Toward Accurate in Silico Prediction of Antigen Binding Affinities for Antibody Engineering

Abstract

In clinical applications and life sciences research, antibodies represent an important diagnostic and therapeutic potential thanks to their high target affinity, specificity, and broad developability. While the antigen affinity, one of the primary success assessors of an antibody, can be measured at reasonably high precision and reliability, the scalability of the measurements can be cumbersome and limited. This is troubling because the affinity must be monitored throughout all steps of the developability approaches such as affinity maturation and humanization of an antibody. In this context, in silico approaches present a lucrative opportunity at a fraction of the cost and time typically invested in a comparable wet lab undertaking. In addition to their high-throughput potential, in silico approaches can provide an invaluable side product, i.e., identifying the molecular driving forces behind affinity. Here, we investigated the performance of six different high-throughput servers in two settings common in antibody engineering applications: (i) de novo prediction of the experimental antibody-antigen binding constants, and (ii) the free energy change in binding due to single point mutations. We find that the accuracy of these tools can be significantly low in the two regimes relevant to antibody development: (i) prediction of high-affinity binding, and (ii) prediction of favorable mutations. These issues are intricately related to the training sets used in the underlying models of these tools where high-affinity complexes and favorable point mutations are typically underrepresented. We showed that biophysical characteristics of single point mutations, especially changes in bulkiness and hydrophobicity, increase the prediction error. We argue that while the prediction of mutational impact can be predicted within one kcal per mol using re-parameterized versions of the present in silico tools, the de novo prediction of the affinity likely requires revisiting the underlying physical models behind these tools. © 2025