| --- |
| 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 |
|
|
