Developing a therapeutic antibody is a long, tedious, and expensive process. Many obstacles need to be overcome, such as biophysical properties (issues of solubility, stability, weak production yields, etc.), as well as cross-reactivity and subsequent toxicity, which are major issues. No in silico method exists today to solve such issues. We hypothesized that if we were able to properly measure the similarity between the CDRs of antibodies (Ab) by considering not only their evolutionary proximity (sequence identity) but also their structural features, we would be able to identify families of Ab recognizing similar epitopes. As a consequence, Ab within the family would share the property to recognize their targets, which would allow (i) to identify off-targets and forecast the cross-reactions, and (ii) to identify new Ab specific for a given target. Testing our method on 238D2, an antagonistic anti-CXCR4 nanobody, we were able to find new nanobodies against CXCR4 and to identify influenza hemagglutinin as an off-target of 238D2.

Antibody discovery is a lengthy and labor-intensive process, requiring extensive laboratory work to ensure that an antibody demonstrates the appropriate efficacy, production, and safety characteristics necessary for its use as a therapeutic agent in human patients. Traditionally, this process begins with phage display or B-cells isolation campaigns, where affinity serves as the primary selection criterion. However, the initial leads identified through this approach lack sufficient characterization in terms of developability and epitope definition, which are typically performed at late stages. In this study, we present a pipeline that integrates early-stage phage display screening with AI-based characterization, enabling more informed decision-making throughout the selection process. Using immune checkpoints TIM3 and TIGIT as targets, we identified five initial leads exhibiting similar binding properties. Two of these leads were predicted to have poor developability profiles due to unfavorable surface physicochemical properties. Of the remaining three candidates, structural models of the complexes formed with their respective targets were generated for 2: T4 (against TIGIT) and 6E9 (against TIM3). The predicted epitopes allowed us to anticipate a competition with TIM3 and TIGIT binding partners, and to infer the antagonistic functions expected from these antibodies. This study lays the foundations of a multidimensional AI-driven selection of lead candidates derived from high throughput analysis.