mol-filters

mol-filters — SA_Score + medicinal-chemistry structural alerts

Fills a real gap: the rdkit MCP server exposes descriptors/fingerprints but NOT Ertl's synthetic-accessibility score or RDKit's FilterCatalog (PAINS/BRENK/NIH alerts). This skill adds both as a small CLI, so you can triage screening/docking hits without a Schrodinger license.

Run

Use the interpreter that has RDKit — the rdkit MCP server's venv:

"C:\Users\Prekzursil\source\repos\rdkit-mcp-server\.venv\Scripts\python.exe" \
  C:\Users\Prekzursil\.agents\skills\mol-filters\scripts\mol_filters.py <INPUT> [--json]

Input modes:

• single SMILES: ... mol_filters.py "CC(=O)Oc1ccccc1C(=O)O"
• one-per-line: ... mol_filters.py --smiles-file mols.smi (blank / # lines skipped; <SMILES> <name> ok)
• SDF library: ... mol_filters.py --sdf library.sdf
• add --json for machine-readable output.

What it reports (per molecule)

1. SA_Score — Ertl & Schuffenhauer (2009) synthetic accessibility, 1 = easy … 10 = hard to synthesize. Loaded from RDKit's Contrib tree; if a wheel omits Contrib the score degrades to null with a note and the alert filters still run.
2. Structural-alert hits across PAINS_A/B/C, BRENK, NIH, ZINC — each hit names the catalog and the matched alert (e.g. PAINS_B: rhod_sat_A(33)). Zero hits = "Alerts: none".

Invalid SMILES are flagged [INVALID] and skipped, never crash the run.

Acceptance (verified 2026-07-04, rdkit 2026.03.3)

• "CC(=O)Oc1ccccc1C(=O)O" (aspirin) → SA_Score ≈ 1.58, no alerts.
• "O=C1CSC(=S)N1" (rhodanine) → SA_Score ≈ 3.46, alerts: PAINS_B: rhod_sat_A(33), Brenk: Thiocarbonyl_group. A correct run shows a low SA_Score for a simple drug and fires ≥1 alert on a known PAINS/BRENK scaffold.

Notes

• SA_Score is NOT importable by default; the script locates sascorer via RDConfig.RDContribDir/SA_Score and caches it. A high SA_Score does not mean "bad" — read it alongside the alerts (a clean, easy-to-make hit is ideal).
• For descriptors (TPSA, logP, HBD/HBA, rings) use the rdkit MCP or the cheminformatics-rdkit skill; for docking/ADMET use structure-based-screening / schrodinger-maestro-driver.