InstructPLM

Design protein sequences following structure instructions. Read the InstructPLM paper.

Setup

We recommend using docker for a quick start. You can launch an instance of instructPLM with the following commands:

docker pull jundesiat/instructplm:mpnn-progen2-xlarge
docker run --gpus all -it -v /path/to/input_output:/workspace/ jundesiat/instructplm:mpnn-progen2-xlarge
cd /root/InstructPLM

Or you can run InstructPLM from the source code, clone this repo and install dependence:

git clone --recurse-submodules https://github.com/Eikor/InstructPLM.git
cd InstructPLM
pip install -r requirements.txt

Usage

Code organization:

• run_eval.py - gives a minimal code of model evaluation (LM-Loss and perplexity).
• run_generate.py - example of generate protein sequence.
• recovery_rate.py - example of calculating recovery rate of generated sequences.
• pdbs - input PDB files.
• structure_embeddings - input preprocessed structure embeddings.

[!IMPORTANT] Make sure you have obtained structure embedding before running InstructPLM, you can construct preprocessed structure embeddings by python structure_embeddings/preprocess.py. This script will process protein pdbs stored in pdbs/ and save the result in structure_embeddings/.

Protein Design

For protein design, run python run_generate.py --total 10 --save_suffix test. This script will read embeddings automatically in structure_embeddings/ and save the result at the path specified by --save_prefix. For generating fix-length proteins, setting --fix_length=True.

[!TIP] Large language models some times suffer from Hallucinations, so as pLMs :thinking: . You may need to generate a large set of candidates and a select policy (e.g.,