Sumaiya Iqbal, Wei Jiang, Eric Hansen, Tonia Aristotelous, Shuang Liu, Andrew Reidenbach, Cerise Raffier, Alison Leed, Chengkuan Chen, Lawrence Chung, Eric Sigel, Alex Burgin, Sandy Gould, Holly H. Soutter

npj Drug Discovery  2, 5 (2025)

https://doi.org/10.1038/s44386-025-00007-4

Abstract

DNA-encoded library (DEL) technology enables ultra-high-throughput screening of millions to billions of compounds in a cost-effective manner. Coupled with machine learning (ML), DEL data can be used to train predictive models for virtual screening of drug-like compounds. This study presents a systematic evaluation of 15 DEL+ML combinations—three DELs and five ML models—for hit discovery against two therapeutic targets, CK1α and CK1δ. Among 808 predicted binders tested, 10% were confirmed as true binders in biophysical assays, including two nanomolar hits. Additionally, 94% of predicted non-binders were confirmed as true negatives. The study highlights the importance of chemical diversity in training data and model generalizability over accuracy. All models and data have been made publicly available for further use.

Summary

This study introduces a five-module DEL+ML pipeline: DEL screening, data preparation, ML model development, hit prediction, and experimental validation. Three DELs (MS10M, HG1B, DD11M) and five ML models (MLP, SVM, RF, XGB, ChemProp) were evaluated for their ability to identify orthosteric binders of CK1α/δ. The HG1B DEL-trained ChemProp model achieved the best performance, with a 15% hit rate and identification of two nanomolar binders. The results demonstrate that chemical diversity of the DEL training set is more critical than library size, and that deep learning-based models outperform traditional ML algorithms.

Highlights

1. First systematic evaluation of 15 DEL+ML combinations for hit discovery. 2. HG1B DEL-trained ChemProp model achieved the highest hit rate (15%). 3. Two nanomolar binders were identified (CK1α: 308 nM; CK1δ: 187 nM and 69.6 nM). 4. 94% of predicted non-binders were experimentally confirmed as true negatives. 5. Chemical diversity of training data is more important than library size. 6. Neural network models (ChemProp, MLP) outperformed traditional models (RF, SVM, XGB). All models and training data are publicly available for community use.

Conclusion

This study demonstrates the effectiveness of integrating DEL screening with machine learning for efficient and scalable hit discovery. By systematically comparing different DELs and ML models, we found that the chemical diversity, drug-like properties, and balanced representation of binders and non-binders in the training data are key determinants of model performance. The HG1B DEL-trained ChemProp model emerged as the top performer, successfully identifying multiple potent binders, including nanomolar-level hits. The DEL+ML pipeline not only accelerates hit identification but also effectively filters out non-binders, reducing experimental costs. We recommend using chemically diverse and drug-like DELs for training and leveraging deep learning models for optimal performance. All models and data have been released as open-source resources to support further research and development.