PSICHIC: physicochemical graph neural network for learning protein-ligand interaction fingerprints from sequence data [Preprint]

Environment Setup

Currently, PSICHIC is validated for use on MacOS (OSX), Linux and Windows. We recommend installation via conda, or even better, using the faster mamba package and environment manager. Mamba can be installed with the command conda install mamba -n base -c conda-forge. For setup using either conda or mamba, please refer to the relevant code line provided below.

## OSX 
conda env create -f environment_osx.yml  # if mamba: mamba env create -f environment_osx.yml
## LINUX or Windows GPU
conda env create -f environment_gpu.yml # if mamba: mamba env create -f environment_gpu.yml
conda activate psichic_fp
pip install torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.1.0+cu118.html
## LINUX or Windows CPU
conda env create -f environment_cpu.yml  # if mamba: mamba env create -f environment_cpu.yml
conda activate psichic_fp
pip install torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.1.0+cpu.html

Alternatively, command lines that can be helpful in setting up the environment (tested on linux with python 3.8).

conda create --name psichic_fp python=3.8
conda install pytorch==2.0.0 torchvision==0.15.0 torchaudio==2.0.0 pytorch-cuda=11.7 -c pytorch -c nvidia
conda install pyg -c pyg
conda install -c conda-forge rdkit==2022.09.5
pip install scipy biopython pandas biopandas timeout_decorator py3Dmol umap-learn plotly mplcursors lifelines reprint
pip install "fair-esm"

Virtual Screening Platform

• Only Sequence Data: Protein Sequence + Ligand SMILES pairs is all you need.
• Quick Screening: Up to 100K compounds in an hour.
• Deep Analysis: Uncover molecular insights with PSICHIC-powered pharmacophore and targeted mutagenesis analysis.

Start exploring. Your next discovery could be just clicks away!

UPDATE: We now included selectivity online platform (beta version) in selectivity subfolder that demonstrate how PSICHIC can be used for selectivity profiling.

BYO-PSICHIC with Annotated Sequence Data

Create a train, valid and test csv file in a datafolder (for examples, see the dataset folder). The datafolder should contain at least a train.csv and test.csv file. Depending on your annotated labels, you want to use --regression_task True if it is a continuous value label (e.g., binding affinity), --classification_task True if it is a binary class label (e.g., presence of interaction) and --mclassification_task C where C represents the number of classes in your multiclass labels (e.g., 3 if you are using our protein-ligand functional response dataset). Note, you can have a dataset with multiple label types and we will train PSICHIC on predicting multiple protein-ligand interaction properties (see PSICHIC-MultiTask below)

python main.py --datafolder annotated_folder --result_path result/annotated_result --regression_task True 

BYO-PSICHIC using a benchmark dataset, for example, the PDBBind v2020 benchmark:

python main.py --datafolder dataset/pdb2020 --result_path result/PDB2020_BENCHMARK --regression_task True 

Model and optimizer configurations are consistent across all benchmark datasets, except PDBBind v2016 where you want to change the optimizer's number of training iterations, betas and eps to 30000, "(0.9,0.99)" and 1e-5 respectively, i.e. add to the commandline: --total_iters 30000 --betas "(0.9,0.99)" --eps 1e-5. For binary classification task, replace --regression_task True to --classification_task True. For protein functional effect dataset, replace --regression_task True to --mclassification_task 3. Feel free to adjust the model hyperparameters in the config.json file, let us know if you find any interesting results!

Dataset Structure and BYO Formatting Guidelines

All datasets referenced in our manuscript are available on Google Drive (Dataset). For the datasets used in the benchmark evaluation of PSICHIC, we have train, valid, and test CSV files that have been created based on established split settings. A separate README.md in the dataset section is dedicated to explaining the purpose of each dataset in the Google Drive Link (this is similar to Extended Data Table 1 in our manuscript).

BYO-PSICHIC Dataset: Each file should look something like this if you are interested in training BYO-PSICHIC. A validation CSV file is not required if you don't have one, for instance, if you plan to apply the results in external experiments.

Binding Affinity Regression

Protein	Ligand	regression_label
ATCGATCG....	C1CCCCC1	6.7
GCTAGCTA....	O=C(C)Oc1ccccc1C(=O)O	4.0
...	...	...
TACGTACG	CCO	8.1

python main.py --datafolder BYO_DATASET --result_path BYO_RESULT --regression_task True 

Binary Interaction Classification

Protein	Ligand	classification_label
ATCGATCG....	C1CCCCC1	1
GCTAGCTA....	O=C(C)Oc1ccccc1C(=O)O	0
...	...	...
TACGTACG	CCO	1

python main.py --datafolder BYO_DATASET --result_path BYO_RESULT --classification_task True

Functional Effect Classification (Three-way Classification)

Protein	Ligand	multiclass_label
ATCGATCG....	C1CCCCC1	-1
GCTAGCTA....	O=C(C)Oc1ccccc1C(=O)O	0
...	...	...
TACGTACG	CCO	1

python main.py --datafolder BYO_DATASET --result_path BYO_RESULT --mclassification_task 3

Multi Task PSICHIC

Protein	Ligand	regression_label	multiclass_label
ATCGATCG....	C1CCCCC1	6.7	-1
GCTAGCTA....	O=C(C)Oc1ccccc1C(=O)O	4.0	0
...	...	...
TACGTACG	CCO	8.1	1

python main.py --datafolder BYO_DATASET --result_path BYO_RESULT --regression_task True --mclassification_task 3

Strategically Split Your Dataset? Jupyter notebook in dataset folder is available to illustrate how we perform random splits, unseen protein splits, and unseen ligand scaffold splits to evaluate the generalizability of PSICHIC or other methods. This can be useful in evaluating whether the BYO-PSICHIC works on your annotated sequence data.

PSICHIC_XL: Multitask Prediction Training on Large-scale Interaction Dataset

The PSICHIC_XL was previously referred to as the pre-trained multi-task PSICHIC. The PSICHIC_A1R was previously referred to as the fine-tuned multi-task PSICHIC. We changed the name to clarify that PSICHIC_XL can be used as is without any additional training. However, PSICHIC_XL can potentially improve its ranking capabilities in virtual screening when fine-tuned on data specific to a protein target, e.g., the PSICHIC_A1R we show below using A₁R-related data.

Training PSICHIC_XL (AKA Pre-trained PSICHIC in Preprint)

python main.py --datafolder dataset/large_scale_interaction_dataset --result_path PSICHIC_MultiTask_Pretrain --lrate 1e-5 --sampling_col pretrain_sampling_weight --regression_task True --mclassification_task 3 --total_iters 300000 --evaluate_step 25000

Fine-tune PSICHIC_XL into PSICHIC_A1R (AKA Fine-tuned PSICHIC in Preprint)

We finetune only the application layers of PSICHIC_XL for 1000 iteration on A₁R-related protein using the following command:

python main.py --regression_task True --mclassification_task 3 --datafolder dataset/A1R_FineTune --result_path PSICHIC_A1R_FineTune --lrate 1e-5 --total_iters 1000 --finetune_modules "['reg_out','mcls_out']" --trained_model_path trained_weights/multitask_PSICHIC

We have renamed the PSICHIC version trained on the extensive interaction dataset as PSICHIC_XL, and the subset focusing on A₁R data as PSICHIC_A1R. Previously, PSICHIC_XL and PSICHIC_A1R were known as pre-trained PSICHIC and fine-tuned PSICHIC, respectively. This change more accurately reflects PSICHIC_XL's broad applicability and PSICHIC_A1R's specific emphasis on A1R.

For any other proteins, you can filter out irrelevant proteins and the non-binders in large-scale interaction dataset to apply PSICHIC for other experiments.

References

For more information, please refer to our work:

PSICHIC: physicochemical graph neural network for learning protein-ligand interaction fingerprints from sequence data
Huan Yee Koh, Anh T.N. Nguyen, Shirui Pan, Lauren T. May, Geoffrey I. Webb
bioRxiv 2023.09.17.558145; doi: https://doi.org/10.1101/2023.09.17.558145