VivoSim and Chemveda put human tissue at the centre of AI drug discovery

Generative models now propose molecules far faster than anyone can test them. Safety is where the queue forms.

If human-tissue screens work as advertised, biotech clients could drop bad compounds earlier and cut animal testing. Nobody has shown yet that better lab predictions lift clinical success rates.

The deal: human tissue as the test bench

Under the collaboration, Chemveda will design, synthesize and optimize AI-derived compounds for pharmaceutical and biotechnology clients. VivoSim will run selected molecules through its NAMkind Liver and NAMkind Intestine models, which are built from primary human cells. Why put human tissue in the loop at all? Because the chemist needs the toxicity signal while the molecule can still be changed. VivoSim calls this a rapid design-make-test cycle, with tissue data feeding back into synthesis rather than arriving after a candidate is locked. Keith Murphy, VivoSim's executive chairman, put the problem in blunt terms in the announcement. AI has handed the industry "a firehose producing new molecules", he said, but not a way to know which will work and which may harm patients.

Why safety testing is the bottleneck

The pitch rests on a familiar failure mode. Liver and gastrointestinal toxicity remain among the most common reasons promising drugs fail in the clinic, often after years of work and hundreds of millions of dollars. Generative models can now propose thousands of novel molecules in the time it once took to design a handful. VivoSim's evidence is a clinical dataset of 92 small-molecule compounds. On that set its NAMkind Liver model predicted drug-induced liver injury with 91% accuracy, 90% sensitivity and 95% specificity, the company says. It also claims more than 90% sensitivity where animal models and traditional methods detect only 50% to 65% of true liver toxicants.

These are company figures. They were published in a company announcement and have not been independently replicated.

The regulatory tailwind is real, though. The US Food and Drug Administration actively encourages human-relevant new approach methodologies in place of animal testing, according to the same announcement. That matters commercially: a screen regulators accept is worth more than one that only convinces a chemist.

Other groups are building the same validation layer

VivoSim is not alone in selling validation rather than generation. QurieGen and Rayca Precision said on 3 November that they will link target discovery, AI-powered molecular design and cellular validation in one iterative workflow. QurieGen measures gene expression, surface and intracellular proteins and phosphoproteins in individual cells over time. Rayca combines AI with physics-based simulations to model how proteins move and to find binding sites, including temporarily accessible ones. A pilot between them starts with one biological question, a disease area and agreed constraints. When results differ from expectations, the data is meant to separate a bad target from a bad molecule. The partners say the loop is designed to cut uncertainty before critical R&D decisions.

Money is moving into the same layer. Vitroscope raised $7 million, a $5 million seed round plus $2 million in non-dilutive funding from the Research Council of Norway, and has raised $8.8 million to date.

Its vitro.alive benchtop platform runs six independent, parallel experiments, and Vitroscope and Instant Bio kept vascularised human brain organoids under perfusion for seven days. A 96-experiment version, vitro.scale, is in development. Data is the other route. Parse Biosciences and Nomic Bio are building Cascade, a baseline dataset of 100 million cells that pairs single-cell RNA sequencing with more than 11,000 secretory protein profiles across more than 150,000 cytokine and drug conditions. GSK, Pfizer, Sanofi, Graph Therapeutics and Turbine have signed on as partners, according to the companies.

What would count as proof

What would proof look like? VivoSim says its partners will know in weeks, not years, which molecules deserve to move forward.

That is a claim about speed. The harder claim, that human-relevant screens raise the share of candidates surviving trials, will take years to test. Big pharma's own deals show the timeline. AstraZeneca signed a strategic drug discovery partnership with Insitro covering metabolic liver disease and heart muscle conditions, structured as milestone-based payments. Kalkine Media noted that any commercial benefit from that alliance would lie many years ahead. Vitroscope's new money is earmarked for commercial rollout of vitro.alive and for higher-throughput work with vitro.scale. Parse and Nomic call Cascade the largest resource of its kind ready for AI applications to model the immune system. Murphy said the aim is to replace guesswork with human-relevant results. None of these groups has published a clinical success rate. That number, when it arrives, will settle the argument.