Drug discovery requires efficient exploration of vast chemical spaces while minimizing computational cost. Fragment-based drug design offers a promising route for generating novel ligands through recombination of fragments, yet the resulting combinatorial search spaces rapidly become too large for exhaustive evaluation. Here, we present an active learning framework for navigating kinase-focused fragment recombination spaces derived from KinFragLib. Our approach combines subpocket-aware fragment recombination, structure-based docking, machine learning surrogate models, and genetic algorithm-inspired exploration through fragment additions, deletions, and substitutions around previously evaluated ligands. Rather than screening a fully enumerated recombination library, the framework iteratively generates and evaluates promising fragment modifications, enabling scalable exploration of large recombination spaces. Using protein kinase A (PKA) as a case study, we evaluate the framework on a recombination space containing approximately 210,000 ligands. Active learning identifies high-scoring candidates while evaluating only about 0.5% of the available ligands with the computationally expensive docking workflow. A greedy acquisition strategy recovers 52% of the global top-50 ligands, substantially outperforming random sampling. In addition, a diversity-aware acquisition strategy increases fragment coverage and exploration of alternative hinge-binding motifs while retaining competitive optimization performance. These results demonstrate that active learning can efficiently navigate fragment recombination spaces and substantially reduce computational screening costs. More broadly, the proposed framework provides a transferable strategy for kinase-focused ligand design and can be extended to other targets when suitable fragment libraries are available.
More information can be found in our preprint: https://chemrxiv.org/doi/full/10.26434/chemrxiv.15011770/v1
