Generative AI-driven de novo design of high-affinity, epitope-specific antibodies targeting the CD93-IGFBP7 axis for cancer immunotherapy

ABSTRACT Background CD93 is an emerging immune checkpoint. Blocking its interaction with IGFBP7 restores anti-tumor immunity, yet discovering high-affinity antibodies against this specific functional epitope remains a formidable challenge for traditional methods. Methods We employed a proprietary generative AI model, GeoFlow, to de novo design antibodies targeting the predicted binding epitope of a reference antibody 7F3. A library constructed from 105 sequences was screened via phage display, and the lead candidate, 7-4_com10, was validated through biolayer interferometry (BLI), flow cytometry, and in vivo efficacy studies. Results To dissect the contribution of endothelial CD93 to immunotherapy resistance, we utilized an endothelial-specific conditional knockout model in MB49 bladder cancer, where the loss of endothelial CD93 significantly sensitized tumors to PD-1 inhibition. The genetic findings identified endothelial CD93 as a fundamental driver of immunotherapy resistance and provided the rationale for our GeoFlow generative AI platform, which was employed to de novo design approximately 105 candidates targeting the IGFBP7-binding interface on CD93. Our lead candidate, 7-4_com10, demonstrated superior affinity and more potent functional blockade of the CD93-IGFBP7 axis compared to the reference 7F3. Crucially, 7-4_com10 sensitized H22 solid tumors to anti-PD-1 therapy in vivo, which confirmed that 7-4_com10 successfully targeted the identified immunosuppressive axis and pharmacologically phenocopied the benefits of genetic endothelial ablation. Conclusions This study validated generative AI for designing epitope-specific antibodies with superior potency, offering a robust candidate for next-generation immunotherapy.