Investigating AlphaFold’s handling of nanobody-antigen complex prediction

AlphaFold2 revolutionized our ability to predict a protein’s structure from its sequence. However, predicting antibody structures and antibody-antigen complexes remains a challenge. To investigate this issue further, we analyzed AlphaFold’s accuracy in predicting nanobody-antigen complexes. Nanobodies are the smallest and simplest antibody derivatives capable of binding antigens. They represent the variable portions of a special type of antibody, called a heavy-chain antibody, which lacks light chains, reducing the structural complexity of their bound complexes. We first benchmarked AlphaFold’s poor performance across 40 known nanobody-antigen complexes. Using the DockQ metric, 35 of 40 complexes ranked in the lowest accuracy category: incorrect. We then hypothesized that an inability to accurately predict nanobody contact sites may contribute to AlphaFold’s poor overall nanobody-antigen modeling. By comparing contact sites of predicted nanobody-antigen complexes against those of their experimentally verified reference structures, we concluded that AlphaFold does not accurately identify nanobody contacts; 32% of the reference contacts were missed and 38% of the predicted contacts were wrong. We further hypothesized that poor intuition for nanobody structural flexibility may limit AlphaFold's ability to predict nanobody contact sites. However, by comparing nanobody conformations of AlphaFold models generated with or without antigens, we found that AlphaFold modifies nanobody structure to accommodate antigen binding, suggesting that AlphaFold may incorrectly identify nanobody contact sites despite an apparent understanding of nanobody structural flexibility. These results motivate future work in this area, such as examination of the effect of incorporating nanobody contact site information as modeling input on AlphaFold complex prediction accuracy.