Scaffold Analysis Usage Guide

Overview

Analyze chemical libraries by underlying scaffolds (Bemis-Murcko, generic frameworks), cluster compounds by chemotype, perform R-group decomposition, run matched molecular pair analysis (mmpdb), and produce scaffold-balanced QSAR train/test splits to prevent data leakage.

Prerequisites

pip install rdkit mmpdb scikit-learn datamol

Quick Start

Tell the AI agent what to do: - "Compute Bemis-Murcko scaffolds for my compound library and report cluster sizes" - "Generate a scaffold-balanced 80/20 train/test split for QSAR" - "Run MMPA on the bioassay data to find activity-improving transformations" - "Decompose this 50-compound series into R-groups around the central scaffold" - "Identify analog series of size >= 5 in this library"

Example Prompts

Library chemotype analysis

"Compute Bemis-Murcko scaffolds for all 10k compounds in library.csv. Output a table of scaffold SMILES with member count. Report the top 20 most-populated scaffolds."

Scaffold-balanced ML split

"Apply Bemis-Murcko scaffold split (80% train, 20% test) to qsar_data.csv. Ensure no scaffold appears in both train and test. Output train.csv and test.csv."

MMPA mining

"Run mmpdb on hERG_data.csv. Report pair counts, contexts, effect estimates, and uncertainty without applying universal reliability thresholds."

R-group decomposition

"Given scaffold 'c1ccc(-[:1])cc1-[:2]' and 50 analogs, decompose into R-group table. Output CSV with columns R1, R2, activity for downstream Free-Wilson analysis."

What the Agent Will Do

1. Parse SMILES library; standardize where needed.
2. Compute Bemis-Murcko scaffolds (and generic framework if requested).
3. Group compounds by scaffold; report chemotype distribution.
4. For ML splits: assign whole scaffolds to train or test; output partitioned data.
5. For MMPA: fragment compounds, index by mmpdb, run transform queries.
6. For R-decomposition: align compounds to scaffold, extract R-group SMILES.

Tips

• Bemis-Murcko gives empty scaffold for linear molecules; supplement with linear features.
• Generic framework loses heteroatom info; use only for topology comparison.
• MMPA evidence depends on matched-pair count, independence, context diversity, and validation, not a universal dataset-size cutoff.
• Choose scaffold, time, random, or other splits to match deployment; always audit overlap and label balance.
• Preserve original row identifiers and unmatched indices during R-group decomposition.

Related Skills

• chemoinformatics/molecular-io - Parse compounds
• chemoinformatics/molecular-standardization - Standardize before scaffold extraction
• chemoinformatics/reaction-enumeration - Free-Wilson on R-decomp output
• chemoinformatics/similarity-searching - 2D scaffold-hopping
• chemoinformatics/qsar-modeling - Scaffold-aware splitting for QSAR
• chemoinformatics/generative-design - Scaffold-decoration generation