UniMoMo

Generative binder design has split into three largely disconnected literatures. Structure-based drug design models generate small molecules into a pocket, peptide models generate short chains of amino acids, and antibody models generate CDR loops onto a framework. Each has its own representation, its own benchmarks, and its own weights — even though the physics they are modeling is identical: the same non-covalent interactions, the same bond lengths and angles, the same steric constraints. UniMoMo, from Tsinghua University with collaborators at Renmin University of China and ByteDance, is a single generative model that designs binders in all three domains.

The unifying move is representational. Every molecule — a drug-like ligand, a peptide, an antibody variable domain — is expressed as a graph of blocks, where a block is either a standard amino acid or a molecular fragment extracted by the principal subgraph algorithm. This keeps the hierarchical priors that make peptides tractable while retaining full-atom geometry, which is what transfers across domains. Generation is then a geometric latent diffusion process over those blocks. The paper was published at ICML 2025 and released with MIT-licensed code and a trained checkpoint.

#Key Features

• Block-graph representation: Amino acids and fragments are treated as the same kind of object, so one model consumes peptides, antibodies, and small molecules without domain-specific encoders.
• Iterative full-atom autoencoder: Each block is compressed into a latent point — a low-dimensional hidden state plus a spatial coordinate — and reconstructed to full-atom geometry through two-stage decoding, so the diffusion model never iterates over individual atoms.
• E(3)-equivariant latent diffusion: The reverse process generates E(3)-invariant latent states alongside equivariant coordinates, guaranteeing that designs respect rotational and translational symmetry of the binding site.
• Cross-domain transfer: One checkpoint trained on all three data sources outperforms the same architecture trained on any single domain, on nearly every metric within that domain.
• Multi-modality design for one target: Given a single binding site, the model can propose a small molecule, a peptide, and an antibody CDR as alternative therapeutic formats.

#Technical Details

Latent coordinates are centered on the binding-site point cloud and rescaled before diffusion, which is what makes generation efficient enough to run over antibody-sized systems. Training draws on PepBench (4,157 protein-peptide complexes) plus 70,498 ProtFrag synthetic samples for peptides, 9,473 SAbDab entries for antibodies, and 99,900 CrossDocked2020 complexes for small molecules; full training takes two to three days on eight 80GB GPUs. On the LNR set of 93 protein-peptide complexes, UniMoMo reaches 39.45% amino acid recovery, 2.19 Å complex RMSD and 1.27 Å ligand RMSD, and a pyRosetta binding energy of −34.35 against a native-peptide reference of −37.25, with the lowest internal clash rate of the models compared. On CDR-H3 design over 60 RAbD complexes, sampling 100 candidates per complex gives 52.34% recovery, 1.04 Å RMSD, and improved binding energy over the native CDR in 65% of complexes. On CrossDocked2020 evaluated through the CBGBench protocol, it takes the highest overall ranking score across substructure, chemical property, geometry, and interaction categories. Ablations against a single-domain variant show multi-domain training improving binding energy and dihedral-angle distributions in every domain, and a GPCR case study shows the model importing hydrogen-bonding motifs from small-molecule data and wide-spanning amide scaffolds from peptide and antibody data into the same pocket.

#Applications

The practical draw is format-agnostic target evaluation. A team with a validated binding site can generate candidate small molecules, peptides, antibody CDR-H3 loops, and nanobody CDRs against it from one tool and one configuration, rather than assembling three pipelines whose outputs are not directly comparable. The released repository ships demonstrations for KRas G12C across all three formats, plus a nanobody workflow that co-folds target and nanobody framework before CDR design. For antibody engineers, CDR design against a docked framework is the standard entry point; for medicinal chemists, pocket-conditioned small-molecule generation with realistic geometry and competitive docking scores is the equivalent.

#Impact

UniMoMo makes the case that multi-domain training is not merely a convenience but an accuracy gain — the all-domain model beats its own single-domain ablation inside each benchmark, which is the strongest available evidence that binder design knowledge genuinely transfers across molecular classes. The authors are explicit that the released version is a computational proof-of-concept and that its current form is simple relative to real-world requirements. Every result is in silico, scored with pyRosetta energies, Vina docking, and geometric rationality metrics rather than measured affinities, and CDR-H3 design still needs 100 samples per complex to reach its headline recovery. The framework also lacks explicit hydrogen-bond constraints, which the authors identify as the next improvement. The trained checkpoint is distributed from the repository's release page. The write-up is the one restricted artifact in an otherwise unusually open release: the PMLR v267 proceedings version states no reuse licence and the arXiv posting sits under arXiv's default non-exclusive distribution terms, so the paper's benchmark tables cannot be redistributed the way the MIT-licensed code and weights can.