CoaDTI: multi-modal co-attention based framework for drug-target interaction annotation.

Area of Science:

• Computational biology
• Drug discovery
• Bioinformatics

Background:

• Accurate drug-target interaction (DTI) identification is crucial for in silico drug discovery.
• Manual DTI annotation is laborious and time-consuming.
• Existing high-throughput prediction methods often rely on error-prone manual features.

Purpose of the Study:

• To develop an efficient and interpretable deep learning framework for drug-target annotation.
• To improve the accuracy and reduce the reliance on manual features in DTI prediction.

Main Methods:

• Developed CoaDTI, an end-to-end deep learning framework utilizing a co-attention mechanism.
• Employed transformer for protein representation learning from amino acid sequences.
• Utilized GraphSage for extracting molecule graph features from SMILES.
• Implemented transfer learning with pre-trained transformers to address limited labeled data.

Main Results:

• CoaDTI achieved competitive performance on three public datasets compared to state-of-the-art models.
• Transfer learning significantly boosted prediction performance.
• Identified novel DTIs, including potential interactions with SARS-CoV-2 associated proteins.
• CoaDTI's co-attention mechanism provided interpretable insights into predicted interactions.

Conclusions:

• CoaDTI offers an efficient and interpretable solution for drug-target annotation.
• Transfer learning is a valuable strategy for enhancing DTI prediction with scarce data.
• The framework has the potential to accelerate drug discovery by identifying novel DTIs.