Compositional Deep Probabilistic Models of DNA-Encoded Libraries.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Bioinformatics

Background:

• DNA-encoded libraries (DEL) are powerful for screening small molecules.
• DEL experiments generate complex data that can obscure important signals.
• Machine learning is needed to extract insights from complex DEL data.

Purpose of the Study:

• Introduce DEL-Compose, a deep probabilistic model for DEL data analysis.
• Improve observation models for DEL count data using covariate factors.
• Enhance the interpretability and robustness of DEL data analysis.

Main Methods:

• Decomposed molecular representations into monosynthon, disynthon, and trisynthon building blocks.
• Modeled latent reactions between embedded synthons using a hierarchical structure.
• Integrated covariate factors into observation models to account for data noise.

Main Results:

• DEL-Compose demonstrated strong performance on public benchmark datasets (CA-IX and HRP).
• The model successfully enriched correct pharmacophores, indicating effective binder identification.
• The intrinsic interpretable structure provided valuable insights into the DEL data.

Conclusions:

• DEL-Compose offers a robust computational tool for analyzing complex DNA-encoded library data.
• The model's approach enhances the discovery of potent binders and provides interpretable results.
• This work advances the application of machine learning in DEL-based drug discovery.