Enhancing Molecular Contrastive Prediction by Recognizing Metric Distribution Separation

Molecular machine learning has attracted much attention in molecular property prediction and AI-aided Drug Discovery. Recent advancements have significantly improved the efficiency and speed of searching for potential drug candidates. The primary challenge in molecular property prediction is how to transfer potential knowledge from known molecules to a large set of unseen candidates. To address this challenge, this work explores a metric distribution contrastive prediction method. This method establishes metric distributions among ordered categorical molecule populations, ensuring that all samples within the same class, whether known or unknown, share a common metric distribution. This technique comprises constructing metric distributions, extracting class prototypes as the foundational representations for each class, and employing a Bhattacharyya distribution transformation distance metric to distinguish pairwise metric distributions. Notably, this approach minimizes the distances between intra-class samples while increasing the distances between inter-class samples. This method integrates pre-training and subsequent prediction phases into a unified neural computation framework. Through thorough experiments on various diverse benchmark molecular datasets, the effectiveness of this method is validated, yielding notable improvements in prediction accuracy compared to existing models.