DeepCDA: deep cross-domain compound-protein affinity prediction through LSTM and convolutional neural networks

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning in drug discovery

Background:

• Accurate prediction of compound-protein binding affinity is crucial for drug discovery.
• Standard computational methods often fail when test data distributions differ from training data.
• Real-world drug discovery involves diverse and often unseen compound-protein interactions.

Purpose of the Study:

• To develop a robust deep learning approach for predicting compound-protein binding affinity.
• To address the challenge of domain shift between training and testing datasets in drug discovery.
• To improve the reliability of computational models for novel compound-protein pairs.

Main Methods:

• A three-step deep learning framework combining convolutional and long-short-term memory layers for novel representations.
• Implementation of a two-sided attention mechanism to encode substructure interaction strengths.
• Utilizing adversarial domain adaptation to learn a feature encoder for the test domain.

Main Results:

• The proposed method demonstrated improved model reliability on the test domain across KIBA, Davis, and BindingDB datasets.
• The approach effectively handles challenging situations with differing data distributions.
• Learned representations capture essential patterns for accurate binding affinity prediction.

Conclusions:

• The developed deep learning approach offers a more reliable solution for binding affinity prediction in drug discovery.
• Adversarial domain adaptation is effective in mitigating the impact of domain shift.
• The method provides a valuable tool for accelerating the identification of potential drug candidates.