File size: 2,471 Bytes
dc5b3f3 c8229ae dc5b3f3 4dc4aad dc5b3f3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 | ---
license: mit
tags:
- biology
- protein
- enzyme
- ec-number
- interpretability
---
# BioBlobs
The BioBlobs model behind the M-CSA results of **Unsupervised discovery of Functional
Substructures** (Xin Wang, Kaiwen Shi and Carlos Oliver, NeurIPS 2026).
[Paper](https://arxiv.org/abs/2510.01632) · [Code](https://github.com/OliverLaboratory/BioBlobs)
The model predicts third-level Enzyme Commission (EC) classes, 242 in total, from frozen ESM2-T30
residue embeddings and the protein structure. It splits each protein into 8 spatially local blobs
and quantizes their embeddings with residual vector quantization. The blobs it relies on most
point to candidate catalytic sites, although training uses only protein-level EC labels.
## Files
| File | Contents |
| --- | --- |
| `model.safetensors` | model weights (59 MB) |
| `config.yaml` | model configuration |
| `classes.json` | EC class names and their indices |
| `training_log.jsonl` | training log |
## Use
```bash
git clone https://github.com/OliverLaboratory/BioBlobs.git
cd BioBlobs
uv sync && source .venv/bin/activate
hf download xwang112358/BioBlobs --local-dir bioblobs-mcsa-rvq-k8
python inference.py --model bioblobs-mcsa-rvq-k8 --pdb-dir my_pdbs --out results
```
For each protein, the output lists the top EC classes, the blobs ranked by importance score with
their residue numbers, and per-residue scores. The repository's `docs/PREDICT.md` describes the
options and outputs, and `docs/REPRODUCE.md` shows how to retrain the model and evaluate it on
M-CSA.
## Training data
Swiss-Prot enzymes from the [EC-Bench](https://github.com/dsaeedeh/EC-Bench) training set, with
AlphaFold DB v6 structures. Training keeps proteins of up to 1,000 residues with a complete
third-level EC number, split into training and validation sets at 30% sequence identity with
MMseqs2. The GitHub repository ships the training and validation ids.
## Limitations
- The model only knows the 242 EC classes in its training data.
- Class probabilities are softmax scores, not calibrated confidence.
- A blob that overlaps annotated catalytic residues is evidence for, not proof of, its role in
catalysis.
## Citation
```bibtex
@inproceedings{wang2026bioblobs,
title = {Unsupervised discovery of Functional Substructures},
author = {Wang, Xin and Shi, Kaiwen and Oliver, Carlos},
booktitle = {Advances in Neural Information Processing Systems},
year = {2026}
}
```
## License
MIT
|