Machine-Learning-Based Data Analysis Method for Cell-Based Selection of DNA-Encoded Libraries.

Area of Science:

• Biochemistry
• Computational Chemistry
• Drug Discovery

Background:

• DNA-encoded library (DEL) technology is crucial for ligand discovery in pharmaceuticals.
• Traditional DEL selections use purified proteins, but cell-based selections face challenges due to noisy data and nonspecific interactions.
• Existing denoising methods' suitability for cell-based DEL data remains uncertain.

Purpose of the Study:

• To introduce and validate a novel machine-learning (ML)-based approach for processing cell-based DEL selection datasets.
• To address the challenge of identifying reliable hits from noisy, complex cell-surface target selections.
• To quantify uncertainties in noisy DEL data using a probabilistic framework.

Main Methods:

• Development of a machine-learning approach utilizing a Maximum A Posteriori (MAP) estimation loss function.
• Application of an extended-connectivity fingerprint (ECFP)-based regression model to a cell-based DEL dataset.
• Calculation of a regularized enrichment metric (MAP enrichment) for outlier suppression.

Main Results:

• The ML-MAP approach successfully identified true binders and reliable structure-activity relationships (SAR) from noisy cell-based DEL data.
• The MAP enrichment metric effectively enhanced the signal-to-noise ratio by suppressing low-confidence outliers.
• Demonstrated proof-of-principle for processing complex cell-based DEL selection datasets.

Conclusions:

• The ML-MAP approach offers a robust solution for analyzing noisy cell-based DEL selection data.
• This method improves the reliability of hit identification and SAR analysis in complex biological systems.
• Future work will focus on applying this method for de novo ligand discovery from cell-based DEL selections.