A Computer Stand-In for Breast Cancer Cells, Built to Predict Drug Responses

Researchers at Westlake University and Peking University have built a computer model of breast cancer cells and trained it on more than 38 million measurements of protein levels recorded as the cells reacted to drugs. Their paper was published in Nature on Sept. 9, 2026.

Rui Sun and Tiannan Guo of Westlake University, Peijie Zhou of Peking University and colleagues call the model ProteinTalks and describe it in the paper as an operational tool for a set of drug-discovery jobs: predicting how well a drug or a drug pair will work, suggesting new combinations, flagging proteins linked to drug resistance, sorting patients by their likely response, and ranking candidate drugs for organoids, which are lab-grown clumps of a patient's own tumor tissue.

To produce that record, the group exposed breast cancer cell lines to drugs and tracked how much of each protein was present at a series of time points. Most existing virtual cell models, the authors write, lack data of this kind.

The model extends beyond cell lines to organoids and to biopsy samples taken from patients, "generally achieving higher performance than the selected benchmark implementations under the evaluated protocols."

It reports predictions and comparisons rather than patient outcomes, and the study remains at the level of laboratory cell lines, organoid cultures, and biopsy tissue.

The raw measurements have been deposited with the ProteomeXchange Consortium through iProX under accession IPX0007409000, the paper says, and the protein matrix behind the model is available for academic and non-commercial use at db.prottalks.com. The analysis code is on GitHub.