Assessing the Generalizability of Machine Learning and Physics-Based Methods with DNA-Encoded Libraries

Area of Science:

• Drug discovery and development
• Computational chemistry
• Machine learning in cheminformatics

Background:

• DNA-encoded libraries (DELs) offer vast chemical space screening.
• Machine learning (ML) models trained on DEL data often fail to generalize to novel chemical structures (out-of-distribution, OOD).

Purpose of the Study:

• To investigate if structural modeling methods can improve the generalization of ML models for DEL data.
• To assess the performance of ML, docking, and co-folding methods for OOD hit discrimination.

Main Methods:

• Systematic assessment of state-of-the-art ML, docking, and co-folding algorithms.
• Evaluation across three diverse protein targets and multiple DEL synthesis formats.
• Analysis of in-distribution and OOD performance for hit identification.

Main Results:

• ML models perform well on in-distribution data but poorly on OOD chemical space.
• Optimal OOD hit discrimination is dependent on the specific protein target and ligand.
• Structural modeling approaches show potential for bridging the generalization gap.

Conclusions:

• Rigorous, system-specific pilot testing is essential for reliable virtual screening predictions.
• Aggregated benchmark performance may not reflect real-world applicability.
• Open-source tools (DEL-iver) are provided to facilitate these workflows.