SPOT-1D-LM
Model checkpoints for SPOT-1D-LM, which predicts protein secondary structure (3- and 8-state), solvent accessibility (ASA), half-sphere exposure, contact number and backbone angles (ฯ, ฯ, ฮธ, ฯ) from a single sequence, using embeddings from two protein language models (ESM-1b and ProtT5) instead of multiple sequence alignments.
Code and instructions: https://github.com/jas-preet/SPOT-1D-LM
Files
| File | Size | SHA-256 |
|---|---|---|
spot_1d_lm.tar.gz |
956 MB | 594e5a140a335cca89640a6bd3c56e8a30e4ebee7e226433b7ce1461721aa97e |
The archive holds a nested archive of TorchScript checkpoints for an ensemble of three models (a bidirectional LSTM, a multi-scale ResNet and a multi-scale ResNet-LSTM), for classification and for regression. ESM-1b and ProtT5 are downloaded separately by their own libraries the first time the pipeline runs.
git clone https://github.com/jas-preet/SPOT-1D-LM.git
cd SPOT-1D-LM
curl -LO https://huggingface.co/Jas-preet/SPOT-1D-LM/resolve/main/spot_1d_lm.tar.gz
echo "594e5a140a335cca89640a6bd3c56e8a30e4ebee7e226433b7ce1461721aa97e spot_1d_lm.tar.gz" | shasum -a 256 -c
tar -xzf spot_1d_lm.tar.gz spot_1d_lm/checkpoints/SPOT-LM-checkpoints.xz
tar -xf spot_1d_lm/checkpoints/SPOT-LM-checkpoints.xz
The archive also contains the training and test data (fasta, dssp, hse and theta folders, plus train, validation and test lists). tar -xzf spot_1d_lm.tar.gz extracts everything.
Training data
Trained on 38,913 proteins from ProteinNet (PDB entries before 2016, resolution < 2.5 ร , clustered at 95% sequence identity). Inputs are limited to 1,024 residues.
Citation
Singh J, Paliwal K, Litfin T, Singh J, Zhou Y. Reaching alignment-profile-based accuracy in predicting protein secondary and tertiary structural properties without alignment. Scientific Reports. 2022;12:7607. doi:10.1038/s41598-022-11684-w