A Virtual Cell That Predicts Your Cancer Treatment

Posted on 13 September, 2026 by Mehrdad Fathi

One of the persistent frustrations in oncology is this: two patients with the same cancer diagnosis, the same staging, even the same tumor type, can respond completely differently to the same drug. For triple-negative breast cancer, that frustration is acute. TNBC lacks the hormone and HER2 receptors that most targeted therapies exploit, leaving clinicians with a narrower toolkit and fewer molecular handles to guide treatment decisions.

A study published in Nature on September 9 introduces an approach that may help change that — a proteomics-trained AI model capable of predicting how individual TNBC tumors respond to drugs, before those drugs are ever given.

What Was Built

The model was developed by Tiannan Guo’s group at Westlake University and trained on more than 38 million protein measurements. The dataset was generated by exposing 18 breast cancer cell lines — 16 of which were TNBC — to 63 FDA-approved antitumor drugs and 59 drug combinations, then measuring the levels of 5,585 protein groups at four time points: before treatment, and at 6, 24, and 48 hours after drug exposure.

That time-course design matters. Most existing virtual cell models are trained on single-cell sequencing data and capture static cell states. Cellular response to a drug is not a static event — it is a cascade. Proteins rise and fall, signaling networks shift, and resistance begins to emerge along a timeline. The multiple measurement windows allow the model to learn those dynamics rather than a single snapshot.

Hani Goodarzi, a systems biologist at the Arc Institute who was not involved in the study, described the scale and modality of the proteomics data as unique: “Modalities that we haven’t had before” for this type of model, he said. The temporal resolution was equally important in his assessment — without a time course, he noted, “you only have a lot of static images.”

What the Model Can Do

Trained on that data, the model learned to predict drug effects on TNBC cells and to identify proteins associated with resistance. Tested on 81 drugs not included it achieved 88% accuracy in predicting cellular response — a strong generalization result for a model of this kind.

The researchers then took the model out of the controlled cell-line setting entirely. Using proteomic profiles of tumor biopsies from 501 people with TNBC collected before chemotherapy, the model accurately predicted the clinical outcomes of the treatments those patients actually received. That is the transition that matters: from in vitro accuracy to real-world clinical concordance.

To stress-test clinical utility further, the team used the model to screen 3,000 compounds — either approved or in clinical trials — fo either approved or in clinical trials — for three individual TNBC patients, working from pre-treatment cell samples maintained in the laboratory. In effective when actually administered. It also suggested three additional compounds predicted to outperform the standard therapy those patients received, with drug-efficacy experiments on the cell samples supporting those recommendations.

What the Model Cannot Yet Do

Guo is explicit about the limits. This is a prototype. Prospective clinical trials — where model recommendations are tested prospectively in real patients before treatment decisions are made — are the necessary next step before this can influence clinical practice.

Goodarzi adds a structural limitation: the current model appears not to fully capture protein-protein interaction dynamics, and the effect of varying drug doses has not been systematically modeled. Both are meaningful gaps. Protein interactions are central to how signaling networks adapt to drug pressure, and dose-response relationships often determine the difference between efficacy and toxicity at the clinical level.

The model also does not currently include immune therapies, which is a significant constraint given the role of immunotherapy in TNBC treatment, particularly for PD-L1-positive tumors. Guo’s team has indicated this is a planned area for development.

The Broader Framing

The proteomics-first approach sets this work apart from the wave of AI models built on genomic or transcriptomic data. Proteins are the functional layer — they are what cells actually use to respond to drugs. Training a predictive model at that layer, with temporal resolution and FDA-approved drug coverage, is a different kind of infrastructure than what most computational oncology tools have been built on.

Goodarzi anticipates that the findings will catalyze joint modeling approaches integrating both protein and gene expression. If that synthesis emerges, the predictive framework could become substantially richer.

For TNBC specifically, which accounts for 15–20% of all breast cancer diagnoses and carries disproportionate mortality burden in younger and Black women, a validated tool for prospective treatment selection would be a meaningful advance. The current paper establishes that the biological logic is sound and the signal is real. What comes next is the harder work of clinical validation.

Source: Cyranoski, D. “AI model predicts which breast-cancer drugs work best.” Nature (2026). DOI: https://doi.org/10.1038/d41586-026-02845-2

References:

1. Sun, R. et al. A virtual cell model for triple-negative breast cancer drug prediction. Nature (2026). DOI: https://doi.org/10.1038/s41586-026-11001-9