DrugDL: dual-modal deep learning framework for multi-property drug prediction and targeted therapy discovery

Area of Science:

• * Computational drug discovery and cheminformatics.
• * Development of novel machine learning frameworks for bioinformatics.

Background:

• * Accurate drug molecule representation and prediction of drug-target interactions are vital but challenging in drug development.
• * Existing methods struggle with limited generalizability, lack of multitask frameworks, and poor performance in cold-start scenarios.

Purpose of the Study:

• * To introduce DrugDL, a comprehensive framework for drug molecule representation and multitask prediction.
• * To address limitations in current drug discovery tools, particularly for cold-start problems.

Main Methods:

• * Jointly learning representations of drug chemical and target protein biological spaces.
• * Integrating cross-modal contrastive learning and single-modal feature enhancement.
• * Employing a multitask prediction framework for diverse downstream tasks.

Main Results:

• * DrugDL consistently outperforms state-of-the-art methods, especially in cold-start tasks.
• * Successfully applied to high-throughput screening, inhibitor identification (SARS-CoV-2, metabolic enzymes), and cancer drug prediction.
• * Validated on EGFR and ALK targets, demonstrating precision in drug discovery.

Conclusions:

• * DrugDL offers end-to-end technical support for drug development by enabling accurate molecular representation and multi-property prediction.
• * The framework significantly accelerates the drug discovery process.
• * Provides a powerful tool for identifying novel drug candidates and understanding drug-target interactions.