AI in Medicine: $1.1B Drug Deal, Robot Prescriptions, and Cancer Breakthroughs

This week in health technology, artificial intelligence moved from research papers into operating rooms, pharmacies, and billion-dollar boardrooms. Five stories—each a signal of where medicine is heading—paint a picture of a field accelerating faster than its regulators and ethicists can keep pace.

METiS TechBio and Eli Lilly Sign a $1.1 Billion AI Drug-Delivery Deal

The largest headline of the week came from METiS TechBio, a Shanghai-based startup that uses AI to design lipid nanoparticles—the tiny fat shells that carry RNA medicines into specific tissues inside the body. Eli Lilly signed a deal worth up to $1.1 billion to access METiS's platform for developing next-generation RNA therapies.

The core problem METiS is solving is a stubborn one. RNA medicines—including mRNA vaccines and RNA interference drugs—are extraordinarily promising, but getting them to the right tissue without degrading or triggering an immune response is enormously difficult. Traditional nanoparticle design has relied on expensive, slow, trial-and-error laboratory work. METiS trains machine-learning models on molecular structure data to predict which lipid formulations will home in on a specific organ, slashing development time and cost.

For Eli Lilly, the deal is a bet that AI-designed delivery vehicles will unlock an entirely new class of RNA drugs for metabolic diseases, cancer, and beyond. For the broader industry, it signals that pharma giants are now willing to pay ten-figure sums for AI infrastructure—not just AI-discovered molecules, but the tools to get those molecules into patients safely.

Utah Becomes the First US State to Let AI Prescribe Without a Doctor

In a move that drew immediate national attention—and sharp criticism from the American Medical Association—Utah quietly amended its regulatory sandbox rules this month to allow Nolla Health's AI system to issue prescriptions for certain acne medications without any physician involvement.

Under the arrangement, a patient answers a digital questionnaire, the AI reviews the responses, and a prescription is generated and sent to a pharmacy automatically. No doctor reviews the case. No clinician signs off.

Proponents argue that for straightforward dermatology cases—a teenager with mild acne seeking a topical retinoid—the friction of mandatory physician review adds cost and delay without meaningfully improving safety. Utah's regulatory sandbox was designed precisely to test such innovations in a controlled way, with the state health department able to monitor outcomes and revoke permissions if problems emerge.

Critics, led by the AMA, counter that removing physician oversight from the prescribing loop sets a dangerous precedent. Acne drugs can interact with other medications; patients may underreport symptoms; AI systems can miss context a trained clinician would catch. The AMA has called on other states not to follow Utah's lead until long-term outcome data are available.

The debate reflects a genuine tension in AI-assisted healthcare: the efficiency gains are real, but so are the risks of removing humans from high-stakes decisions. Utah has effectively volunteered to run the experiment the rest of the country will watch closely.

Evaxion's AI-Designed Cancer Vaccine Holds Steady at Three Years

Denmark-based Evaxion Biotech reported three-year follow-up data from its Phase 2 trial of EVX-01, a personalized melanoma vaccine designed entirely by its PIONEER AI platform. The results were striking: all patients who had responded to the vaccine maintained those responses at the three-year mark, with no tumor progression detected—including in a subset who received the vaccine as a standalone therapy without checkpoint inhibitors.

Personalized cancer vaccines work by identifying mutations unique to a patient's tumor and training the immune system to attack those specific markers. The challenge is that each patient's tumor has a different mutational fingerprint, making design both scientifically complex and time-consuming. Evaxion's AI platform processes genomic sequencing data from a patient's biopsy and predicts which neoantigens—mutated proteins on the tumor's surface—are most likely to trigger a robust immune response. The vaccine is then synthesized and administered in weeks rather than months.

Three-year durability data in oncology is significant. Many immunotherapies show early promise but fade as tumors evolve and evade the immune response. Sustained remission across the entire Phase 2 cohort at this time point strengthens the case for moving EVX-01 into larger trials. Evaxion has indicated it plans to advance toward Phase 3 and explore applications in additional tumor types.

A Photo of Your Face Can Now Estimate Your Cancer Survival Risk

Researchers at Mass General Brigham published a study this week in The Lancet Digital Health introducing FAHR-Face, a foundation model trained on approximately 40 million photographs that estimates a person's biological age and survival risk from a single facial image.

The idea that the face encodes health information is not new—physicians have long noted that patients who look older than their chronological age tend to have worse outcomes. What FAHR-Face does is quantify that intuition at scale. The model was pre-trained on a massive dataset of photographs paired with medical records, learning to extract features—skin texture, facial symmetry, subtle asymmetries—that correlate with biological aging and disease burden. In cancer populations, the model's biological age estimates were significantly more predictive of five-year survival than chronological age alone.

The clinical application is to give oncologists an additional, non-invasive data point when counseling patients on treatment intensity. A patient who is 65 chronologically but whose face scores as biologically 75 might be a candidate for a less aggressive regimen; a 70-year-old who scores 60 might tolerate and benefit from a more intensive approach.

The researchers are careful to frame FAHR-Face as a decision-support tool, not a diagnostic. They also acknowledge that the model was trained predominantly on data from certain demographic groups and will require validation across more diverse populations before broad clinical deployment.

Penn Medicine's Percival Finds Disease Signals Hiding in Ordinary CT Scans

The final story of the week comes from Penn Medicine, whose AI model—named Percival—was trained on hundreds of thousands of CT scans to detect early disease signals that radiologists routinely miss because they fall outside the scan's original clinical purpose.

When a patient gets a chest CT to evaluate a lung nodule, the radiologist focuses on the lungs. Percival, trained to look at the entire image, noticed patterns in bone density, vascular calcification, and body composition that correlate with future cardiovascular events, osteoporosis, and diabetes—conditions the patient may not know they have and that the ordering physician was not looking for.

The model was validated in a study published this week, showing that Percival's incidental findings were predictive of clinical diagnoses made months to years after the original scan. In several cases, early warnings from Percival could have triggered preventive interventions that might have avoided hospitalizations.

Penn Medicine frames Percival as a step toward truly proactive healthcare: a system that transforms every routine imaging study into a broad screening opportunity. The challenge, as the researchers acknowledge, is managing the clinical and psychological consequences of incidental findings—not every patient wants to know about every potential future risk, and not every signal warrants intervention.

---

Taken together, this week's stories describe an industry at an inflection point. AI is no longer a research curiosity in medicine; it is prescribing drugs, designing vaccines, reading faces, and scanning bodies for hidden signals. The question now is not whether AI will reshape healthcare, but how quickly institutions, regulators, and patients will adapt to a world where the machine is often the first—and sometimes the only—clinician in the loop.