DOMINO: Learning Domain Co-occurrence for Multidomain Protein Design

Abstract

Multidomain proteins arise through the reuse and recombination of structural domains, yet natural architectures represent a sparse, structured sample of the possible domain-combination space. Here, we introduce DOMINO, a two-stage framework that learns domain co-occurrence from TED-annotated multidomain proteins and uses the learned patterns to generate new multidomain sequences. DOMIN, a contrastive retrieval model, embeds domains into a latent compatibility space and retrieves candidate partners for a query domain from a TED-derived domain pool, including pairings not observed in the TED-derived co-occurrence set. DOMO, a conditional autoregressive sequence model, converts each retrieved domain pair into a full-length protein sequence by jointly generating the specified domain regions and the non-domain sequence context between and around them. DOMIN recovers hierarchical patterns of natural domain co-occurrence and expands the observed CATH homologous-superfamily co-occurrence network with candidate novel pairings. DOMO realizes both held-out natural pairs and DOMIN-retrieved pairs as proteins with high domain recovery and high AlphaFold-predicted structural confidence. Applied at scale, DOMINO generated 5 million retrieval-derived multidomain proteins, with sampled designs showing recovery of the specified domains, diverse CATH annotations, and sequence novelty relative to UniRef100. Together, these results support domain co-occurrence as a predictive design prior and demonstrate a scalable strategy for exploring multidomain protein architectures through new combinations of existing structural modules.