From Code to Cure: Computationally Designed BMP-2 Binders Using AI-Integrated Pipelines for Controlled Bone Regeneration

Abstract

Nonunion fractures remain a costly and persistent challenge in regenerative medicine, with current treatments limited by donor-site morbidity, restricted graft availability, and severe adverse effects associated with supraphysiological bone morphogenetic protein 2 (BMP-2) delivery, including ectopic ossification and inflammation. Endogenous BMP-2 signaling is tightly regulated in adult tissues, constraining the precision and scalability of approaches based on transcriptional upregulation or bolus growth factor administration. To address these limitations, we developed a two-phase integrated computational-experimental pipeline for the de novo design of protein binders targeting the BMP-2 knuckle epitope, a receptor-binding surface corresponding to BMPR-II engagement, enabling affinity-tuned modulation of BMP-2 activity rather than uncontrolled pathway activation. Phase I employed PyRosetta-based {beta}-strand motif grafting and physics-based docking protocols to generate 264 candidate binders, followed by deep-learning-driven refinement in Phase II using partial RFDiffusion and ProteinMPNN with AlphaFold2 validation, yielding 22 candidates with stable {beta}-sheet architectures consistent with knuckle-epitope targeting. Experimental validation demonstrated dose-dependent BMP-2 binding, with the lead construct exhibiting an apparent KD of 2.07 nM toward BMP-2. Targeted alanine substitutions revealed differential residue contributions, with mutation of T42 significantly disrupting binding, while other substitutions had more modest effects, indicating a partially hotspot-driven interface supported by other interactions.