ProtFlow
January 19, 2026 ยท View on GitHub
Environment creation
conda env create -f environment.yaml
Ensure all paths and hyperparameters in this project are defined according to your specific tasks. Download facebook/esm2_t12_35M_UR50D, Rostlab/prot_t5_xl_uniref50 and Rostlab/prot_t5_xl_half_uniref50-enc from hunggingface and add them at the root folder.
Data Preparation
create a data folder and convert your specific protein sequences into fasta format, and split it into train.fasta and valid.fasta
Training
1. Compute statistics
python -m utils.get_statistics
python -m utils.stat
2. Decoder training
python train_decoder.py
3. Compression-decompression module training
python train_compressor.py
4. FM holder training
torchrun --nproc_per_node=1 --master_port=31345 train_flow_matching.py
Generation
torchrun --nproc_per_node=1 --master_port=31345 generation.py