Pan-Target Generation and Validation of Antibody Sequences Driven by an Iterative, Multimodal Foundation Model AI Approach

Antibody therapeutic discovery has entered the digital age, with AI researchers increasingly emphasizing 1) large-language foundation models that can generate molecules with innately human properties and 2) 3D-diffusion models that can capture aspects of antibody-antigen complex structures in design. Despite these advances in technology for in-silico antibody generation, one-shot hit rates remain prohibitively low for producing novel antibody candidates that deviate from known binding templates at more than a few positions. Multiple iterations over wet-lab and in-silico experimental cycles promise to help anneal binding and property concerns found in initial candidates and drive sequences toward de novo epitopes. We here present an iterative AI and experimental workflow that combines large-language model, graphical, and 3D diffusion AI architectures for antibody candidate generation with a computational and experimental validation approach. We demonstrate our methodology on three diverse but well-characterized antibody targets. We see that each target places unique demands on our foundation model AI workflow based on the richness of available antibody binding data, desired coverage of target epitope space, and antibody novelty preferences. AI-generated candidates also showed favorable non-specific binding profiles. We discuss our progress in various dimensions toward using AI for productively augmenting experimental approaches and driving true de novo antibody design.