DeepAffinity: interpretable deep learning of compound-protein affinity through unified recurrent and convolutional...

Area of Science:

• Computational Biology
• Drug Discovery
• Machine Learning

Background:

• Accurate quantification of compound-protein interactions (CPI) is crucial for drug discovery.
• Existing methods often lack the applicability, accuracy, and interpretability needed for sequence-based prediction of compound-protein affinity.

Purpose of the Study:

• To develop a novel computational method for predicting compound-protein affinity using only sequence information.
• To enhance the accuracy and interpretability of drug-target interaction predictions.
• To integrate domain knowledge with machine learning for improved predictive performance.

Main Methods:

• A semi-supervised deep learning model unifying recurrent and convolutional neural networks was developed.
• Novel representations of structurally annotated protein sequences were utilized.
• Joint encoding of molecular representations and affinity prediction was performed using both labeled and unlabeled data.
• Attention mechanisms were embedded for model interpretability.

Main Results:

• The proposed model achieved relative error within 5-fold for IC50 on test cases.
• Performance improved by 20-fold for protein classes not included in training.
• Transfer learning enhanced predictions for new protein classes with limited labeled data.
• Attention mechanisms provided interpretability for selective drug-target interactions.

Conclusions:

• The integrated approach offers a powerful tool for predicting compound-protein affinity with high accuracy and interpretability.
• The model outperforms conventional methods, demonstrating significant potential in accelerating drug discovery pipelines.
• Further exploration of alternative representations and unified models highlights future research directions.