A 'Virtual Cell' AI Says It Predicted Drug Effects in Cancers It Never Saw Trained On

A Beijing AI-bio startup says its 'virtual cell' model predicted how an ovarian cancer drug would behave in tumors it had never been trained on. The only public validation anyone has seen is the company's own.

The startup, 华源智因 (HuaYuan ZhiYin), has closed a seed round in the 10 million RMB range led by 水木创投 (Shuimu Venture), according to a 36Kr exclusive. Its product, Wise-Perturb, is a virtual cell model: software meant to forecast how a drug will act on a human cell before any patient takes it. The pitch is to replace animal screening and early patient cohorts with cheap computational triage for the pharma industry's standard '10 years, $1B, roughly 10 percent approval' problem.

The founding team is Du Runshi, a UCLA-trained serial AI entrepreneur; Li Yu, a CUHK computer science assistant professor serving as chief AI scientist; and Wang Yixuan, previously lead on BioMap's gene perturbation prediction model xTrimoSC Perturb, as CTO. An advisory board includes researchers from Shenzhen's National GeneBank. The team's training pedigree fits the field. The open question is whether the validation claim survives outside scrutiny.

Two cases sit at the center of the company's pitch, both reported in the 36Kr interview. In the first, the team trained the model on breast cancer patient data without any ovarian cancer exposure, then asked it to predict the behavior of DS-8201, the antibody-drug conjugate behind Enhertu that is approved in breast cancer and has documented ovarian cancer activity, in ovarian tumors grown from patient tissue and implanted in mice (the standard preclinical 'PDX' model). The company says the model's response predictions matched the mouse model readout more closely than a generic virtual cell baseline. In the second case, working with the Cancer Hospital of the Chinese Academy of Medical Sciences on osimertinib, a targeted lung cancer drug, the hospital provided only pre-treatment baseline sequencing and withheld follow-up outcomes. The model independently stratified likely responders, and the company reports the prediction matched the long-term clinical efficacy observed later.

Both cases share a shape: a single zero-shot transfer demonstration, no co-investigator paper, no blinded pre-specified endpoint, no outside replication. The DS-8201 result is the more provocative because the prediction was generated without any ovarian cancer in training, the specific capability readers cannot infer from a breast-cancer-only benchmark.

Two mechanisms would, in principle, make the claim work, according to the company's technical description. The model is cell-type aware: a separate architecture branch identifies which type of cell it is looking at before predicting response. That is what would let the same model handle tissues it was never trained on. The training data stack is layered: large public single-cell sequencing at the base, hundreds of millions of paired perturbation experiments from cell lines and mouse models in the middle, and a top tier of paired pre- and post-treatment clinical sequencing from rare tumor cohorts the company controls. The cell-type branch permits cross-cancer transfer; the layered data gives the prediction any signal.

What would falsify the claim? A blinded comparison where an outside lab runs the same DS-8201 experiment, trains the model on breast cancer data only, and scores the prediction against a fresh set of ovarian patient-derived mouse tumors under pre-registered endpoints. That test has not been reported.

The commercial logic is pipeline triage. HuaYuan ZhiYin says it has signed service contracts with top-tier Chinese three-A hospitals, multinational pharma, and other AI-bio companies, with some fees already collected. The company is also building joint wet-and-dry labs with leading hospitals and plans to deepen ties with at least 30 of them over the next one to three years. The partnership list, deal values, and revenue figures are not in the public source. The current lead investor, Shuimu Venture, is a Chinese early-stage fund whose independent confirmation of the round terms is not in the public source.

Two things to watch. First, whether any of the validation cases get published with outside labs and co-investigators; an independently scored mouse-model replication would shift the read from showcase to capability. Second, whether the next funding round, which the company says it is preparing, brings in a biotech-focused lead willing to diligence the model's claims. The DS-8201 case is what investors and partners are paying for, and right now they are buying the company's word for it.