XPert: Modelling drug-induced cellular perturbation responses with a biologically informed dual-branch Transformer

Systematic mapping of chemical perturbation responses is revolutionizing polypharmacological drug discovery, yet remains constrained by experimental scalability. Here, we introduce XPert, a biologically informed Dual-Branch Transformer model that predicts gene-specific drug responses across dose-time conditions. It generalizes to unseen cells, transfers knowledge to clinical settings, and reveals mechanistic insights, offering a scalable solution for precision medicine and perturbation-based drug discovery.

To facilitate transparency and reproducibility, we provide all relevant resources used in this study. These data are provided as supplements to our code repositories on Zenodo (https://doi.org/10.5281/zenodo.17182939) and GitHub (https://github.com/GSanShui/XPert). A summary of the available datasets is as follows:

XPert.zip: The complete XPert code repository, containing all model architectures and training scripts. This package is identical to the versions hosted on both GitHub and Zenodo.

HG_data.zip: Located under the 'XPert/' directory, this package includes the datasets used for constructing the knowledge-informed heterogeneous graph, together with the corresponding pre-trained node embeddings.

saved_model.zip: Located under the 'XPert/' directory, this package contains the trained model weights and associated pre-trained parameters.

reproducing.zip: Located under the 'XPert/' directory, this package provides the code for reproducing the main analysis and figures reported in the paper.

processed_data.zip: Located under the XPert/ directory, this package contains all processed datasets and meta used in this study.

All other files not included in the compressed archives above are provided under the 'XPert/processed_data/' directory. Users may retrieve these resources as needed and should place them in the designated file paths to ensure full reproducibility of the analyses.