SPOT-Contact-LM

Model checkpoints for SPOT-Contact-LM (also called SPOT-Contact-Single), which predicts a protein's residue–residue contact map from a single sequence. It uses ESM-1b attention maps together with SPOT-1D-Single predictions and one-hot encoding, instead of multiple sequence alignments.

Code and instructions: https://github.com/jas-preet/SPOT-Contact-LM

Files

File	Size	SHA-256
contact_jits.tar.xz	101 MB	fb68555f977543f0f31e30d845f9c927cfb1216fd63dd263bc86a93feecf9e81

The archive holds TorchScript checkpoints for an ensemble of six 12-block ResNets: three predict contacts directly and three predict distance bins, using ESM-1b attention from the last layer, from all layers, or from all layers plus one-hot and SPOT-1D-Single features.

git clone https://github.com/jas-preet/SPOT-Contact-LM.git
cd SPOT-Contact-LM
curl -LO https://huggingface.co/Jas-preet/SPOT-Contact-LM/resolve/main/contact_jits.tar.xz
echo "fb68555f977543f0f31e30d845f9c927cfb1216fd63dd263bc86a93feecf9e81  contact_jits.tar.xz" | shasum -a 256 -c
tar -xf contact_jits.tar.xz

Training data

Trained on 34,691 proteins from ProteinNet (PDB entries before 2016, resolution < 2.5 Å, clustered at 95% sequence identity, up to 500 residues). A contact is a Cβ–Cβ distance of at most 8 Å.

Citation

Singh J, Litfin T, Singh J, Paliwal K, Zhou Y. SPOT-Contact-LM: improving single-sequence-based prediction of protein contact map using a transformer language model. Bioinformatics. 2022;38(7):1888-1894. doi:10.1093/bioinformatics/btac053