Structure-aware deep learning model for peptides toxicity prediction

Table of Contents

• About
• Files
• Installation
• Dependencies
• Inputs
• Usage
• Acknowledgement
• Citation

About

tAMPer is a multi-modal deep learning model that predicts peptide toxicity by integrating the underlying amino acid sequence composition and the predicted three-dimensional (3D) structure. tAMPer adopts a graph-based representation for peptides, encoding their ColabFold-predicted structures. The model extracts structural features using graph neural networks, and employs recurrent neural networks to capture sequential dependencies. The self-attention mechanism is utilized to integrate features from both modalities and weighs the contribution of each amino acid in predicting toxicity.

Files

The project contains the following files and directories:

• checkpoints/: Directory to store trained model checkpoints.
• data/: Directory for storing datasets.
• logs/: Directory for storing log files.
• results/: Directory for storing prediction results.
• src/GConvs.py: Implementation of graph convolutional layers.
• src/embeddings.py: Generating amino acids embeddings.
• src/aminoacids.py: Amino acid-related utility functions.
• src/dataset.py: Dataset loading and preprocessing.
• src/peptideGraph.py: Creating graph representations of peptides.
• src/predict_tAMPer.py: Script for making toxicity predictions on new data.
• src/tAMPer.py: Main implementation of the tAMPer model.
• src/train_tAMPer.py: Training pipeline.
• src/utils.py: General utility functions.

Installation

1. Clone this repository:

   git clone https://github.com/bcgsc/tAMPer.git

2. Navigate to the project directory:

   cd tAMPer

Conda

3. Create a conda environment (optional but recommended):

   conda env create -f environment.yml
   conda activate tAMPer

Pip

3. Create a python virtual environment (make sure python3 is available/loaded):

   pip install --upgrade pip # upgrade pip if necessary
   pip install virtualenv
   virtualenv ENV_ADDRESS
   source ENV_ADDRESS/bin/activate
   pip install -r requirements.txt

Dependencies

• DSSP
• python (>= 3.9)
• pytorch (>= 1.13.1)
• torch-geometric
• torch-cluster
• torch-scatter
• fair-esm (>= 2.0.0)
• numpy
• pandas
• biopython
• wandb (optional)
• loguru

Inputs

Provide the peptide sequences in the FASTA format (.faa). If a peptide has C-terminal amidation, add _AMD at the end of the peptide name.

Example FASTA Format

>peptide_1
MKALIKLPGNRVNGFGRIGR
>peptide_2_AMD
ALWKTLLKKVLKAAA
>peptide_3
GRRPLLLRAR

3D structures

Provide the directory where the ouput of the ColabFold (.result.zip files) is stored. To run ColabFold, please refer to https://github.com/sokrypton/ColabFold. The name of each file should match to its correspoding sequence in the fasta file. Also, add zip_results by ticking its corresponding box which is located within the advanced settings section of AlphaFold2.

If you are using localcolabfold (https://github.com/YoshitakaMo/localcolabfold) for structure predictions, please ensure to include the --amber flag for structure refinement (relaxation / energy minimization) and --zip flag which stores the results in a zip file (.result.zip) in order to utilize tAMPer.

structures
├── peptide_1.result.zip
├── peptide_2_AMD.result.zip
├── peptide_3.result.zip
...

Usage

PROGRAM: train_tAMPer.py & predict_tAMPer.py

USAGE(S): 
   
   ######## TRAIN ##########

   usage: train_tAMPer.py [-h] -tr_pos TR_POS -tr_neg TR_NEG -tr_pdb TR_PDB -val_pos VAL_POS -val_neg VAL_NEG -val_pdb VAL_PDB [-lr LR] [-hdim HDIM] [-gru_layers GRU_LAYERS] [-embedding_model EMBEDDING_MODEL]
                [-modality MODALITY] [-gnn_layers GNN_LAYERS] [-batch_size BATCH_SIZE] [-n_epochs N_EPOCHS] [-gard_acc GARD_ACC] [-weight_decay WEIGHT_DECAY] [-d_max D_MAX] [-lammy LAMMY]
                [-monitor MONITOR] [-pre_chkpnt PRE_CHKPNT] [-chkpnt CHKPNT] [-log LOG]


   train_tAMPer.py script runs tAMPer for training.

   options:
		-h, --help            show this help message and exit
		-tr_pos TR_POS        training toxic sequences fasta file (.faa)
		-tr_neg TR_NEG        training non-toxic sequences fasta file (.faa)
		-tr_pdb TR_PDB        address directory of train structures
		-val_pos VAL_POS      validation toxic sequences fasta file (.faa)
		-val_neg VAL_NEG      validation non-toxic sequences fasta file (.faa)
		-val_pdb VAL_PDB      address directory of val structures
		-lr LR                learning rate
		-hdim HDIM            hidden dimension of model for h_seq and h_strct
		-gru_layers GRU_LAYERS
		                    number of GRU Layers
		-embedding_model EMBEDDING_MODEL
		                    different variant of ESM2 embeddings: {t6, t12, t30, t33, t36, t48}
		-modality MODALITY    Used modality
		-gnn_layers GNN_LAYERS
		                    number of GNNs Layers
		-batch_size BATCH_SIZE
		                    batch size
		-n_epochs N_EPOCHS    max number of epochs
		-gard_acc GARD_ACC    gradient accumulation steps
		-weight_decay WEIGHT_DECAY
		                    weight decay
		-d_max D_MAX          max distance to consider two connect two residues in the graph
		-lammy LAMMY          lammy in the objective function
		-monitor MONITOR      the metric to monitor for early stopping during training
		-pre_chkpnt PRE_CHKPNT
		                    address of pre-trained GNNs
		-chkpnt CHKPNT        address to where trained model be stored
		-log LOG              address to where log file be stored


   ######## PREDICT ##########

   usage: predict_tAMPer.py [-h] -seqs SEQS -pdbs PDBS [-hdim HDIM] [-embedding_model EMBEDDING_MODEL] [-d_max D_MAX] [-chkpnt CHKPNT] [-out OUT]

   predict_tAMPer.py script runs tAMPer for prediction.

   options:
		-h, --help            show this help message and exit
		-seqs SEQS            sequences fasta file for prediction (.fasta)
		-pdbs PDBS            address directory of train structures
		-hdim HDIM            hidden dimension of model for h_seq and h_strct
		-embedding_model EMBEDDING_MODEL
		                    different variant of ESM2 embeddings: {t6, t12}
		-d_max D_MAX          max distance to consider two connect two residues in the graph
		-chkpnt CHKPNT        address of .pt checkpoint to load the model
		-out OUT
		                    address of output folder
                                                                             
EXAMPLE(S):

	train_tAMPer.py -tr_pos ../tAMPer/data/sequences/tr_pos.faa \
		-tr_neg ../tAMPer/data/sequences/tr_pos.faa \
		-tr_pdb ../tAMPer/data/tr_structures/ \
		-val_pos ../tAMPer/data/sequences/tr_pos.faa \
		-val_neg ../tAMPer/data/sequences/tr_pos.faa \
		-val_pdb ../tAMPer/data/val_structures/ \
		-pre_chkpnt ../tAMPer/checkpoints/trained/pre_GNNs.pt \
		-lr 0.0004 \
		-hdim 64 \
		-gru_layers 1 \
		-gnn_layers 1 \
		-d_max 12 \
		-embedding_model t12 \
		-batch_size 32 \
		-n_epochs 100 \
		-gard_acc 1 \
		-weight_decay 1e-7 \
		-lammy 0.2 \
		-chkpnt ../tAMPer/checkpoints/chkpnt.pt \
		-log ../tAMPer/logs/log.npy
      
	predict_tAMPer.py -seqs ../data/sequences/seqs.faa \
		-pdbs ../tAMPer/data/structures/ \
		-hdim 64 \
		-embedding_model t12 \
		-d_max 12 \
		-chkpnt ../tAMPer/checkpoints/trained/chkpnt.pt \
		-out ../tAMPer/results/prediction.csv
      

Acknowledgement

The implementation of portions of the GNNs convolutional layers and the input data pipeline were adapted from Jing et al, ICLR 2021 and Baldassarre et al, Structural bioinformatics 2021.

Citation

If you use tAMPer in your research, please cite:

Ebrahimikondori H, Sutherland D, Yanai A, Richter A, Salehi A, Li C, Coombe L, Kotkoff M, Warren RL, Birol I. 2024. Structure-aware deep learning model for peptide toxicity prediction. Protein Science. https://doi.org/10.1002/pro.5076.