IEEE BIBM 2024

ProtSolM

Protein solubility prediction with multi-modal features — fusing sequence, structure, and hand-crafted physicochemical descriptors for more reliable solubility labels.

Yang Tan, Jia Zheng, Liang Hong, Bingxin Zhou

Paper Code Dataset

What the model does

ProtSolM predicts protein solubility by combining three complementary views: sequence context, 3D structure encodings, and explicit physicochemical / structural feature vectors (amino-acid composition, GRAVY, secondary-structure composition, hydrogen bonds, exposed-residue fraction, pLDDT, and related signals).

The structure pathway builds on pretrained ProtSSN checkpoints; the recommended default for fine-tuning is k20_h512.

Benchmarks and data

Evaluation centers on PDBSol with an ExternalTest hold-out. ESMFold PDBs for both sets are released on Hugging Face (tyang816/ProtSolM_ESMFold_PDB); labels ship with the repository under data/PDBSol and data/ExternalTest.

Feature extraction (get_feature.py), fine-tuning (run_ft.py), and evaluation (eval.py) scripts support both the paper splits and custom PDB + CSV datasets.

Citation

If you use ProtSolM, please cite:

@inproceedings{tan2024protsolm,
  title={Protsolm: Protein solubility prediction with multi-modal features},
  author={Tan, Yang and Zheng, Jia and Hong, Liang and Zhou, Bingxin},
  booktitle={2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM)},
  pages={223--232},
  year={2024},
  organization={IEEE}
}