Inverse Folding Towards Human Proteins

Protein inverse folding predicts amino acid sequences that fold into a specified three-dimensional protein backbone. Protein humanization aims to make nonhuman protein sequences more human-like, an important objective for reducing the immunogenicity of protein therapeutics while retaining their structure and function. In this work, we train HumanMPNN, a structure-conditioned message-passing neural network (MPNN) that prefers human-like protein sequences. We further introduce two complementary measures of sequence humanness: an immunogenicity-based metric derived from predicted immune recognition and the protein language model (PLM) Human Proximity Score, which measures proximity to human proteins in PLM embedding space. We show that both metrics correlate with phylogenetic distance from humans, and that HumanMPNN produces sequences that are more human-like and less immunogenic than standard inverse-folding baselines while preserving structural fidelity.