Home

Awesome

ProCALM

image ProCALM (Protein Conditionally Adapted Language Model) is an autoregressive transformer where ProGen2-base is finetuned with conditional adapters for conditional generation of functional enzymes, based on EC number, taxonomy, or both. Model here is refered to as progen-conditional.

Setup and Installation

We have provided docker/Dockerfile to build a docker image to run ProCALM. Alternatively, model will run in a conda environment created using docker/environment.yml.

Dataset Processing

For convenience, processed datasets used for model training and evaluation (excluding the Uniref dataset) are preloaded in this repository.

Alternatively, you can follow these steps to reproduce our dataset processing. The raw data used for training ProCALM is obtained from Uniprot and other useful files are from the CARE benchmarks. You will need all enzymes with EC numbers and taxonomy lineages from uniprot for Uniref (>20M) and Swissprot (~200k). Use the versions from June 17, 2024 and export EC number, sequence, entry ID, taxonomy, taxonomy ID, lineage, lineage ID as a tsv. Raw data is alternatively uploaded here and can fill in the data/raw_data folder. The database_setup.py script will run sequence clustering on Swissprot (requires MMseqs2) and will build Diamond reference databases along with MMseqs2 databases including taxonomy (requires Diamon BLAST). For building the Diamond reference database, you will also need swissprot.fasta file (~550k sequences from June 17, 2024 download). The latter two reference databases are only needed for analysis of generated sequences.

The final splits used in our study are built and saved to sharded datasets using build_splits.ipynb under scripts. save_sharded.ipynb is a shortcut to save sharded datasets from presaved csv files of each split (in particular the sharded version of uniref). You can also use select_ECs.ipynb to reproduce how we generated the EC numbers that are used for generation and analysis.

Training

Example command for running training:

composer main.py --config config/long-final/ec-onehot-swissprot.yml --debug

Results will be saved under results/. Training will use all available GPUs with DDP. Training with 4 40GB A100s should take on the order of 6 hours for every 1 billion tokens. Other configs to reproduce the results from our study are given under config/long-final/

Generation

Some of the pretrained models are loaded to Huggingface Hub, allowing for easy generation, with corresponding generated sequences preloaded to this respository. Alternatively, example commands to generate sequences with different conditioning:

#generate for a single EC conditioning
python runner.py --model ec-onehot-swissprot --checkpoint 1.5B --ec 4.2.1.20

#generate for EC and taxonomy conditioning
python runner.py --model ec+tax-swissprot --checkpoint 9B --ec 4.2.1.20 --tax bacteria --num_seqs 225

#ZymCTRL baseline for EC 4.2.1.20
python scripts/ZymCTRL_runner.py

Generated sequences will be saved under results/{model_name}/generated/ as fastas. Generation on a single GPU should take on the order of 1 minute per batch, so 990 sequences will take around 20 minutes on a single 40GB A100. Note that ba1100, ba21000, ba42000, and ba63000 refer to checkpoints trained to 1.5, 3, 6, and 9 billion tokens, respectively.

Analysis

CSV files summarizing the quality of generated sequences are preloaded to this repository. Alternatively, example command to get statistics on the generation quality and diversity of generated sequences:

python scripts/run_generation_processing.py --model ec+tax-swissprot --checkpoint 9B --ec 4.2.1.20 --tax bacteria

Results will be saved under results/{model_name}/all_summary_{checkpoint}.csv. Analysis will fun faster on more CPUs but should take on the order of seconds to one minute per combination of unique conditioning (EC/tax).

Perplexities for different splits can be calculated with perplexity_calculation.py under scripts and outputs will be saved under results/{model_name}/perplexity_{checkpoint}.csv. Perplexity calculation requires a single A100 and should take on the order of minutes to one hour for all datasets on a single checkpoint.

The output csvs can be visualized using analysis/visualization.ipynb.

Limitations

Better logging for losses as training progresses could be implemented.