Deep contrastive learning method for drug-target interactions prediction.

Area of Science:

• Computational Biology
• Drug Discovery
• Machine Learning

Background:

• Accurate drug-target interaction (DTI) prediction is crucial for efficient drug discovery.
• Existing methods often struggle with low-coverage datasets and numerical stability.

Purpose of the Study:

• To introduce DeepCL, a novel deep contrastive learning framework for improved DTI prediction.
• To address numerical underflow and enhance separation between interacting and non-interacting pairs.

Main Methods:

• Developed DeepCL, a dual-pathway framework utilizing ESM-2 protein language models and Morgan fingerprints.
• Implemented a generalized sigmoid activation function and a margin-based contrastive loss.
• Aligned heterogeneous protein and molecular features in a shared latent space.

Main Results:

• DeepCL significantly outperformed state-of-the-art methods on three benchmark datasets (Davis, BindingDB, BIOSNAP).
• Achieved superior performance in both standard and zero-shot DTI prediction settings.
• Demonstrated improved AUPR and AUROC scores, indicating higher prediction accuracy.

Conclusions:

• DeepCL offers a robust, scalable, and numerically stable solution for DTI prediction.
• The framework shows particular promise for accelerating drug discovery in low-data scenarios.