Clever Hans in Chemistry: Chemist Style Signals Confound Activity Prediction on Public Benchmarks

2025-12-23

AI Summary

This work reveals a “Clever Hans” effect in chemical machine learning models for activity prediction—models inadvertently rely on chemist-specific stylistic cues (e.g., synthetic preferences, author identity) rather than genuine structure–activity causal relationships. Method: We systematically demonstrate that molecular fingerprints enable high-accuracy author attribution (top-5 accuracy: 60%), and show that activity predictors trained solely on author-probability vectors match the performance of structure-based baselines—confirming severe causal leakage. To mitigate bias, we propose author-disjoint data splits, scaffold-aware partitioning, and explicit author-probability modeling. Contribution/Results: Our findings motivate the “de-intentionalization” of benchmark datasets in cheminformatics—removing confounding author-level signals—and establish theoretical foundations and methodological guidelines for building causally robust models in drug discovery. These practices enhance model generalizability and scientific interpretability by enforcing structural causality over spurious correlations.

Technology Category

• Machine Learning: Causal Learning
• Reasoning under Uncertainty: Causality
• Knowledge Representation and Reasoning: Action, Change, and Causality

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Web Mining and Content Analysis: Large pretrained models with web data
• Semantics and Knowledge: Methods, algorithms and applications for the development of semantic models, knowledge graphs and other forms of structured data models with machine-interpretable semantics

Abstract

Can machine learning models identify which chemist made a molecule from structure alone? If so, models trained on literature data may exploit chemist intent rather than learning causal structure-activity relationships. We test this by linking CHEMBL assays to publication authors and training a 1,815-class classifier to predict authors from molecular fingerprints, achieving 60% top-5 accuracy under scaffold-based splitting. We then train an activity model that receives only a protein identifier and an author-probability vector derived from structure, with no direct access to molecular descriptors. This author-only model achieves predictive power comparable to a simple baseline that has access to structure. This reveals a "Clever Hans" failure mode: models can predict bioactivity largely by inferring chemist goals and favorite targets without requiring a lab-independent understanding of chemistry. We analyze the sources of this leakage, propose author-disjoint splits, and recommend dataset practices to decouple chemist intent from biological outcomes.

Problem

• Models exploit chemist identity to predict bioactivity
• Author-only models match structure-based baseline performance
• Leakage of chemist intent confounds structure-activity relationships

Innovation

• Predicting authors from molecular fingerprints using classifier
• Training activity model with author-probability vector and protein identifier
• Proposing author-disjoint splits to decouple chemist intent