Integrated AI-Guided de novo CDR-H3 Design and Combinatorial Phage Display for Antibody Discovery

Abstract Background Obtaining target-specific therapeutic antibodies is critical for maximizing efficacy, yet conventional library-based approaches cannot guarantee epitope specificity or enable straightforward humanization; computational de novo design of human antibodies targeting specific epitopes has therefore attracted increasing interest. Methods We developed SANDAN (Structure-Aware de Novo Design for ANtibody), an end-to-end, structure-driven workflow that uses the antigen structure and epitope residues without supplying a known binding antibody or its complex pose as design input. Starting from a human framework with diverse CDR-H2 and CDR-L3 variants, the workflow performs CDR-H3-focused design using the diffusion-based method DiffAb on diverse template complex structures predicted by Boltz-1, aiming to obtain antibodies that bind specifically to the target epitope without requiring known complex pose information. Targeting human TIGIT (hTIGIT), we selected approximately 1,000 CDR-H3 design sequences and constructed a phage display library (diversity approximately ${10}^6$) incorporating them. Results Screening yielded six hTIGIT-binding antibodies that competed with the reference antibody MG1131. Their CDR-H3 sequences differed from that reference antibody’s CDR-H3 sequence. Conclusions Combining AI-guided CDR-H3 design with phage display screening yielded hTIGIT-binding antibodies that competed with MG1131 from a combinatorial library containing the designed H3 sequences.