Abstract
The success of machine learning in chemistry is fundamentally underpinned by the information fidelity of molecular representations. Despite their widespread adoption for efficiency and interpretability, Morgan fingerprints harbor a long-overlooked and fundamental flaw: bit collisions. This phenomenon erroneously maps distinct chemical substructures to identical positions, systematically corrupting structure-property relationships and severely compromising model interpretability. To address this challenge, we introduce Collision-Free Morgan Fingerprints (CF-MF), a principled framework that guarantees the integrity of substructure information through an adaptive, data-driven sizing mechanism. Through a comprehensive evaluation across 25 diverse datasets (> 50,000 molecules) and multiple machine learning paradigms, we demonstrate that CF-MF delivers consistent and significant performance gains up to 16.81% RMSE reduction in regression and 11.1% accuracy increase in classification. More critically, by eliminating attribution errors caused by collisions, CF-MF fundamentally restores chemical interpretability and expands the reliable prediction domains of models by 60-100%. Our information-theoretic analysis reveals a strong correlation between collision-induced entropy loss and performance degradation (R2 = 0.854, p < 0.001), establishing information fidelity as a fundamental design principle for next-generation molecular representations. It also achieves performance competitive with state-of-the-art deep learning models while retaining the simplicity and intuitiveness of traditional fingerprints. This work provides a more reliable and trustworthy foundation for AI-driven drug discovery, materials science, and environmental assessment.Scientific contributionWhile bit collisions in Morgan fingerprints have been acknowledged for decades, this study is the first to systematically quantify their impact on machine learning performance and provide a principled, reproducible solution applicable to any molecular dataset. We establish a novel information-theoretic framework that directly links collision-induced entropy loss to predictive degradation, offering the field a quantitative criterion for evaluating molecular representation fidelity. Beyond performance gains, our work uniquely demonstrates that eliminating collisions restores chemically valid SHAP attributions-addressing a critical but previously unrecognized barrier to trustworthy AI interpretation in chemistry.