QSAR Modeling Usage Guide

Overview

Build target-specific QSAR / QSPR models from in-house assay data by comparing chemprop D-MPNN, RandomForest + ECFP4, and, when justified, transformer-based MolFormer / Uni-Mol / ChemBERTa under the same validation design. Apply OECD 5 principles with deployment-relevant splits, ensemble uncertainty, applicability-domain assessment, and conformal prediction where its assumptions and interface are satisfied.

Prerequisites

pip install chemprop rdkit scikit-learn mapie shap

These examples target chemprop 2.2.x (note the major API change from 1.x).

Quick Start

Tell the AI agent what to do: - "Train chemprop D-MPNN classifier on this hERG dataset with scaffold split" - "Build RandomForest QSAR on 150 compounds (small data); use ECFP4 + nested CV" - "Add conformal prediction intervals to my chemprop classifier" - "Apply applicability domain filter (kNN distance) to new predictions" - "Compute SHAP atomic attribution to interpret hERG predictions"

Example Prompts

Small-data baseline

"Build RF + ECFP4 binary classifier on 150-compound dataset (hERG_blocker column). 5-fold scaffold-split cross-validation. Report AUC, sensitivity, specificity, applicability domain by kNN Tanimoto."

chemprop classifier

"Train chemprop 2.x D-MPNN ensembles on hERG_data.csv. Use a scaffold-balanced split, five replicates, and five ensemble members per replicate. Include current rdkit_2d features. Report test AUC and calibration, keeping replicate and ensemble aggregation explicit."

Conformal prediction

"Use MAPIE with my sklearn QSAR baseline to target 90% marginal coverage, or implement and test an sklearn-compatible wrapper before applying it to chemprop. State the exchangeability assumptions."

Multi-task CYP inhibition

"Train chemprop multitask classifier on CYP1A2, CYP2C9, CYP2C19, CYP2D6, CYP3A4. Scaffold split. Compare to per-target single-task models."

What the Agent Will Do

1. Standardize input data (recommend chemoinformatics/molecular-standardization).
2. Choose and document a split that represents deployment (for example scaffold, time, external-series, or prospective).
3. Featurize with ECFP4 + RDKit 2D descriptors (for chemprop hybrid).
4. Train replicated ensembles for chemprop or bootstrap models for RF, keeping validation partitions independent of model selection.
5. Compute test metrics + calibration plots.
6. Build applicability domain assessment (kNN Tanimoto distance or conformal prediction).
7. Optionally compute SHAP feature importance.

Tips

• Use scaffold splitting when transfer to unseen scaffold groups is the question; use time, external, or prospective splits when they better match deployment. Label random splits as interpolation-focused.
• Always benchmark a fingerprint baseline and a graph model on the same deployment-relevant split; add pretrained transformers when their representation and compute cost are justified.
• Check calibration on held-out data; apply Platt or isotonic calibration only with a proper calibration split.
• Conformal prediction targets marginal coverage under exchangeability; verify empirical coverage on relevant held-out data.
• Use ensemble variance as one uncertainty diagnostic, not as a complete applicability-domain definition.

Related Skills

• chemoinformatics/molecular-descriptors - Featurization choices
• chemoinformatics/molecular-standardization - Mandatory upstream
• chemoinformatics/scaffold-analysis - Scaffold split implementation
• chemoinformatics/admet-prediction - ADMET-specific QSAR
• chemoinformatics/generative-design - QSAR as scoring component
• machine-learning/model-validation - General ML validation
• machine-learning/biomarker-discovery - Adjacent ML approaches