signals. Binding-affinity estimates require an affinity model or a free-energy method.
• For molecular glues, degraders, covalent binders, and macrocycles, choose tools that match the mechanism. A single-pocket docking workflow is often the wrong starting point for those systems.
• When an experimental ternary structure exists, consider perturbation and scoring on that scaffold before asking a co-folding model to rediscover it.
• Pair pocket-conditioned generators with a makeability step. Most structure- based generators do not guarantee a synthesis route.
• Re-check live upstream repositories for recent tools, especially pre-1.0 models and 2026 additions.

Pipeline Pattern

A full agent loop usually crosses both layers:

1. Get or predict the receptor structure.
2. Generate, dock, co-fold, or score candidate molecules for the target.
3. Filter for affinity, QSAR, ADMET, and selectivity.
4. Project survivors into synthesizable space.
5. Plan or validate routes.
6. Re-check licenses and data terms before claiming deployability.

Load the detailed reference file for install commands, model weights, benchmark context, GPU needs, and current caveats.