awslabs Cheminformatics (RDKit) — Pipeline Skill (Thin Scaffold)
Overview
Adds fingerprint selection logic, filter cascade ordering, and version-specific gotchas for RDKit workflows. The agent already knows RDKit syntax — this skill prevents parameter mismatches and deprecated API usage.
Usage
- Activate when filtering compound libraries (Lipinski/Veber/PAINS)
- Activate for fingerprint similarity search or clustering
- Activate for matched molecular pair (MMP) analysis with mmpdb
Core Concepts
Filter cascade order: Lipinski → Veber → PAINS → similarity (cheapest first, most expensive last).
Fingerprint API: Use rdFingerprintGenerator module exclusively (since RDKit 2024.03). The AllChem.GetMorganFingerprintAsBitVect() path is deprecated and will be removed.
SMILES validation: Always parse with sanitize=False then explicit SanitizeMol() in try/except. Default MolFromSmiles silently returns None on partial failures.
Decision Logic
Fingerprint Selection
| Goal | Fingerprint | Parameters | Tanimoto threshold |
|---|---|---|---|
| General similarity | Morgan (ECFP4) | radius=2, fpSize=2048 | ≥ 0.7 |
| High-specificity matching | Morgan (ECFP6) | radius=3, fpSize=4096 | ≥ 0.6 |
| Substructure screening | MACCS 166 keys | — | ≥ 0.8 |
| Pharmacophore similarity | FCFP | radius=2, GetMorganFeatureAtomInvGen() |
≥ 0.6 |
| Large-scale clustering | Morgan | radius=2, fpSize=1024 | Speed over precision |
Filter Thresholds
| Filter | Rule | Threshold | Notes |
|---|---|---|---|
| Lipinski MW | ≤ 500 Da | Allow ≤1 violation total | Not binary pass/fail |
| Lipinski LogP | ≤ 5 | — | — |
| Lipinski HBD | ≤ 5 | — | — |
| Lipinski HBA | ≤ 10 | — | — |
| Veber TPSA | ≤ 140 Ų | Hard cutoff | — |
| Veber RotBonds | ≤ 10 | Hard cutoff | — |
| PAINS | A+B+C catalogs | HTS triage only | Don't apply to advanced compounds |
MMP Analysis Decision
| Scenario | Tool | Key parameter |
|---|---|---|
| SAR from compound pairs | mmpdb fragment → index → transform | --min-pairs 3 |
| Property prediction | mmpdb predict | Requires indexed DB with property data |
| Core identification | rdFMCS.FindMCS() |
ringMatchesRingOnly=True, completeRingsOnly=True |
| Large libraries (>10K) | mmpdb CLI | --max-variable-size 10 to limit fragmentation |
Critical Parameters
| Parameter | Value | When to change |
|---|---|---|
| Morgan radius | 2 (ECFP4) | 3 for higher specificity |
| fpSize | 2048 | 1024 for speed; 4096 for precision |
| includeChirality | False | True for stereo-sensitive applications |
| Lipinski max_violations | 1 | 0 for strict; exempt biologics entirely |
| mmpdb --min-pairs | 3 | Lower for sparse datasets |
| mmpdb --max-variable-size | 10 | Increase for macrocycle transformations |
| PAINS catalogs | A+B+C (all three) | Never use partial sets |
Common Mistakes
Wrong:
AllChem.GetMorganFingerprintAsBitVect()→ Right:rdFingerprintGenerator.GetMorganGenerator(radius=2, fpSize=2048).GetFingerprint(mol)→ Why: AllChem FP functions deprecated since RDKit 2024.03, will be removedWrong: Default
MolFromSmiles(smi)without error handling → Right:MolFromSmiles(smi, sanitize=False)+SanitizeMol()in try/except → Why: Default silently returns None; explicit sanitize catches partial failuresWrong: Lipinski as hard binary filter → Right: Allow ≤1 violation → Why: Many approved drugs violate one rule; binary filtering rejects valid candidates
Wrong: PAINS on clinical candidates → Right: PAINS only for HTS library triage → Why: Designed for screening artifacts; flags valid advanced compounds
Wrong: Morgan radius=1 → Right: radius=2 minimum (ECFP4) → Why: Radius 1 loses structural information; too many false-positive similarities
Wrong: Comparing FPs with different radius or nBits → Right: Identical parameters for both FPs → Why: Tanimoto only meaningful in same parameter space
Wrong: Ignoring stereochemistry for chiral compounds → Right:
includeChirality=Truein generator → Why: Default treats enantiomers as identicalWrong: Fixed 0.7 threshold for all FP types → Right: Validate threshold per fingerprint type and chemical series → Why: 0.7 convention is ECFP4-specific; other types need different cutoffs
Wrong: Raw SMILES into mmpdb → Right: Pre-canonicalize with
MolToSmiles(MolFromSmiles(smi))→ Why: mmpdb expects canonical SMILES; non-canonical causes fragmentation errorsWrong: Confusing LogP with LogD → Right: Use LogD at pH 7.4 for ionizable compounds; RDKit computes LogP only → Why: LogP overestimates lipophilicity for ionizable molecules at physiological pH
Response Format
- Lead with the command or code the user needs — explain after
- Structure as: confirm inputs → working code → key parameters explained → gotchas
- One complete working example per task; do not show every alternative
- Keep code comments minimal and functional
- Target: 50-100 lines of code with brief surrounding explanation