MDNN-DTA: a multimodal deep neural network for drug-target affinity prediction

Area of Science:

• Computational Biology
• Bioinformatics
• Drug Discovery

Background:

• Drug-target affinity (DTA) determination is crucial in drug discovery.
• In silico methods, particularly AI and deep learning, offer efficient alternatives to experimental DTA prediction.
• Predicting DTA from large-scale biological sequence data presents a significant challenge.

Purpose of the Study:

• To introduce MDNN-DTA, a novel multimodal deep neural network for DTA prediction.
• To develop a model capable of predicting DTA directly from drug and protein sequences, eliminating the need for 3D protein structures.
• To enhance feature extraction from protein sequences using advanced AI techniques.

Main Methods:

• Utilized Graph Convolutional Networks (GCN) for drug molecule feature extraction.
• Employed Convolutional Neural Networks (CNN) for protein sequence feature extraction.
• Integrated an ESM pre-trained model and a custom Protein Feature Extraction (PFE) block for comprehensive protein sequence analysis, further enhanced by a Protein Feature Fusion (PFF) block.

Main Results:

• MDNN-DTA demonstrated effective DTA prediction directly from sequence data.
• The model successfully extracted high-dimensional features from both drug and protein sequences.
• Ablation studies confirmed the performance and efficacy of individual components within the MDNN-DTA architecture.

Conclusions:

• MDNN-DTA offers a powerful and efficient AI-driven approach for DTA prediction in drug discovery.
• The model's ability to utilize sequence data overcomes limitations associated with the unavailability of protein 3D structures.
• MDNN-DTA represents a significant advancement in computational drug discovery, improving efficiency and reducing costs.