OneProt

A sequence encoder never sees the pocket geometry that decides whether a ligand binds, and a structure encoder never sees the UniProt annotation naming the reaction the enzyme catalyses. Training one joint model on all of it usually demands proteins for which every modality is present at once — and in practice those slices barely overlap. Predicted structures, detected pockets, multiple sequence alignments and curated text annotations each cover a different subset of sequence space, so the complete-tuple requirement throws most of the data away.

OneProt sidesteps that by borrowing the ImageBind recipe. Every modality gets its own encoder, and each is aligned by an InfoNCE contrastive loss to a single anchor — the amino acid sequence — using only pairs, never complete tuples. A protein with a structure but no detected pocket still trains the sequence-structure alignment. Because all encoders map into the same latent space through that anchor, modality pairs never seen together during training still align: the paper measures this emergent effect directly, recovering the matching pocket for a text query at a median rank of 64 out of a 4,000-protein test set despite text and pocket never being paired.

The model was built at the Jülich Supercomputing Centre of Forschungszentrum Jülich with Helmholtz Munich, the Technical University of Munich and Heinrich Heine University Düsseldorf, appearing as a preprint in November 2024 and in PLOS Computational Biology a year later. Two checkpoints are released: OneProt-5, spanning all five modalities, and OneProt-4, which omits the structure-token encoder.

#Key Features

• Pairwise alignment instead of complete tuples: Each modality is paired only with sequence, so a protein missing an annotation or a pocket still contributes to every alignment it does have.
• Binding pockets as a pretraining modality: An all-atom ProNet graph encoder over P2Rank-predicted pockets is trained from scratch alongside the others. Ablations single this encoder out as the one that most improves protein-protein interaction and metal-ion binding performance.
• Emergent cross-modal retrieval: Untrained modality pairs such as structure-token to structure-graph reach a median rank of 2, far ahead of the rank of 2,000 that random alignment would give.
• Transfer through frozen embeddings: Downstream use precomputes embeddings from the fixed checkpoint and fits an MLP head of at most 1.5M parameters, so no encoder is refit per task.
• Room for additional modalities: Adding a modality means training one more encoder against the sequence anchor, without retraining or re-pairing the existing ones.

#Technical Details

The sequence anchor is a frozen ESM-2 650M encoder. Around it sit a 35M-parameter ESM2 transformer trained from scratch on Foldseek structure tokens, two 2.6M all-atom ProNet graph networks for backbone structure and for pockets, and a BiomedBERT text encoder adapted with LoRA — 802.2M parameters in total, of which 40.6M are trainable. Pretraining ran for 33,000 optimizer steps on 64 NVIDIA A100 GPUs of the JUWELS Booster supercomputer. Training pairs come from the OpenProteinSet and UniProtKB/Swiss-Prot clustered at 50% sequence identity, with structures from the AlphaFold Protein Structure Database, pockets from P2Rank, and text mined from UniProt: 1.04M sequences, 1M structure tokens, 656K structure graphs, 546K text annotations and 341K pockets.

Across ten downstream benchmarks evaluated over six runs each, OneProt-4 reaches 0.668 Spearman on FLIP thermostability, 88.8% accuracy on HumanPPI and 0.877 Fmax on Enzyme Commission number prediction, matching or beating SaProt, ESM-3, OpenFold in SoloSeq mode and ProTrek despite training on roughly 1M paired datapoints against ProTrek's 40M. On the ProSPECCTs binding-site comparison benchmark it leads every baseline without any task-specific tuning.

#Applications

The practical output is an embedding service: a protein sequence goes in, and a vector comes out that already carries structural, pocket and functional-annotation signal, ready for a small classifier. That suits enzyme function assignment, subcellular localisation, thermostability ranking and metal-ion binding screens where a group has a few thousand labelled proteins and no budget to fine-tune a large encoder. Cross-modal retrieval supports the inverse query — finding proteins whose pockets match a described function — the shape of a hit-list problem in drug repurposing and biocatalytic reaction planning.

#Impact

OneProt is a careful demonstration that modality coverage can substitute for scale: a 40.6M trainable-parameter alignment over about a million pairs holds its own against contrastive models trained on forty times the data, and the ablations make the case that the binding-pocket encoder is doing much of that work. Weights for both checkpoints are on HuggingFace under an MIT license — OneProt-4 downloadable outright, OneProt-5 behind an automatic click-through agreement — with the pretraining splits archived on Zenodo, though inference requires cloning the repository rather than a packaged library. The retrieval numbers also show the limits of pairwise alignment: text is the weakest-aligned modality, and OneProt-5's alignment scores sit below OneProt-4's, indicating that more modalities need more training rather than less.