Small Molecule Design Tools

Small-Molecule Design Tools

This skill routes small-molecule design tasks to suitable open or publicly documented tools. It directs the agent to a focused reference and keeps the applicable license and data terms visible during tool selection.

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 and validators such as AiZynthFinder, RENKIN, 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, including RetroAgent.
• 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 locally runnable 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 terms-review checklist.
• references/worked-example-kras-glue.md: public worked example on daraxonrasib, KRAS, and CypA using PDB 9BG6.
• references/watchlist.md: tools with missing code, missing weights, unclear license terms, or early-stage releases.

Routing Rules

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

First Picks

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

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. For an LLM fine-tune, record both the project terms and the original base-model 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 model the complete mechanism. A single-pocket docking workflow omits important components of these systems.
• When an experimental ternary structure exists, perturb and score that structure before asking a co-folding model to predict it again.
• Pair pocket-conditioned generators with a makeability step. Most structure- based generators do not guarantee a synthesis route.
• Re-check upstream repositories for pre-1.0 models and entries with recent verification dates.

Pipeline Pattern

A full design loop 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 deployment.

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