Alignment-free structure prediction using protein language models
The prediction pipeline uses Python3 and requires the following modules:
- numpy
- matplotlib
- torch (1.9.0 recommended)
- transformers (4.6.0 recommended)
The adapted trRosetta folding pipeline additionally requires pyRosetta to be installed.
Clone the repository and install the dependencies listed above.
The ProtT5 protein language model will be downloaded automatically on first use.
For a FASTA file containing one or more protein sequences and an output directory of your choice, run the pipeline via
python predict.py -i <FASTA_file> -o <output_directory>
The ProtT5 model will be downloaded on first use and stored by default in the directory 'ProtT5-XL-U50'. You can change this directory with the --t5_model parameter.
You can trade speed with prediction quality by modifying the cropping stride used during inference (default: 16) with the --stride parameter (see publication for details).
If you run out of GPU memory and/or want to compute predictions for long protein sequences, you might want to lower the default batch-size of 200 with the --batch_size parameter.
You can create a PDB structure from a predicted distogram using the adapted trRosetta folding scripts in the 'folding' directory:
python trRosetta.py -m 0 -pd 0.05 <distogram_file> <FASTA_file> output.pdb
Please note that the FASTA file for the folding script should only contain a single sequence corresponding to the distogram. It is recommended to create multiple decoys with different cutoffs (-pd [0.05, 0.5]) and modes (-m {0,1,2}). Please refer to trRosetta for additional details on the folding pipeline.
Predictions for all human proteins smaller than 3000 residues are available at EMBER2_human.
Konstantin Weißenow, Michael Heinzinger, Burkhard Rost
Technical University Munich
Weissenow, K., Heinzinger, M., Rost, B.
Protein language model embeddings for fast, accurate, and alignment-free protein structure prediction.
Structure (2022) link