A comprehensive comparison of molecular feature representations for use in predictive modeling

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Machine learning accelerates material and drug design by predicting molecular properties.
• Effective molecular representation is crucial for machine learning model performance.
• Numerous methods exist for calculating molecular features, including traditional fingerprints, descriptors, and learnable neural network representations.

Purpose of the Study:

• To comprehensively compare various molecular feature representations for machine learning.
• To evaluate the performance of different representations across diverse predictive tasks.
• To identify the most effective molecular representations for accelerating material and drug design.

Main Methods:

• Evaluation of traditional molecular features (fingerprints, molecular descriptors) and learnable representations (neural networks).
• Testing on 11 benchmark datasets for predicting properties like mutagenicity, melting point, activity, solubility, and IC50.
• Comparative analysis of performance across different feature types and datasets.

Main Results:

• Several molecular features demonstrate similar performance across datasets.
• Spectrophores showed significantly worse performance.
• PaDEL molecular descriptors excelled in predicting physical properties.
• MACCS fingerprints offered strong overall performance despite their simplicity.
• Learnable representations achieved competitive results but offered no significant advantage over expert-based methods.
• Task-specific representations (graph convolutions, Weave) provided minimal benefit despite higher computational cost.
• Combining different representations did not typically improve performance.

Conclusions:

• Traditional molecular representations remain highly effective for many machine learning tasks in material and drug design.
• Learnable representations show promise but do not consistently outperform established methods.
• The choice of molecular representation significantly impacts predictive accuracy, with no single method universally superior.
• Further research may focus on optimizing specific representations or exploring hybrid approaches.