Benchmarking Sequence-Based Compound-Protein Interaction Prediction through Constructing a Debiased Data Set CDPN.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Accurate prediction of compound-protein interactions (CPIs) is crucial for accelerating drug discovery.
• Existing datasets often contain biases, such as over-represented molecular scaffolds and imbalanced label distributions, which can lead to machine learning shortcuts and hinder model generalization.
• Current debiasing methods may compromise dataset diversity.

Purpose of the Study:

• To address biases in compound-protein interaction (CPI) datasets.
• To introduce a novel protocol, Clustering-based Down-sampling and Putative Negatives (CDPN), for constructing a debiased CPI benchmark.
• To systematically benchmark deep learning-based CPI models, particularly protein language models, using the CDPN dataset.

Main Methods:

• Developed the Clustering-based Down-sampling and Putative Negatives (CDPN) protocol.
• CDPN mitigates biases via compound cluster-level down-sampling.
• Generated putative negatives from unexplored chemical spaces to ensure balanced label distributions.

Main Results:

• Systematically benchmarked deep learning CPI models on the CDPN dataset.
• Identified limitations in attention interpretability for protein language models on PDBbind.
• Discovered KPGT-Ankh as a superior model with enhanced generalization and virtual screening performance through ablation studies on the CDPN dataset.

Conclusions:

• The CDPN protocol effectively creates a debiased CPI benchmark.
• KPGT-Ankh demonstrates superior performance for CPI prediction.
• Top-performing models were integrated into DeepSEQreen, a no-code web server, to enhance accessibility and facilitate community feedback for drug discovery research.