A geometry-over-coevolution principle governs protein complex assembly in AlphaFold

Abstract

AlphaFold has revolutionized protein complex structure prediction, yet how it assembles intermolecular interfaces remains poorly understood. Contrary to the prevailing view that inter-protein coevolution drives complex prediction, we uncover a geometry-over-coevolution principle governing assembly in AlphaFold-Multimer and AlphaFold3. Through systematic perturbation of evolutionary and structural inputs and development of residue-level constraint propagation mapping to trace the emergence and propagation of geometric information within the network, we show that prediction accuracy is governed primarily by monomer-derived structural geometry and interface-specific sequence-geometry compatibility, rather than direct inter-protein coevolutionary signals. Our mapping reveals a hierarchical assembly mechanism in which monomer-level geometric representations are established first and progressively propagated to constrain cross-chain interfaces. This mechanism further explains why antigen-antibody complexes are predicted less accurately, as their intrinsic interface plasticity and non-canonical architectures limit the propagation of geometric constraints across interfaces. Together, these findings establish a mechanistic framework for understanding how artificial intelligence models assemble protein complexes and provide principles for improving next-generation structure prediction.