HKD-CPI: high-order knowledge distillation enhanced inductive compound-protein interaction prediction

Area of Science:

• Computational Chemistry
• Bioinformatics
• Drug Discovery

Background:

• Accurate compound-protein interaction (CPI) identification is crucial for drug discovery.
• Current deep learning methods often neglect high-order interaction patterns in molecules.

Purpose of the Study:

• To propose HKD-CPI, a novel framework for enhanced inductive compound-protein interaction prediction.
• To improve generalization capabilities for predicting interactions with unseen compound-protein pairs.

Main Methods:

• HKD-CPI integrates molecular graph features with large language model (LLM) embeddings via molecular graph tokenization.
• A hypergraph representation models high-order relationships between feature-similar compound/protein groups.
• Knowledge distillation transfers high-order interaction knowledge to a lightweight model for efficient prediction.

Main Results:

• HKD-CPI demonstrates superior performance in inductive CPI prediction tasks compared to state-of-the-art methods.
• Achieved an average improvement of 4.94% in AUROC and 4.90% in AUPRC across five benchmark datasets.
• The framework effectively infuses sequence-derived semantics into structural representations.

Conclusions:

• HKD-CPI offers a robust and efficient approach for predicting compound-protein interactions.
• The proposed methods enhance the understanding and prediction of complex molecular interactions.
• This framework holds significant potential for accelerating drug discovery pipelines.