CAT-CPI: Combining CNN and transformer to learn compound image features for predicting compound-protein interactions

Area of Science:

• Computational chemistry and cheminformatics
• Bioinformatics and computational biology
• Artificial intelligence in drug discovery

Background:

• Compound-protein interaction (CPI) prediction is crucial for drug discovery but is traditionally time-consuming and expensive.
• Deep learning models, particularly those leveraging computer vision techniques, offer a promising avenue to enhance the efficiency and accuracy of CPI prediction.
• Characterizing molecular structures and protein sequences effectively is key to improving predictive models.

Purpose of the Study:

• To propose a novel deep learning model, CAT-CPI, for accelerated and accurate compound-protein interaction prediction.
• To utilize molecular images and protein sequences as input for a hybrid Convolutional Neural Network (CNN) and Transformer-based architecture.
• To introduce a Feature Relearning (FR) module to capture complex interaction features between compounds and proteins.

Main Methods:

• Employed Convolutional Neural Networks (CNNs) to extract local features from molecular images.
• Utilized Transformer encoders to capture semantic relationships from both molecular image features and protein sequence k-grams.
• Developed a Feature Relearning (FR) module to learn and integrate compound-protein interaction patterns.

Main Results:

• CAT-CPI demonstrated competitive performance compared to state-of-the-art predictors on three benchmark datasets (Human, Celegans, and Davis).
• The model effectively learns from molecular image representations and protein sequence information.
• Drug-Drug Interaction (DDI) experiments validated the potential of molecular image-based methods and the FR module.

Conclusions:

• CAT-CPI offers a powerful and efficient deep learning approach for compound-protein interaction prediction, leveraging molecular imaging.
• The integration of CNNs, Transformers, and the FR module provides a robust framework for drug discovery acceleration.
• This study highlights the significant potential of image-based molecular representations in computational drug development.