CPI-Pred: A deep learning framework for predicting functional parameters of compound-protein interactions.

Area of Science:

• Computational Biology
• Biochemistry
• Machine Learning

Background:

• Deep learning advances genome annotation and enzyme discovery.
• Predicting compound-protein interactions (CPI) is challenging due to data complexity and sparsity.
• Accurate CPI prediction is crucial for understanding enzyme function and regulation.

Purpose of the Study:

• Introduce CPI-Pred, a versatile deep learning model for predicting compound-protein interaction function.
• Leverage novel deep learning architectures for enhanced CPI prediction.
• Compile the largest dataset of enzyme kinetic parameters for training and evaluation.

Main Methods:

• CPI-Pred integrates compound representations from message-passing neural networks and enzyme representations from protein language models.
• Utilizes innovative sequence pooling and cross-attention mechanisms.
• Trained and evaluated on a comprehensive dataset of enzyme kinetic parameters (K_M, k_cat, k_cat/K_M, K_I).

Main Results:

• CPI-Pred accurately predicts diverse CPI types using only enzyme amino acid sequences and compound structural representations.
• Outperforms state-of-the-art models on unseen compounds and structurally dissimilar enzymes.
• Demonstrates the model's generalization capabilities across a wide range of interactions.

Conclusions:

• CPI-Pred offers a valuable tool for metabolic engineering challenges, including novel enzyme and compound design.
• The curated dataset serves as a benchmark for machine learning models in enzyme activity and promiscuity prediction.
• Facilitates the discovery of new enzymes and metabolites with potential applications in various biological contexts.