Sort & Slice: a simple and superior alternative to hash-based folding for extended-connectivity fingerprints

Area of Science:

• Cheminformatics
• Molecular Machine Learning
• Computational Chemistry

Background:

• Extended-connectivity fingerprints (ECFPs) are widely used for molecular representation in cheminformatics and machine learning.
• Current methods for converting ECFP substructures into bit vectors, like hash-based folding, can lead to information loss and bit collisions.
• Graph neural networks offer advanced atom feature aggregation, but ECFP vectorization remains a bottleneck.

Purpose of the Study:

• To introduce a general mathematical framework for structural fingerprint vectorization, termed substructure pooling.
• To present Sort & Slice, a novel, bit-collision-free alternative to hash-based folding for ECFP pooling.
• To computationally evaluate Sort & Slice against existing methods for molecular property prediction.

Main Methods:

• Developed a general framework for substructure pooling, encompassing various vectorization techniques.
• Implemented Sort & Slice, which ranks substructures by prevalence and selects the top L features.
• Compared Sort & Slice with hash-based folding, filtering, and mutual-information maximization using ECFP-based molecular property prediction tasks.

Main Results:

• Sort & Slice demonstrated robust and substantial performance improvements over hash-based folding across diverse prediction tasks.
• The proposed method consistently outperformed other investigated substructure-pooling techniques.
• Performance gains were observed across different machine learning models, data splitting strategies, and ECFP hyperparameters.

Conclusions:

• Sort & Slice is a simple yet highly effective method for vectorizing ECFPs, offering a superior alternative to hash-based folding.
• The substructure pooling framework provides a generalized approach to structural fingerprint vectorization.
• Sort & Slice is recommended as the default method for vectorizing ECFPs in supervised molecular machine learning to enhance predictive accuracy.