Dataset, Model Weights and Code for: PAIRWISE: Deep Learning-based Prediction of Effective Personalized Drug Combinations in Cancer Response

Abstract

PAIRWISE: Deep Learning-based Prediction of Effective Personalized Drug Combinations in CancerPAIRWISE predicts whether a drug pair acts synergistically in a specific tumour sample. It fuses three modalities — molecular graphs of the two compounds, theirdrug–target interaction profiles propagated over a protein–protein interactionnetwork, and the sample transcriptome — through an attention encoder into a singlesynergy probability. This repository contains the model, the seven benchmarked baselines, and the code that reproduces figure and table in the manuscript. Also you can refer to the code repo in the GitHub repository. Interactive predictions: synergy explorer and BTKi explorer Contents of this depositdata/ — the p13 benchmark corpus (~30K drug–drug–cell-line records, 1,275 drugs × 163 cell lines across 15 lineages, harmonised from 13 public screens with Loewe/Bliss/ZIP/HSA scores recomputed in SynergyFinder v3.0), plus chemical, drug–target and transcriptome feature banks, the STRING PPI network, and gene sets.weights/ — trained checkpoints for PAIRWISE and eleven baselines (seven published deep-learning methods and four classical ML models). best_model_pairwise.pth is the main model (held-out test AUROC 0.8444).results/ — out-of-fold and held-out predictions for every model.paper/ — per-stage inputs, outputs and scripts behind each figure and table, one directory per analysis (benchmark, wet-lab screen, network/pathway, patient stratification, external DLBCL validation, NCI-DREAM comparison, ablation, SHAP, supplementary). Each has its own README.example/ — a runnable notebook with sample data, sample weights and the original run logs; so you can freely make a prediction under the instruction.UsageClone the repository and install: git clone https://github.com/Mew233/pairwise.git && cd pairwise conda create -n pairwise python=3.10 && conda activate pairwise pip install torch==2.3.0 --index-url https://download.pytorch.org/whl/cu121 pip install dgl==2.4.0 -f https://data.dgl.ai/wheels/torch-2.3/cu121/repo.html pip install -e . Download and place the data. Extract thei zip file you will automatically dump data/, weights/ and results/ into the repository root, or leave them elsewhere and point the package at them: export PAIRWISE_DATA_ROOT=/path/to/deposit/data export PAIRWISE_WEIGHTS_DIR=/path/to/deposit/weights Run a quick test. Open example/example2run.ipynb, which walks through prediction, feature extraction, fine-tuning and training on the bundled sample data.To retrain from scratch (5-fold CV on p13): python -m pairwise.main --model pairwise --synergy_df p13 --train_test_mode train For environment setup, the full training pipeline and per-analysis instructions, see the README.md in the GitHub repository and the stage READMEs under paper/.NotesData files remain subject to the terms of their original sources (DrugComb and the 13 constituent screens, CCLE/DepMap, TCGA, STRING, DrugTargetCommons, DrugBank, the NCI-DREAM Challenge, and Griner et al.). Code is released under the MIT license.Associated Publication: Xu C., et al. "PAIRWISE: Deep Learning-based Prediction of Effective Personalized Drug Combinations in Cancer", under revision (2026).