An Explainable Molecular Token Estimation Method for Knowledge-Aware Drug-Drug Interaction Prediction.

Area of Science:

• Computational chemistry
• Machine learning
• Bioinformatics

Background:

• Molecular representation learning (MRL) uses tokens like atoms and motifs to represent molecules for Graph Neural Networks (GNNs).
• While GNNs with tokens show promise in drug-drug interaction (DDI) prediction, the impact of token choice on molecular embedding expressiveness is not well understood.

Purpose of the Study:

• To theoretically define MRL from a frequency domain perspective and analyze the influence of token count on model performance.
• To propose SimMotifPro, an efficient motif-based method for DDI prediction that incorporates theoretical insights.

Main Methods:

• Developed an axiomatic definition of MRL from a frequency domain perspective to establish theoretical upper bounds for model convergence.
• Proposed SimMotifPro, utilizing a DeeperGCN encoder, a motif-motif knowledge graph, and a Motif Ranker module for enhanced DDI prediction.

Main Results:

• Demonstrated that SimMotifPro aligns with the derived theoretical upper bounds, validating the proposed theory's general applicability.
• Achieved state-of-the-art performance on multiple benchmarks for DDI prediction, showcasing the efficacy of the motif-based approach.

Conclusions:

• The number of tokens significantly impacts MRL model performance, with theoretical insights guiding the development of effective molecular embedding strategies.
• SimMotifPro offers a robust and efficient solution for DDI prediction, highlighting the importance of motif-based representations and advanced GNN architectures.