Active learning-guided docking enables efficient exploration of ultra-large chemical libraries otherwise inaccessible due to the cost of exhaustive screening. By iteratively selecting the most informative compounds, it prioritizes candidates with high predicted docking scores and model uncertainty, ensuring balanced exploration of chemical space. Only a small fraction of molecules is sampled, reducing computational demands while maintaining strong enrichment of promising candidates.
In this work, the method was applied to LasR, a key regulator of bacterial virulence. Targeting virulence factors such as LasR is particularly attractive because it may reduce the selective pressure associated with traditional antibiotics and limit the emergence of resistance. Moreover, the scarcity of known ligands for LasR makes structure-based approaches especially relevant.
Overall, active learning-guided docking enables scalable and flexible navigation of chemically diverse, billion-scale libraries, supporting the identification of novel drug candidates and providing a practical framework for integration with experimental validation workflows. Importantly, experimental validation of selected candidates confirmed the models predictive capability, highlighting its promise as an effective tool for accelerating drug discovery.
Importantly, the iterative nature of active learning allows continuous refinement of the model as new data is incorporated. This closed-loop workflow efficiently allocates computational resources to the most relevant regions of chemical space and bridges the gap between in silico screening and real-world drug discovery.

Acknowledgements: This work was supported by the National Science Centre (NCN), Poland, under the SONATA-19 grant no. 2023/51/D/NZ7/0259.

Co-authors: Anna Górska-Ratusznik, Agata Barzowska-Gogola, Joanna Budziaszek, Adam Sułek, Tomasz Kościółek, Barbara Pucelik