Accurate prediction of molecular targets using a self-supervised image representation learning framework.

Area of Science:

• Computational drug discovery
• Artificial intelligence in pharmacology
• Molecular target identification

Background:

• Drug efficacy and safety depend on molecular targets, but proteome-wide evaluation is challenging.
• Current methods for assessing drug properties and targets are complex and time-consuming.
• Developing efficient computational tools is crucial for advancing drug discovery.

Approach:

• Developed ImageMol, an unsupervised deep learning framework using 8.5 million unlabeled molecular images.
• ImageMol pre-trains chemical representations from pixel-based local and global molecular structures.
• The framework leverages self-supervised learning for image processing-based molecular analysis.

Key Points:

• ImageMol accurately predicts molecular properties like metabolism, brain penetration, and toxicity.
• The framework demonstrates high performance in identifying molecular target profiles, including for HIV.
• ImageMol identified anti-SARS-CoV-2 molecules and re-prioritized 3CL inhibitors for COVID-19 treatment.

Conclusions:

• ImageMol offers a powerful, image processing-based strategy for computational drug discovery.
• The framework accelerates the identification of drug candidates for various human diseases.
• This approach provides a valuable toolbox for accelerating the development of new therapeutics.