Learning Molecular Representation in a Cell.

Area of Science:

• Computational Biology
• Drug Discovery
• Machine Learning

Background:

• Predicting drug efficacy and safety requires understanding biological responses to small molecules.
• Current molecular representation learning methods lack comprehensive cell state views and struggle with noise, limiting generalization.

Purpose of the Study:

• To introduce the Information Alignment (InfoAlign) approach for learning robust molecular representations.
• To improve the prediction of drug efficacy and safety by enhancing molecular representations.

Main Methods:

• InfoAlign integrates molecules and cellular response data into a context graph.
• It utilizes an information bottleneck method with minimality and sufficiency objectives.
• The sufficiency objective aligns latent representations with molecular neighborhood feature spaces.

Main Results:

• InfoAlign's sufficiency objective is demonstrated to be tighter than existing contrastive methods.
• Representations were validated in molecular property prediction, outperforming 27 baseline methods.
• Zero-shot molecule-morphology matching was also successfully performed.

Conclusions:

• InfoAlign provides a powerful new method for learning molecular representations from cellular responses.
• This approach enhances the prediction of molecular properties and biological effects.
• InfoAlign offers a promising direction for improving drug discovery and development.