As scientists argue over what a virtual cell is, what it can be, and whether it will eventually open the portal to Narnia, AI × bio just reached its biggest clinical milestone yet: an AI-originated drug is entering Phase III.

Anthropic may have just started talking about developing drugs of its own, but Insilico Medicine has a twelve-year head start. Rentosertib, its fibrosis drug, is now moving into the trial stage Alex Zhavoronkov calls “the ultimate test.”

In this issue we also have Biohub using AI after a large human-cell CRISPR screen to find unexpected psoriasis targets, ATHENA training medical agents to look up evidence before recommending treatment, and OpenDDE making a strong open-source move in biomolecular structure prediction.

Let’s dive in.

Insilico Medicine’s AI-designed drug rentosertib has entered Phase III for idiopathic pulmonary fibrosis, a progressive, age-related fibrotic lung disease. Insilico CEO Alex Zhavoronkov hailed it as a “major milestone” and “the ultimate test”.

“To my knowledge, in our AI drug discovery industry, no other company took the drug for a novel target with a novel molecule into Phase III before so it is a big milestone for the entire industry,” Zhavoronkov wrote in a post on X.

The Phase III trial will enroll 320 people at 47 centers in China. The main measure is forced vital capacity - how much air you can exhale after a full breath - tracked over 52 weeks.

Insilico published Phase IIa results in Nature Medicine last year. The 71-patient study found rentosertib was tolerable and showed a dose-dependent lung-function signal.

The news comes after a period of large Insilico deal announcements. In March, Lilly signed a deal worth up to $2.75 billion. In recent weeks, Insilico has also announced agreements with SK Biopharmaceuticals, worth up to $2.5 billion, and Takeda, worth up to $600 million.

Why it matters: AI drug discovery has plenty of deals, demos and platform claims. But Zhavoronkov is right: this is the ultimate test.

Did you know? Insilico is hiring.

The 2026 Virtual Cell Challenge kicks off August 20. But you can already bring out the popcorn: scientists are now feuding in public over whether “virtual cell” is a real destination or a hype label for narrower prediction tools.

The spark was a new paper from Shift Bioscience researchers, with Bo Wang - the University of Toronto researcher recently promoted to Chief AI Scientist at Xaira - as a co-author. Presented at the International Conference on Machine Learning, ICML 2026, it looks at perturbation prediction: asking what happens to a cell’s gene activity after a gene is switched off or otherwise disturbed.

It addresses the uncomfortable point BAIO has covered several times: on common benchmarks, simply predicting the average cell response can beat much fancier single-cell models. Not a good look for something supposedly taking us toward virtual cells.

So what’s going on? Wang summarized the paper’s finding on X like this: the models aren’t bad, it’s actually the benchmarks that are wrong.

The paper’s explanation is that most genes do not change after a perturbation. So broad scoring methods can reward models for getting the boring majority right while missing the few genes that carry the signal. The authors propose evaluation metrics that put more weight on those perturbation-specific genes.

Then Wang widened the argument: “Perturbation prediction is a means, not the end.” The larger goal, he argued, is a model that helps drug discovery: target finding, mechanism, toxicity, biomarkers and therapeutic design.

Anshul Kundaje, Stanford computational genomicist and frequent AI × bio hype critic, did not buy it. “None of these objectives will deliver a virtual cell,” he wrote, arguing that each task can be solved by shortcut models trained on the right data.

“Mark my words. These same folks were arguing that scFMs were virtual cells. They have been consistently wrong,” Kundaje added.

Wang replied that he has no problem with simpler models. “Science should reward what works, not what is most sophisticated,” he wrote. But he defended virtual cells, foundation models and world models as useful high-level concepts, adding: “Rigorous validation will decide […] I don't think we should dismiss a promising research direction simply because the terminology sounds aspirational.”

Why it matters: There is a lot of work and creativity going into virtual cells right now. Since BAIO started in February, we have covered Xaira’s X-Cell, PerturbGen, Tripso, Biohub’s $500 million virtual-biology push, VCHarness, and Tahoe’s Rhaister. But for all that energy, there is also a sense that virtual cells have not delivered yet, amid repeated warnings that models fail against simple baselines. We’ll keep you up to date as the story develops (which it does, as you can see in the screenshot below, taken just as I write this).

Did you know? Adding fuel to the flames, perhaps, and highlighted by Anshul Kundaje, is a new paper called “Tabular foundation models are competitive cellular perturbation predictors across biological scales.” Take a look.

Biohub used CRISPR and AI to find unexpected drug targets for psoriasis - including the oxytocin receptor, better known for its involvement in childbirth and social bonding than skin inflammation. The findings have been published in Nature Communications.

Psoriasis is driven in part by keratinocytes, the outer skin cells that become overactive and inflammatory. These cells have been hard to study with CRISPR screens because delivery chemicals are toxic to them.

The team got around that with a centrifuge-based delivery method, then knocked out roughly 19,000 genes in skin cells from two human donors. They measured which knockouts changed IL17RA, a receptor that helps skin cells respond to psoriasis inflammation.

That produced a long hit list. To find the overlooked-but-plausible ones, the team used VirtualCRISPR, a language model trained on functional-genomics literature, to ask which hits the field would already have expected.

Two stood out: ALOX5, already targeted by the asthma drug zileuton, and OXTR, the oxytocin receptor. The team validated both targets in primary human skin cells, showing that blocking or knocking out ALOX5 and OXTR reduced IL17RA and dampened inflammatory signals. Then they made topical gels with zileuton and cligosiban, an OXTR blocker, and tested them in a mouse psoriasis model, where both reduced inflammation about as well as an injected anti-IL17RA antibody.

Why it matters: This is a great example of AI working where biology actually needs help: after a large human-cell experiment has produced too many possible leads.

Did you know? VirtualCRISPR code is on GitHub.

AI models can be lazy. Ask about a paper and they may give you a confident fictional version instead of reading it. Ask for a current fact and they may pull something stale from memory instead of checking.

Often the remedy is to make the model look things up.

That is sort of the idea behind ATHENA-R1, a new open medical agent from a multi-institutional team led by Harvard’s Marinka Zitnik, described in a preprint. It is built to decide when it needs evidence before recommending treatment.

ATHENA can call a library of biomedical tools tied to FDA drug labels and other curated sources. It pulls evidence on indications, contraindications, drug interactions and patient-specific restrictions, then leaves a trace of what it checked and how it used that evidence.

In a comparison, OpenAI’s GPT-5 was also given optional access to the same tool library. In one test, it used tools in only about 1% of treatment cases. ATHENA used them throughout, because that behavior was trained into it through generated reasoning traces and reinforcement learning with scientific feedback.

In the preprint, ATHENA reached 94.7% accuracy on 3,168 drug-reasoning questions and 82.9% on 456 patient-specific treatment cases, beating GPT-5 in both. And rare-disease experts from 28 organizations preferred ATHENA over reference models across all eight evaluation criteria, with the biggest gains for traceability and the usefulness of its rationale. Practicing physicians also rated its recommendations favorably on complex hospitalized-patient cases.

Why it matters: Tools only help if the model actually uses them. ATHENA is an attempt to get medical agents into the habit of doing so.

Did you know? ATHENA-R1 has a handy project page with links to paper and code.

OpenDDE is Aureka’s open-source all-atom co-folding model for predicting biomolecular complexes, from proteins and antibodies to nucleic acids and small molecules. It reportedly beats AlphaFold3, Protenix-v1 and ESMFold2 on one antibody-antigen benchmark. It is still a preview, and the authors are clear that small-molecule docking and affinity prediction are not solved. Try it with Python 3.11 or Docker; weights are on Hugging Face, code on GitHub.

☑️ “Curing All Disease by next century is too conservative.”

☑️ ScrambleBench: a workflow for comparative assessment of structure-based de novo generative models.

☑️ “Well, today seemed like as good a moment as any to have Claude analyze my genome.”

☑️ “We built the lab that's able to go from AI-led drug design to data in 24h.”