Enumerate analogs around a lead compound for SAR expansion

Hand Claude Code a lead SMILES; get back a deduplicated, standardized set of close-in analogs — tautomer and stereoisomer variants plus fragment-substituted neighbors — each scored for drug-likeness and similarity to the parent, ready for the medicinal chemist to triage into the next synthesis round.

   
Problem class Hypothesis generation
Subject areas Chemistry, Drug Repurposing and Discovery
Evidence level Proposed
Complexity One skill or MCP
Availability Fully open
Compute Laptop

Problem

Once a project has a confirmed hit or a lead, the recurring question is what should we make next? The medicinal chemist wants a structured set of close neighbors — the same scaffold with single-point modifications (a methyl walk, a halogen scan, a bioisosteric ring swap), plus the tautomers and stereoisomers that a single drawn SMILES silently collapses. Doing this by hand is slow and inconsistent: the enumeration order, the standardization rules, and the similarity cut-offs vary run to run, so two chemists expanding the same lead produce different sets. The output also has to be clean — salts stripped, charges neutralized, duplicates removed — before any of it is worth scoring or ordering.

Solved looks like: one lead SMILES in, a deduplicated table of candidate analogs out, each row carrying its Tanimoto similarity to the parent, a QED drug-likeness score, and the transformation that produced it, in a few minutes on a laptop. This is a hypothesis-generation step — the table is a menu of testable structures, not a ranked list of predicted winners.

  1. Install the Datamol skill — it bundles the standardization, enumeration, and featurization recipes:

    /plugin marketplace add K-Dense-AI/claude-scientific-skills
/plugin install datamol@claude-scientific-skills

    Confirm with /plugin list. The RDKit skill is implicitly available (Datamol is RDKit-built); install it explicitly only if you want raw reaction-SMARTS enumeration beyond what Datamol exposes.

  2. Standardize and expand the stereo/tautomer space first. A minimal prompt:

    Lead: CC(=O)Oc1ccccc1C(=O)O  (or paste your SMILES)

Use the Datamol skill:
  1. Standardize the lead (dm.fix_mol, dm.sanitize_mol,
     dm.standardize_smiles); keep the largest fragment and
     neutralize charges. Print the standardized canonical SMILES.
  2. Enumerate tautomers (dm.enumerate_tautomers) and
     stereoisomers (dm.enumerate_stereoisomers), capping at
     32 forms each. Deduplicate by InChIKey.
Report the parent plus each enumerated form as a table:
smiles | inchikey | relation (tautomer/stereoisomer) | n_atoms.

  3. Generate fragment-substituted neighbors. Add a follow-up that walks single-point modifications around the core:

    Now generate close analogs of the standardized lead:
  - For each aromatic H position, enumerate substitution with a
    small fragment set: -F, -Cl, -CH3, -OCH3, -CF3, -CN, -OH.
  - Use RDKit reaction SMARTS or dm scaffold tools; standardize
    each product and drop anything that fails to parse.
  - Deduplicate the full set by InChIKey against the parent and
    against each other.

    Keep the fragment set small and chemically sensible for the chemotype — a kinase hinge-binder warrants a different substituent palette than a GPCR amine. The point is breadth of single-point ideas, not exhaustive combinatorics.

  4. Score and rank the analog set. Drive Datamol’s featurization plus a drug-likeness pass:

    For every analog (and the parent), compute with the Datamol skill:
  - ECFP4 (Morgan radius 2, 2048-bit) fingerprint
  - Tanimoto similarity to the standardized parent
  - QED drug-likeness score
  - molecular weight, logP, HBD, HBA, rotatable bonds
Write analogs/<lead>_expansion.csv with columns:
smiles | inchikey | transformation | tanimoto_to_parent | qed |
mw | logp | hbd | hba | rotb.
Sort by tanimoto_to_parent descending, then qed descending.
Print the top 20 rows and the total count.

  5. Hand off to filtering and bioactivity. The enumeration deliberately stops at “plausible neighbors” — it does not judge developability. Pipe analogs/<lead>_expansion.csv straight into the Filter a virtual screening hit list recipe (Lipinski / Veber / PAINS / BRENK cascade via the MedChem skill) to strip non-drug-like and alert-bearing structures, then look up any analog that already exists in ChEMBL with the Profile a compound’s polypharmacology recipe before committing synthesis effort.

Why this assembly

Rung 2 of the simplicity ladder. The whole problem — standardize, enumerate tautomers/stereoisomers, walk single-point substitutions, featurize, and score similarity — is exactly the surface the Datamol skill bundles (its catalog page lists “analog generation in lead optimisation” and “similarity searching” as primary use cases, with parallel batch operations on top of RDKit). Plain Claude Code with raw RDKit (rung 1) can do it, but every session re-derives the standardization order and the enumeration caps, and the InChIKey-deduplication step is easy to forget — which is precisely what makes hand-rolled analog sets irreproducible. A multi-tool harness (rung 3) adds nothing here: there is no second data source and no orchestration across heterogeneous APIs, only one cheminformatics library applied in sequence. Generative-model or autonomous-chemist escalation (rung 4) is reserved for de-novo scaffold design and synthesis planning — a different, larger problem than expanding the immediate neighborhood of a known lead.

Availability

Fully open. The Datamol and RDKit skills ship via the K-Dense-AI/claude-scientific-skills marketplace under MIT (skills) / Apache-2.0 and BSD-3-Clause (underlying libraries). No subscription, no institutional licence, no account beyond a Claude plan.

Compute requirements

Laptop. Enumerating tautomers, stereoisomers, and a single-point substitution scan for one lead produces hundreds to low-thousands of analogs and runs in well under a minute on a single CPU core. Fingerprinting and similarity scoring scale linearly; even a 50 000-analog combinatorial expansion fits in <2 GB RAM with Datamol parallelism (n_jobs=-1). No GPU.

Evidence

Proposed. No published benchmark of an LLM-driven Datamol analog-enumeration workflow is known. The closest documented analogue is the K-Dense lead-optimisation workflow, which positions Datamol (standardization, enumeration, featurization) immediately upstream of the MedChem filtering step this recipe hands off to — the same rdkit → datamol → medchem ordering. The underlying enumeration and similarity primitives are textbook medicinal cheminformatics: ECFP/Morgan fingerprints and Tanimoto similarity (Rogers & Hahn, J. Chem. Inf. Model. 50:742, 2010); the QED drug-likeness score (Bickerton et al., Nature Chemistry 4:90, 2012); and the matched-molecular-pair concept that motivates single-point SAR expansion (Griffen et al., J. Med. Chem. 54:7739, 2011). Each component has independent validation; the agent-orchestrated assembly does not. Treat the output as a hypothesis menu to be confirmed by filtering, bioactivity lookup, and ultimately synthesis and assay.

Alternatives considered

  • Plain Claude Code, no skill (rung 1). Fine for enumerating a handful of analogs you can eyeball, but Claude re-derives the standardization and enumeration boilerplate each session and the dedup-by-InChIKey step is easy to drop. Reach for it only for one-off, sub-dozen expansions.
  • RDKit skill alone. The RDKit skill gives reaction-SMARTS enumeration and fingerprints but not Datamol’s higher-level standardization and batch helpers. Use it when you need bespoke reaction transforms Datamol does not wrap; otherwise Datamol is the lower-friction rung.
  • A generative chemistry model or autonomous chemist (rung 4). De-novo generators (and synthesis-planning agents such as ChemCrow) propose new scaffolds and routes rather than close neighbors of a known lead. Reach for them when the project needs scaffold hopping or retrosynthesis, not single-point SAR expansion. They are heavier to set up and harder to audit per structure.

See also

Sources

Tried this recipe?

Share feedback — what worked, what didn’t, what you’d change. The form opens with this recipe pre-selected and a link back to this page.