Bioactivity descriptors for uncharacterized chemical compounds.

Area of Science:

• Chemoinformatics
• Computational Biology
• Drug Discovery

Background:

• Chemical descriptors are crucial for chemoinformatics, representing molecular properties.
• Bioactivity data enriches compound representation but is scarce for most molecules.
• Existing bioactivity descriptors limit analysis to a few thousand well-characterized compounds.

Purpose of the Study:

• To develop deep neural networks (signaturizers) for inferring bioactivity signatures for any compound.
• To enable the use of bioactivity signatures as replacements for chemical descriptors in chemoinformatics.
• To improve the prediction of biological activities and enhance the exploration of chemical space.

Main Methods:

• A collection of deep neural networks was developed to infer bioactivity signatures.
• Signaturizers were trained to predict 25 different types of bioactivities.
• Signature-activity relationship (SigAR) models were implemented and compared to chemistry-based classifiers.

Main Results:

• Inferred bioactivity signatures can be used as drop-in replacements for chemical descriptors.
• Signatures enable biologically relevant navigation of chemical space and reveal organization in natural products.
• Enrichment of uncharacterized libraries for activity against the drug-orphan target Snail1 was demonstrated.
• SigAR models showed substantial performance improvements over chemistry-based classifiers in activity prediction benchmarks.

Conclusions:

• Deep learning models can effectively infer bioactivity signatures for compounds lacking experimental data.
• Inferred bioactivity signatures offer a powerful tool for chemoinformatics and drug discovery.
• This approach significantly enhances biological activity prediction and chemical space exploration.