SynProtX

An official implementation of our research paper "SynProtX: A Large-Scale Proteomics-Based Deep Learning Model for Predicting Synergistic Anticancer Drug Combinations".

SynProtX is a deep learning model that integrates large-scale proteomics data, molecular graphs, and chemical fingerprints to predict synergistic effects of anticancer drug combinations. It provides robust performance across tissue-specific and study-specific datasets, enhancing reproducibility and biological relevance in drug synergy prediction.

Setting up environment

We use Miniconda to manage Python dependencies in this project. To reproduce our environment, please run the following script in the terminal:

conda env create -f env.yml
conda activate SynProtX

Downloading raw data

Datasets, hyperparameters, and model checkpoints can be downloaded through .

Using SynProtX for inference

SynProtX allows the prediction of synergistic effects between drug combinations through inference using the SynProtX model. It leverages various tissue-specific and study datasets to make these predictions.

Usage Example

To perform inference, you can run the following command:

python synprotx_inference.py --smi1 "CCOc1ccc2c(c1)N=C(N)N(c3ccc(Cl)cc3)S2" --smi2 "CN1CCC(CC1)Nc2nccc3c2ncn3C" \
--dataset ALMANAC-Breast --cell_line MCF7 --task classification --thr 0.5

In this example:

• --smi1 and --smi2 represent the SMILES strings of the two drug compounds being tested.
• --dataset specifies the dataset to use (e.g., ALMANAC-Breast).
• --cell_line indicates the cell line to consider (e.g., MCF7).
• --task defines the type of task: classification for synergy/antagonism prediction or regression for raw score prediction.
• --thr sets the threshold for classification tasks, used to differentiate between synergistic and antagonistic interactions.

Available Options

Option	Description
--smi1	SMILES string of the first compound (required)
--smi2	SMILES string of the second compound (required)
--cell_line	Cell-line identifier (e.g. MCF7) (required)
--dataset	Dataset to use (default: ALMANAC-Breast). Available options are:
	- For Tissue Datasets: ALMANAC-Breast, ALMANAC-Lung, ALMANAC-Ovary, ALMANAC-Skin
	- For Study Datasets: FRIEDMAN, ONEIL
--task	Task type (default: regression). Options:
	- classification, regression
--device	Device for computation (default: cpu). Options:
	- cpu, cuda:0 (or another CUDA device string)
--thr	Threshold for classifying synergy vs antagonism (only for classification task). Default: 0.5

Dataset availability

Dataset	Cell Lines
ALMANAC-Breast	BT-549, MCF7, MDA-MB-231, MDA-MB-468
ALMANAC-Lung	A549, EKVX, HOP-62, HOP-92, NCI-H226, NCI-H460, NCI-H522
ALMANAC-Ovary	OVCAR-4, OVCAR-5, OVCAR-8, SK-OV-3
ALMANAC-Skin	SK-MEL-2, SK-MEL-5, SK-MEL-28, UACC-257
FRIEDMAN (Skin)	A2058, G-361, IPC-298, RVH-421, SK-MEL-2, SK-MEL-5, SK-MEL-28, UACC-257
ONEIL (Several Tissues)	A2058 (skin), NCI-H460 (lung), SK-OV-3 (ovary), A2780 (ovary), A427 (lung), RKO (large intestine), SW837 (large intestine)

Generating dataset

A tarball will be obtained after download. After file extraction, move all nested folders to the root of this project directory. You might need to move all files in data/export up to data folder. Otherwise, you will run the Jupyter Notebook files to generate mandatory data. Let’s take a look at ipynb folder. Run the following files in order if you want to replicate our exported data.

• 01_drugcomb_clean.ipynb → cleandata_cancer.csv
• 02_CCLE_gene_expression → CCLE_expression_cleaned.csv
• 03_omics_preprocess → protein_omics_data_cleaned.csv
• 04_drugcomb_gene_prot_clean → data_preprocessing_gene.pkl, data_drugcomb.pkl, data_preprocessing_protein.pkl
• 05_graph_generate.ipynb → nps_intersected folder
• 06_smiles_feat_generate.ipynb → smiles_graph_data.pkl
• 07_to_ecfp6_deepsyn.ipynb → deepsyn_drug_row.npy, deepsyn_drug_col.npy

If the console shows an error indicating that SMILES are not found, you MUST run the file 06_smiles_feat_generate.ipynb again to regenerate data.

Training and testing

To execute a training and testing task for our model, run the following script

python synprotx/<model>.py -d <database> -m <mode>

Possible options are listed below.

• model represents the name of the model to run. Must be one of gat, gcn, attentivefp and gatfp.
• --database/-d specifies data source to train the model on. Must be one of almanac-breast, almanac-lung, almanac-ovary, almanac-skin, friedman, oneil.
• --mode/-m input must be either clas, for classification task, or regr, for regression task. Default to clas
• Flags --no-feamol, --no-feagene, --no-feaprot disable the molecule branch, gene expression branch, and protein expression branch, respectively, when propagate through the model.

Note: There are more options to configure. Execute python synprotx/<model>.py -h for a more detailed description.

Results

The performance evaluation per repeated fold can be looked up in the folder "results". This folder includes a comprehensive list of all results files obtained from the training process. The models in comparison are XGBoost, DeepDDS, DeepSyn, SynProtX variations, and AttenSyn. The type of split includes random, cold-start for (leave-one-out) drugs, drug combinations, and cell lines, and ablation (gene and protein) on both classification and regression tasks.

Disclaimer: The CSV files in the "results" folder are not covered by the same MIT license as the source code. These data files are dedicated to the public domain under CC0.

Citations

Research Article

@article{boonyarit2025synprotx_gigascience,
  author       = {Boonyarit, Bundit and
                  Kositchutima, Matin and
                  Phattalung, Tisorn Na and
                  Yamprasert, Nattawin and
                  Thuwajit, Chanitra and
                  Rungrotmongkol, Thanyada and
                  Nutanong, Sarana},
  title        = {SynProtX: a large-scale proteomics-based deep learning model for predicting synergistic anticancer drug combinations},
  journal      = {GigaScience},
  volume       = {14},
  pages        = {giaf080},
  year         = {2025},
  month        = {08},
  issn         = {2047-217X},
  doi          = {10.1093/gigascience/giaf080},
  url          = {https://doi.org/10.1093/gigascience/giaf080},
  eprint       = {https://academic.oup.com/gigascience/article-pdf/doi/10.1093/gigascience/giaf080/64028448/giaf080.pdf},
}

Zenodo

@online{boonyarit2025synprotx_zenodo,
  author       = {Boonyarit, Bundit and
                  Kositchutima, Matin and
                  Phattalung, Tisorn Na and
                  Yamprasert, Nattawin and
                  Thuwajit, Chanitra and
                  Rungrotmongkol, Thanyada and
                  Nutanong, Sarana},
  title        = {SynProtX: A Large-Scale Proteomics-Based Deep Learning Model for Predicting Synergistic Anticancer Drug Combinations},
  year         = {2025},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.15603481},
  url          = {https://doi.org/10.5281/zenodo.15603481},
  note         = {[Dataset]}
}

WorkflowHub

@online{boonyarit2025synprotx_workflowhub,
  author       =  {Boonyarit, Bundit and
                  Kositchutima, Matin and
                  Phattalung, Tisorn Na and
                  Yamprasert, Nattawin and
                  Thuwajit, Chanitra and
                  Rungrotmongkol, Thanyada and
                  Nutanong, Sarana},
  title        =  {SynProtX},
  year         =  {2025}
  url          =  {https://workflowhub.eu/workflows/1726?version=3},
  DOI          =  {10.48546/WORKFLOWHUB.WORKFLOW.1726.3},
  publisher    =  {WorkflowHub}
}

Software Heritage

@online{boonyarit2025synprotx_software,
  author       = {Boonyarit, Bundit and
                  Kositchutima, Matin and
                  Phattalung, Tisorn Na and
                  Yamprasert, Nattawin and
                  Thuwajit, Chanitra and
                  Rungrotmongkol, Thanyada and
                  Nutanong, Sarana},
  title        = {SynProtX: A Large-Scale Proteomics-Based Deep Learning Model for Predicting Synergistic Anticancer Drug Combinations (Version 1)},
  year         = {2025},
  note         = {[Computer software]},
  url          = {https://archive.softwareheritage.org/swh:1:snp:750d09d4ed20b1628cef1f20cf0d2b2e518c4a3b;origin=https://github.com/manbaritone/SynProtX}
}