Abstract
Antibody-antigen complexes challenge our understanding, as analyses to datefailed to unveil the key determinants of binding affinity and interaction specificity. We par-tially fill this gap based on novel quantitative analyses using two standardized databases, theIMGT/3Dstructure-DB and the structure affinity benchmark.First, we introduce a statistical analysis of interfaces which enables the classification of ligand types(protein, peptide, chemical; cross-validated classification error of 9.6%), and yield binding affinitypredictions of unprecedented accuracy (median absolute error of 0.878 kcal/mol). Second, weexploit the contributions made by CDRs in terms of position at the interface and atomic packingproperties to show that in general, VH CDR3 and VL CDR3 make dominant contributions tothe binding affinity, a fact also shown to be consistent with the enthalpy - entropy compensationassociated with pre-configuration of CDR3.Our work suggests that the affinity prediction problem could be solved from databases of highresolution crystal structures of complexes with known affinity.