Toward generalizable predictive models for DNA-encoded libraries.

Area of Science:

• Drug Discovery
• Computational Chemistry
• Biotechnology

Background:

• DNA-encoded libraries (DELs) combined with machine learning (ML) present a powerful approach for hit identification in drug discovery.
• Sequencing data from DELs can be noisy and biased, leading to ML models that overfit to specific chemical libraries.

Purpose of the Study:

• To critically evaluate the capabilities and limitations of DEL-ML for hit identification.
• To illustrate challenges in DEL-ML using Aurora Kinase A (AURKA) affinity selection data.
• To propose strategies for building robust DEL-ML models.

Main Methods:

• Review of existing DEL-ML methodologies.
• Analysis of Aurora Kinase A (AURKA) DEL affinity selection data.
• Evaluation of denoising strategies and domain adaptation techniques.

Main Results:

• Standard ML models often fail to generalize to new chemical spaces due to library-specific structural constraints.
• Noisy and biased sequencing data significantly impacts DEL-ML model performance.
• Specific structural constraints in combinatorial libraries limit model generalization.

Conclusions:

• Rigorous denoising strategies are essential for effective DEL-ML.
• Techniques like domain adaptation can help mitigate limitations and improve model robustness.
• A roadmap is proposed for developing DEL-ML models capable of exploring diverse chemical space effectively.