Protein Retrieval Foundation Model

Cryo-electron microscopy now produces protein density maps faster than the function of those proteins can be annotated. A structural biologist holding a newly resolved map often wants a simple thing: show me proteins that look like this and tell me what they do. Sequence search does not help when only a map is in hand, and geometric structure comparison returns neighbors without saying anything about function.

This model, from Carnegie Mellon University and Purdue University, attacks the problem the way vision-language models attack image search. It is a CLIP-style dual encoder that embeds a 3D protein structure and a natural-language description of that protein's function into a shared space, trained by contrastive learning so that matched structure-text pairs are close and mismatched pairs are far apart. At inference, a query structure is embedded once and ranked against a database by cosine similarity, returning both functionally related proteins and the text annotations attached to them, with no task-specific fine-tuning. The work was accepted to the CVPR 2025 Workshop on Multimodal Foundation Models for Biomedicine.

The model is a retrieval counterpart to text-protein alignment systems such as ProteinDT, which pairs sequence with text; here the visual modality is the protein's 3D surface, which is what an experimental density map actually gives you.

#Key Features

• Structure-to-function retrieval: A query structure returns functionally similar proteins and their annotations, turning an unannotated map into a ranked list of functional hypotheses rather than a list of geometric neighbors.
• Cross-database zero-shot transfer: The model is trained on PDB-derived structures but evaluated on experimentally determined EMDB density maps that it never saw during pretraining, testing whether the shared embedding space survives the shift from idealized to observed density.
• Surface point-cloud representation: Structures are converted to volumetric density maps and then to surface meshes and point clouds, so PDB entries and experimental cryo-EM maps enter the model through the same representation.
• Leaf-level GO captions: Captions are built from Gene Ontology terms with generic parent terms stripped out, keeping only the most specific leaf-level annotations so the text signal is compact and discriminative.

#Technical Details

The structure branch uses I2P-MAE, a 3D point-cloud encoder pretrained with supervision from a 2D vision transformer, initialized from a Point-Bind checkpoint. The text branch uses all-mpnet-base-v2 from SentenceTransformers and is kept frozen throughout training. Structures from the PDB are converted to density maps with EMAN2's pdb2mrc, surfaces are extracted by marching cubes after intensity normalization, and 16,384 points are uniformly sampled per protein. EMDB entries are processed identically, using each entry's author-recommended contour level as the iso-value. The training corpus is approximately 200,000 protein-caption pairs.

Training runs for 40 epochs with AdamW at a 1e-3 initial learning rate under cosine annealing, a global batch size of 64 across four NVIDIA A40 GPUs, with GO-term shuffling and geometric jitter, rotation, and translation as augmentations. On in-domain zero-shot PDB retrieval the model reaches 35.18% Top-1 and 59.70% Top-5 accuracy. On the harder cross-database EMDB-to-PDB setting it reaches roughly 17% Top-1 and 36% Top-5 at a combined cross-correlation and overlap threshold of 1.5. The paper reports no comparison against prior retrieval baselines.

#Applications

The natural workflow is cryo-EM structure interpretation: a map is solved, its surface is embedded, and the nearest annotated neighbors in PDB space suggest what the protein family and molecular function are likely to be. The same mechanism supports functional annotation of predicted structures, structure-based literature and database search where sequence homology is too weak to be informative, and exploratory browsing of large structure repositories by functional query text. Groups building automated pipelines around structure determination are the immediate beneficiaries, since retrieval runs at embedding-comparison speed over precomputed databases.

#Impact

The contribution is mainly one of framing: it establishes multimodal structure-text retrieval as a task, supplies a 200,000-pair dataset construction recipe for it, and shows that a straightforward CLIP transplant produces non-trivial zero-shot accuracy on experimentally observed density maps. The results are early-stage — this is a four-page workshop paper, absolute Top-1 accuracy is modest, the cross-database numbers drop sharply, and the absence of baseline comparisons makes it hard to place against structure-search tools. No code, weights, or dataset release accompanies the preprint. Its value at present is as a demonstration that the vision-language retrieval recipe transfers to protein surfaces at all, and as a template others can build stronger encoders into.