5 Science News Stories — October 24, 2026

 5 NEWS OF THE DAY  |  24 OCTOBER 2026

AI Takes on Diagnosis, a Cancer Vaccine Faces Its Test, and Black Holes Challenge Physics

Five research stories shaping the future of medicine, space technology, fundamental physics and climate forecasting. AI-generated editorial concept art; not photographs of the reported experiments.

01  /  AI / MEDICINE  /  OCTOBER 23, 2026

Microsoft’s AI Works Through Diagnoses Step by Step — But Is Not a Doctor Yet

A patient rarely arrives with a tidy list of test results and one multiple-choice question. A clinician has to ask what hurts, consider possible causes, order useful tests and revise the diagnosis as new evidence appears. Microsoft’s AI Diagnostic Orchestrator, or MAI-DxO, was designed to simulate that sequence rather than guess from a fully completed case file.

The system coordinates several AI-generated clinical viewpoints that propose explanations, challenge one another and decide what information to request next. Microsoft evaluated it using SDBench, a benchmark built from 304 unusually difficult New England Journal of Medicine cases and a separate set of emergency-department presentations. On the difficult cases, the company reports 83% diagnostic accuracy. In the comparison, physicians reached 36% with reference resources and 22% without them. On the emergency-department cases, the reported figures were 80% for MAI-DxO and 47% for physicians.

Those numbers are striking, but they are not evidence that the system is a safer physician in a real hospital. These were structured simulated work-ups with a particular selection of cases, physician participants and estimated costs. Real care also requires physical examination, patient communication, incomplete or misleading histories, monitoring over time and accountability for mistakes. The system’s reported lower investigation costs are benchmark estimates, not proof of lower hospital bills.

The research points toward a more useful role for clinical AI: helping doctors compare hypotheses and choose tests without automatically ordering everything. That connects to our broader look at AI assistants for doctors, where human review remains central. Microsoft says prospective evaluation with hospital partners is next. The company published its report on October 22 in the United States; the peer-reviewed study drew fresh coverage on October 23. Clinical validation, safety monitoring and regulatory review are still ahead.

PRIMARY SOURCE

Microsoft AI — Building Medical Superintelligence (Oct. 22, 2026)

02  /  CANCER / CLINICAL TRIAL  /  OCTOBER 23, 2026

A KRAS Cancer Vaccine Triggers Immune Responses — Without Yet Preventing Recurrence

Cancer vaccines aim to train immune cells to recognize something distinctive about a tumor. On October 23, Mayo Clinic reported results from a randomized phase 2 trial of a vaccine targeting seven common mutations in KRAS, a gene altered in most pancreatic ductal adenocarcinomas. Unlike a fully personalized vaccine, this candidate is designed for multiple patients whose tumors carry the targeted mutations.

The trial enrolled 144 people with stage 1–3 KRAS-mutated pancreatic cancer who had already undergone surgery and chemotherapy. Imaging showed no visible cancer when they joined. Researchers assigned 96 participants to vaccination and 48 to observation. Among 90 vaccinated participants who could be evaluated for immune response, 89 — 98.9% — developed increased KRAS-specific T-cell responses. That is an impressive biological signal: the immune system appears to have learned to recognize the intended targets.

Yet the study’s main clinical outcome was disappointing. Median disease-free survival was 23.8 months in the vaccine group and 21.2 months in the observation group; the difference was not statistically significant. The trial therefore did not meet its primary endpoint. Stronger immune responses were associated with longer periods without recurrence in exploratory analyses, but those findings cannot establish that the vaccine caused the difference or identify who will benefit.

The investigators also examined circulating tumor DNA, or ctDNA, which can signal residual disease after treatment. Patterns involving baseline ctDNA and previous radiation therapy may help researchers design the next trial. A phase 3 study is planned, and overall-survival results are not yet available. The careful conclusion is neither 'cancer vaccine works' nor 'cancer vaccine fails': it generated strong immune recognition, but has not yet demonstrated a clear recurrence-prevention benefit.

PRIMARY SOURCE

Mayo Clinic — Investigational KRAS Vaccine Phase 2 Results (Oct. 23, 2026)

03  /  SPACE / MATERIALS  /  OCTOBER 23, 2026

Could Layers of Everyday Mylar Help Keep Future Moon Bases Warm?

A Moon habitat must retain heat through brutal temperature swings, survive vacuum conditions and tolerate years of handling. Insulation that works beautifully in a delicate laboratory sample may fail when engineers bend, compress or transport it. Researchers at the University of Pennsylvania have tested a deceptively simple alternative: carefully stacked sheets of Mylar, a thin polyester film familiar from reflective emergency blankets.

Their hybrid layered insulating film stacks, or HyLIFS, combine many thin films, some treated with an extremely thin alumina coating. In vacuum tests, the best samples showed strong resistance to heat transfer while remaining flexible. Coated stacks also retained their insulating performance after ten repeated compression cycles at stresses of roughly 20 megapascals. This addresses a central engineering trade-off: the material needs to be a good insulator without being too fragile to use.

The surprising detail was microscopic. Commercial Mylar contains tiny 'antiblocking' particles that keep adjacent sheets from sticking together. The researchers found that those same particles can provide small contact points through which heat moves between layers. That suggests manufacturers could eventually tune particle size and distribution to improve insulation, rather than treating them as an incidental manufacturing ingredient.

But the samples were only about one square centimeter and assembled painstakingly by hand. The work has not established how meter-scale panels would behave after lunar dust exposure, radiation, repeated bending or long-term installation. The underlying paper appeared in September; Penn Engineering’s new report was published October 23. Our guide to Moon bases explains why durable thermal control is only one part of the larger survival system. The next step is scalable manufacturing and testing under realistic lunar conditions, not a claim that Moon habitats already use this material.

PRIMARY SOURCE

Penn Engineering — Mylar Insulation for the Moon (Oct. 23, 2026)

04  /  COSMOLOGY / THEORY  /  OCTOBER 23, 2026

Can One Black Hole Split Into Two? A New Calculation Finds a Narrow Loophole

Astronomers have observed black holes merging, but the reverse sounds almost impossible: one black hole somehow dividing into two. A theoretical paper discussed by the University of Southern Denmark on October 23 asks whether known physical constraints rule out that possibility in every case. The answer is nuanced: ordinary black holes should not split, yet extremely fast rotation may open a narrow mathematical possibility.

The first obstacle comes from entropy. In black-hole thermodynamics, the entropy of a nonrotating black hole is proportional to the area of its event horizon. Splitting one into two smaller holes would normally lower the combined entropy, conflicting with the expected direction of physical evolution. The researchers examined whether the energy and angular momentum of a spinning black hole can change this balance.

Their calculations indicate a possible window for a rapidly rotating object. But a split would only count as a genuine fragmentation if the two resulting black holes could move apart instead of promptly rejoining. Imposing that condition closes much of the window, leaving at most a possibility near the maximum rotation rate. The researchers also explored more speculative theories with extra spatial dimensions, where instability can behave differently.

This is a theoretical investigation, not an observation of a black hole splitting and not a prediction that such events are common. The authors do not calculate a reliable event probability, and the answer may depend on physics beyond Einstein’s general relativity. Their 'pair-of-pants' term describes the hypothetical shape of a horizon traced through spacetime. Future observations might constrain such ideas indirectly, but today the result is a provocative boundary test for gravitational theory — not a discovered cosmic event.

PRIMARY SOURCE

University of Southern Denmark — Can a Black Hole Split in Two? (Oct. 23, 2026)

Physical Review D — Original Theoretical Paper (Oct. 22, 2026)

05  /  CLIMATE / FORECASTING  /  OCTOBER 23, 2026

Scientists Find When El Niño Forecasts May Be Trustworthy a Year Ahead

El Niño and La Niña are large-scale changes in tropical Pacific Ocean temperatures and atmospheric circulation. They influence rainfall, drought and temperature patterns far beyond the Pacific. Forecasters can often identify an approaching event several months in advance, but predictions stretching to a full year are much harder. A study published October 23 in Science Advances asks a useful question: can we tell beforehand which long-range forecasts deserve more trust?

Researchers associated with Australia’s CSIRO and the University of Melbourne examined more than 7,500 historical forecasts produced by climate models from nine countries. They compared forecast skill with the state of the Pacific at the moment each prediction began. A clear pattern emerged: one-year forecasts initialized during La Niña conditions were consistently more reliable across the nine models, while forecasts starting from neutral conditions were substantially less dependable.

That matters because a forecast should not be presented as equally certain in every situation. Farmers, water managers and disaster planners may benefit more from knowing when a warning is unusually credible than from hearing one unqualified prediction. The study also suggests that statistical and machine-learning approaches may help under conditions where traditional physics-based models struggle, although this remains an avenue for further testing.

The researchers have not created a universal system that predicts every El Niño twelve months ahead. Their result concerns the reliability of existing forecasts under different starting conditions, not guaranteed event detection. They say operational forecasting agencies will need further validation before adopting the method. The longer-term opportunity is better calibrated early warnings: communicating both what models predict and how much confidence the public should place in those predictions.

PRIMARY SOURCE

Science Advances — Using Initial States to Identify Successful ENSO Forecasts (Oct. 23, 2026)