trill-proteins 0.2.4

Sandbox (in progress) for Computational Protein Design

                          _____________________.___.____    .____     
                          \__    ___/\______   \   |    |   |    |    
                            |    |    |       _/   |    |   |    |    
                            |    |    |    |   \   |    |___|    |___ 
                            |____|    |____|_  /___|_______ \_______ \
                                             \/            \/       \/

TRILL

TRaining and Inference using the Language of Life

Set-Up

1. I recommend using a virtual environment with conda, venv etc.
2. Run $ pip install trill-proteins torch pytorch-lightning
3. $ pip install pyg-lib torch-scatter torch-sparse torch-cluster torch-spline-conv torch-geometric -f https://data.pyg.org/whl/torch-1.13.0+cu117.html

Examples

Default (Fine-tuning ESM2)

1. The default mode for TRILL is to just fine-tune the base esm2_t12_35M_UR50D model from FAIR with the query input.

$ trill fine_tuning_ex 1 --query data/query.fasta

Embed with base esm2_t12_35M_UR50D model

2. You can also embed proteins with just the base model from FAIR and completely skip fine-tuning. The output will be a CSV file where each row corresponds to a single protein with the last column being the fasta header.

$ trill base_embed 1 --query data/query.fasta --noTrain

Embedding with a custom pre-trained model

3. If you have a pre-trained model, you can use it to embed sequences by passing the path to --preTrained_model.

$ trill pre_trained 1 --query data/query.fasta --preTrained_model /path/to/models/pre_trained_model.pt

Distributed Training/Inference

4. In order to scale/speed up your analyses, you can distribute your training/inference across many GPUs with a few extra flags to your command. You can even fit models that do not normally fit on your GPUs with sharding, CPU-offloading etc. Below is an example slurm batch submission file. The list of strategies can be found here (https://pytorch-lightning.readthedocs.io/en/stable/extensions/strategy.html). The example below utilizes 16 GPUs in total (4(GPUs) * 4(--nodes)) with Fully Sharded Data Parallel and the 650M parameter ESM2 model.

#!/bin/bash
#SBATCH --time=8:00:00   # walltime
#SBATCH --ntasks-per-node=4
#SBATCH --nodes=4 # number of nodes
#SBATCH --gres=gpu:4 # number of GPUs
#SBATCH --mem-per-cpu=60G   # memory per CPU core
#SBATCH -J "tutorial"   # job name
#SBATCH --mail-user="" # change to your email
#SBATCH --mail-type=BEGIN
#SBATCH --mail-type=END
#SBATCH --mail-type=FAIL
#SBATCH --output=%x-%j.out
master_addr=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
export MASTER_ADDR=$master_addr
export MASTER_PORT=13579

srun trill distributed_example 4 --query data/query.fasta --nodes 4 --strategy fsdp --model esm2_t33_650M_UR50D

You can then submit this job with:

$ sbatch distributed_example.slurm

More examples for distributed training/inference without slurm coming soon!

Generating protein sequences using inverse folding with ESM-IF1

5. When provided a protein backbone structure (.pdb, .cif), the IF1 model is able to predict a sequence that might be able to fold into the input structure. The example input are the backbone coordinates from DWARF14, a rice hydrolase. For every chain in the structure, 2 in 4ih9.pdb, the following command will generate 3 sequences. In total, 6 sequences will be generated.

$ trill IF_Test 1 --query data/4ih9.pdb --if1 --genIters 3

Generating Proteins using ProtGPT2

6. You can also generate synthetic proteins using ProtGPT2. The command below generates 5 proteins with a max length of 100. The default seed sequence is "M", but you can also change this. Check out the command-line arguments for more details.

$ trill Gen_ProtGPT2 1 --protgpt2 --gen --max_length 100 --num_return_sequences 5

Fine-Tuning ProtGPT2

7. In case you wanted to generate certain "types" of proteins, below is an example of fine-tuning ProtGPT2 and then generating proteins with the fine-tuned model.

$ trill FineTune 2 --protgpt2 --epochs 100

$ trill Gen_With_FineTuned 1 --protgpt2 --gen --preTrained_model FineTune_ProtGPT2_100.pt

Arguments

Positional Arguments:

1. name (Name of run)
2. GPUs (Total # of GPUs requested for each node)

Optional Arguments:

• -h, --help (Show help message)
• --query (Input file. Needs to be either protein fasta (.fa, .faa, .fasta) or structural coordinates (.pdb, .cif))
• --nodes (Total number of computational nodes. Default is 1)
• --lr (Learning rate for adam optimizer. Default is 0.0001)
• --epochs (Number of epochs for fine-tuning transformer. Default is 20)
• --noTrain (Skips the fine-tuning and embeds the query sequences with the base model)
• --preTrained_model (Input path to your own pre-trained ESM model)
• --batch_size (Change batch-size number for fine-tuning. Default is 1)
• --model (Change ESM model. Default is esm2_t12_35M_UR50D. List of models can be found at https://github.com/facebookresearch/esm)
• --strategy (Change training strategy. Default is None. List of strategies can be found at https://pytorch-lightning.readthedocs.io/en/stable/extensions/strategy.html)
• --logger (Enable Tensorboard logger. Default is None)
• --if1 (Utilize Inverse Folding model 'esm_if1_gvp4_t16_142M_UR50' to facilitate fixed backbone sequence design. Basically converts protein structure to possible sequences)
• --temp (Choose sampling temperature. Higher temps will have more sequence diversity, but less recovery of the original sequence for ESM_IF1)
• --genIters (Adjust number of sequences generated for each chain of the input structure for ESM_IF1)
• --LEGGO (Use deepspeed_stage_3_offload with ESM. Will be removed soon...)
• --profiler (Utilize PyTorchProfiler)
• --protgpt2 (Utilize ProtGPT2. Can either fine-tune or generate sequences)
• --gen (Generate protein sequences using ProtGPT2. Can either use base model or user-submitted fine-tuned model)
• --seed_seq (Sequence to seed ProtGPT2 Generation)
• --max_length (Max length of proteins generated from ProtGPT)
• --do_sample (Whether or not to use sampling ; use greedy decoding otherwise)
• --top_k (The number of highest probability vocabulary tokens to keep for top-k-filtering)
• --repetition_penalty (The parameter for repetition penalty. 1.0 means no penalty)
• --num_return_sequences (Number of sequences for ProtGPT2 to generate)

Misc. Tips

• Make sure there are no "*" in the protein sequences
• Don't run jobs on the login node, only submit jobs with sbatch or srun on the HPC
• Caltech HPC Docs https://www.hpc.caltech.edu/documentation