TransformerCPI: improving compound-protein interaction prediction by sequence-based deep learning with self-attention...

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Predicting compound-protein interactions (CPI) is vital for drug discovery and chemogenomics.
• Many potential biological targets lack 3D structures, necessitating sequence-based CPI prediction methods.
• Existing sequence-based models often suffer from dataset issues, ligand bias, and improper data splitting, leading to overestimated performance.

Purpose of the Study:

• To address limitations in sequence-based CPI prediction.
• To develop a robust model for predicting CPI using only protein sequence information.
• To introduce rigorous validation methods to ensure models learn true interaction features.

Main Methods:

• Construction of novel, specific datasets for CPI prediction.
• Development of a novel transformer neural network architecture, TransformerCPI.
• Implementation of a rigorous label reversal experiment for model validation.

Main Results:

• TransformerCPI demonstrated significantly improved performance on newly designed experiments.
• The model effectively identifies important interacting regions in protein sequences and compound atoms.
• The label reversal experiment confirmed that TransformerCPI learns genuine interaction features.

Conclusions:

• TransformerCPI offers a reliable approach for sequence-based CPI prediction.
• The model's interpretability aids in understanding molecular interactions.
• This work provides valuable guidance for ligand design and chemical biology research.