Treating cancer is, in many ways, still an educated guess. The same tumour type behaves differently from patient to patient, and a drug that works brilliantly for one person may fail entirely for another. When multiple drugs enter the mix, the uncertainty compounds. Meanwhile, the cancer keeps growing. Now, a team of Chinese researchers has built something that could begin to change that: a virtual cell powered by AI that predicts, with remarkable accuracy, which drugs will work best for an individual patient’s cancer.
The system, called ProteinTalks, was developed at Westlake University in Hangzhou and published this month in Nature. It focuses on one of the most difficult cancers to treat — triple-negative breast cancer — a form that evades many standard therapies and leaves patients with far fewer options than other breast cancer types.
Why Proteins, Not Genes
Most attempts to build AI models of cell behaviour have used transcriptomics — measurements of which genes are active at any given moment. It’s a reasonable proxy, but an imperfect one. Gene activity doesn’t always reflect what proteins are actually doing inside the cell, and proteins are where the real biochemical action happens.
ProteinTalks cuts out the middleman entirely, training directly on proteins. The dataset behind it is formidable: the team treated 18 breast cancer cell types with 63 FDA-approved drugs and 59 common drug combinations, then measured thousands of proteins at four time points — before treatment and at 6, 24, and 48 hours afterwards. The result was more than 38 million protein measurements, now published as an open-source database. It is one of the largest resources of its kind ever assembled.
What the Virtual Cell Can Do
The model’s capabilities go well beyond simple drug response prediction.
ProteinTalks identified over 800 proteins whose levels shifted in response to drug treatment, flagging a rapidly changing subset that could serve as early warning signals of whether a drug is working. It also identified proteins likely involved in tumour resistance — the process by which cancers evolve to evade treatments over time — pointing toward potential new targets.
When challenged with 81 drugs it had never seen during training, the model predicted protein changes with 88% accuracy, outperforming several existing models. When trained on over 900 drug combinations, it gave higher scores to pairs that had already been validated in clinical practice — a reassuring sanity check suggesting the model is tracking genuine biology rather than generating plausible-looking noise.
The most striking demonstration came when the team applied ProteinTalks to real patients. Using proteomics data from three people with triple-negative breast cancer, the model screened 3,000 molecules and identified treatment regimes matching what had previously kept each patient’s disease under control — and suggested three additional molecules predicted to be even more effective. When those predictions were tested in actual cancer cells from the patients, the drugs inhibited tumour growth at lower doses than standard therapies.
“This is the first time that a virtual cell model goes out of the laboratory and is tested in a clinical scenario,” said study author Tiannan Guo.
Beyond Breast Cancer
Though trained on breast cancer, ProteinTalks showed it could adapt. When fed cancer-specific protein data from melanoma, colorectal, lung, and pancreatic cancer cells, it identified thousands of relevant protein changes — including patterns unique to each cancer type. The underlying architecture appears broadly applicable, not locked to a single disease.
This matters because the field of virtual cell modelling is rapidly expanding. Google DeepMind is developing an AI-powered virtual nucleus. The Chan Zuckerberg Initiative is partnering with Nvidia on virtual cell tools. Sweden’s Science for Life Laboratory has secured funding through 2033 for its AlphaCell programme. ProteinTalks joins a growing ecosystem of efforts to simulate cellular biology in digital form — each tackling the problem from a different angle.
The Honest Limitations
The researchers are careful not to overstate what they’ve built. ProteinTalks doesn’t account for how proteins interact with each other, with DNA, or with other molecules — and those interactions matter enormously. Drugs often work by disrupting the brief molecular handshakes between proteins, triggering ripple effects the model can’t yet see. The current version can only evaluate two-drug combinations, not more complex regimes.
Most importantly, lab dish results are not clinical results. The predictions have been validated in cancer cells grown in culture, not in patients. Animal models and then clinical trials are the necessary next steps before ProteinTalks can influence real treatment decisions.
Why It Still Matters
Despite those caveats, the proof of concept is significant. A virtual cell that can scan thousands of molecules, match them to a patient’s specific tumour profile, and identify treatments that work better than the standard of care — even imperfectly, even in a dish — represents a genuinely new way of approaching one of medicine’s most persistent problems.
Trial and error in cancer treatment has costs measured in time, side effects, and lives. A tool that narrows the guesswork, even partially, is worth building on. ProteinTalks is not yet that tool. But it’s a credible step toward it.