From Vaccines to Virtual Trials: AI's Biggest Health Week Yet

One week, six stories, one unmistakable signal: artificial intelligence is no longer a supporting player in medicine — it is designing the drugs, reading the scans, and simulating the trials.

The World's First AI-Designed Vaccine Enters Human Trials

Cambridge researchers have announced the first-in-human trial of a vaccine whose antigen was designed entirely by artificial intelligence. The shot targets a broad range of coronaviruses — not just the strains we already know, but family members that could seed the next pandemic. By training on the structural patterns of coronavirus spike proteins, the AI generated a synthetic antigen no human team had conceived. If it works, the implications go far beyond COVID: it would establish a repeatable playbook for rapidly generating pandemic-proof immunisations before an outbreak even begins.

This marks a philosophical shift as much as a technical one. Until now, AI has been used to accelerate vaccine development — helping researchers sift through candidates faster. Here, the AI is the inventor.

FDA Clears an AI That Reads Breast Ultrasounds and Writes the Report

RadNet subsidiary DeepHealth has earned FDA 510(k) clearance for an AI system that reads breast ultrasounds end-to-end and auto-generates the radiologist's report. In clinical validation, the tool boosted breast-cancer detection sensitivity by 8 percentage points and cut radiologist read-times by 37%. DeepHealth plans to deploy it across more than 400 imaging centres.

The detection gain matters most. Breast ultrasound is often ordered for women with dense tissue, where mammography is less reliable — exactly the population where misses carry the highest cost. An 8% sensitivity improvement in that group is not a marginal gain; it is additional cancers caught at an earlier, more treatable stage. The read-time reduction is a secondary benefit that makes the economics work for stretched radiology departments.

Routine Sleep Studies Are Hiding Heart and Dementia Risks — AI Can See Them

A study published in Nature Communications by Cleveland Clinic and UW Medicine found that an AI model trained on standard polysomnography data — the overnight sleep studies already routinely ordered for patients with suspected apnea — can identify patient subtypes whose five-year mortality risk is double that of their low-risk peers.

The critical finding: the risk signal is invisible to the apnea-hypopnea index, the conventional metric clinicians use to grade sleep-disordered breathing. The AI is picking up on patterns in heart rate variability, oxygen desaturation dynamics, and other signals embedded in data that today gets read for apnea severity and little else. Sleep labs are not ordering new tests; they are already generating this data. The open question is whether health systems will start running this model alongside their standard reads to flag patients for earlier cardiac and neurology follow-up.

Pathos AI Bets $125 Million on Two Cancer Drug Candidates

Pathos AI, an oncology startup built around an AI-agent drug development platform, has committed $125 million upfront across two licensing deals. The first acquires JSKN016 — a TROP2/HER3 bispecific antibody-drug conjugate — from Alphamab, targeting a receptor combination that has shown promise in hard-to-treat solid tumours. The second is a partnership with AstraZeneca on a PROTAC protein degrader aimed at oestrogen receptor-positive breast cancer that has grown resistant to standard hormone therapy.

Pathos's thesis is that its AI agents can compress the time from licensed candidate to clinical trial readiness — managing the preclinical data package, regulatory strategy, and trial design faster than a conventional biotech team. A $125 million upfront commitment signals that investors believe the thesis is credible enough to back at scale.

QuantHealth Raises $45 Million to Run Trials Before Patients Enrol

Israeli startup QuantHealth has closed a $45 million Series B, bringing its total raise to $70 million, to expand its virtual clinical trial platform. The technology builds AI-generated patient populations from real-world data — electronic health records, claims, and genomics — and simulates how those synthetic patients would respond to an experimental drug. The company reports accuracy rates of up to 90% against actual trial outcomes.

The target problem is the pharmaceutical industry's notorious 90% drug-failure rate, much of which is discovered expensively late in Phase II and Phase III. If a virtual trial can flag a probable failure — or identify the responding patient subgroup — before a sponsor spends hundreds of millions on a human study, the value is substantial. QuantHealth is in discussions with several large pharma companies to integrate its platform into early development decisions.

DiffuDose Personalises Prostate Cancer Radiation in 23 Seconds

Researchers at UMass Amherst have developed DiffuDose, a diffusion-model AI that generates personalised radiation dose maps for prostate cancer patients undergoing radiopharmaceutical therapy — in under 23 seconds. Current FDA-approved treatment uses standardised dosing, a practical compromise made because patient-specific dosimetry calculations take hours and require specialist medical physicists that most centres cannot staff at volume.

DiffuDose produces gold-standard dosimetry nearly instantly. The team validated it against manual calculations performed by expert physicists and found agreement tight enough for clinical use. If adopted broadly, it would let every patient receiving radiopharmaceutical therapy get a dose tailored to their anatomy and disease distribution rather than a population average — potentially improving tumour control while reducing radiation exposure to healthy tissue.

---

The pattern across all six stories is the same: AI finding signal in data that already exists, AI compressing timelines that were previously measured in years, and capital following both bets. The research-to-clinic pipeline is being rebuilt around machine speed. Whether it consistently improves patient outcomes — the only question that ultimately matters — will take years of follow-up data to answer. But the regulatory approvals and investment rounds are not waiting for that verdict.