Machine Learning on DNA-Encoded Library Count Data Using an Uncertainty-Aware Probabilistic Loss Function.

Area of Science:

• Computational chemistry
• Cheminformatics
• Drug discovery

Background:

• DNA-encoded library (DEL) screening and quantitative structure-activity relationship (QSAR) modeling are key techniques in drug discovery.
• Current QSAR modeling of DEL data often uses binary classifiers on aggregated 'disynthons', which can lose information and fail to capture enrichment levels.

Purpose of the Study:

• To develop a regression approach for analyzing DEL selection data that overcomes limitations of binary classification.
• To enable more nuanced understanding of molecular enrichments and facilitate the identification of novel drug candidates.

Main Methods:

• Developed a regression model to learn DEL enrichments of individual molecules.
• Implemented a custom negative-log-likelihood loss function to denoise sparse and noisy DEL data.
• Modeled the Poisson statistics of the sequencing process inherent in DEL workflows.

Main Results:

• The regression approach effectively denoises DEL data and visualizes structure-activity relationships.
• Models can identify low-confidence outliers by accounting for data uncertainty.
• Demonstrated on large DEL datasets screened against carbonic anhydrase (CAIX), soluble epoxide hydrolase (sEH), and SIRT2.

Conclusions:

• The proposed regression method offers a powerful tool for analyzing DEL data, improving the identification of structure-activity trends and enriched pharmacophores.
• This uncertainty-aware regression approach is applicable to other sparse or noisy datasets with known stochasticity.
• Enhances the utility of DEL screening in drug discovery pipelines.