A Multi-Task Self-Supervised Strategy for Predicting Molecular Properties and FGFR1 Inhibitors.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Understanding molecular properties and target interactions is crucial for drug development.
• Effective molecular representations are essential for predicting properties and designing high-affinity ligands in computer-aided drug discovery.
• Developing robust multi-task and self-supervised pretraining strategies for molecular representation learning remains a challenge.

Purpose of the Study:

• To propose MTSSMol, a multi-task self-supervised deep learning framework for pretraining molecular representations.
• To leverage approximately 10 million unlabeled drug-like molecules for pretraining.
• To identify potential inhibitors of fibroblast growth factor receptor 1 (FGFR1).

Main Methods:

• Utilizing a graph neural networks (GNNs) encoder for learning molecular representations during pretraining.
• Implementing a multi-task self-supervised pretraining strategy to capture comprehensive structural and chemical knowledge of molecules.
• Validating MTSSMol's performance on 27 diverse datasets for molecular property prediction.

Main Results:

• MTSSMol demonstrated exceptional performance in predicting molecular properties across various domains.
• The framework successfully identified potential FGFR1 inhibitors.
• Validation involved molecular docking using RoseTTAFold All-Atom (RFAA) and molecular dynamics simulations.

Conclusions:

• MTSSMol offers an effective algorithmic framework for enhancing molecular representation learning.
• The study validates MTSSMol as a valuable tool for identifying potential drug candidates and accelerating drug discovery.
• The developed framework and code are publicly available to support further research.