The Parts That Do More Than We Thought

"Organoid intelligence" is an emerging way to build computers out of lab-grown clusters of human brain cells instead of silicon. The neurons come from one of two starting points, which raise different questions about where they came from even though they end up capable of similar things: leftover embryos from IVF (embryonic stem cells), or a person's own adult cells, chemically reprogrammed backward into a stem-cell-like state (induced pluripotent stem cells, or "iPS cells") and then coaxed into becoming neurons. Once a stable organoid line exists from either source, it can in principle be kept alive and growing indefinitely.

The appeal for computing specifically: today's AI chips keep memory and computation in physically separate places, and shuttling data between them is a large part of what makes AI hardware so power-hungry. A biological neuron does both jobs in the same piece of tissue, and runs on far less energy — the core technical case for growing a computer instead of manufacturing one.

The piece isn't reporting a new technical result — the novelty is the ethical gap it surfaces. Donating cells for research today means giving broad, open-ended consent: whoever collects the tissue can use it for whatever they judge "reasonable," an arrangement that trusts researchers' good judgment rather than spelling out what's actually allowed. Nobody donating tissue for general biomedical research was specifically asked whether they're fine with their neurons being used to build something that computes the way a brain does — arguably edging toward a system capable of some rudimentary form of awareness. The sharper version of the question this raises: once that's a real possibility, is this still just "research," in the sense the original consent covered?

This is a consent framework that hasn't caught up with what's actually being done with donated tissue. Open-ended consent assumed a world where "biomedical research" meant reasonably foreseeable things; organoid computing is a use case nobody was picturing when that consent model became standard, and the field is moving faster than the ethics conversation around it.

The heart has its own small, semi-independent nervous system — an "intrinsic cardiac nervous system" living not in the heart muscle itself but in the fat pad surrounding it. It's a tiny fraction of the heart's cells, roughly 0.01%, and until this study, nobody knew what these neurons actually did day to day.

Researchers identified two functionally distinct populations within this network: neurons marked by the gene Npy, which lower heart rate when activated, and a separate population marked by Ddah1, which provide resistance to physiological stress. Both sets of marker genes also show up in human cardiac tissue, not just the mice studied — a meaningful detail, since it means the finding isn't automatically limited to mice.

This mirrors a pattern already established elsewhere in the body: the gut has its own well-known "second brain" (the enteric nervous system) that does more than just relay signals from the brain. This is the same shape of finding, just in the heart — these peripheral neural clusters run their own local, specialized jobs, with different neuron types doing distinctly different things, rather than being simple message-passing relays.

Because the marker genes are conserved in humans, this isn't just a mouse curiosity — it opens a real, specific path to targeting these exact neuron populations therapeutically, for stress resistance or heart-rate control, without needing to act on the central nervous system at all. It's also a reminder of how little is still understood about this tissue: exactly how researchers pinned down these markers, and what other functions this network might have, are open questions the study doesn't fully answer yet.

Polio spreads through the gut — transmitted orally or via feces — so stopping transmission requires "mucosal immunity": immune defenses stationed directly in the intestinal lining, where the virus actually lives, rather than relying on the blood-based immunity (T cells, B cells) most people picture. The oral vaccine (OPV) induces this kind of gut immunity using a live, weakened virus, but that live virus can occasionally mutate back toward a dangerous form as it replicates in the gut — "vaccine-derived polio." The injected vaccine (IPV) avoids that risk by using killed virus, but only protects the individual from disease — it doesn't stop them from still carrying and spreading the virus.

The piece's actual point is that the remaining polio cases in Pakistan and Afghanistan aren't a story about vaccine science failing — both vaccines work as designed. It's a story about logistics: conflict disrupting vaccination campaigns, and vaccine skepticism in affected populations, not a gap in the biology.

This matters because it points at a different kind of "Plan B" than a scientific one — no next-generation vaccine fixes a delivery and trust problem. It's a reminder that the last mile of eradicating a disease is frequently a public health and political challenge, not a biomedical one, and throwing more vaccine science at it won't be what closes the gap.

An exome analysis of over 1 million people found that rare mutations disrupting the FNIP1 gene track with a roughly 60% lower risk of cardiometabolic disease, plus a favorable body composition — more muscle, less abdominal fat. Follow-up experiments in mice and human liver cells showed that suppressing the FNIP1 pathway directly triggers fat breakdown and improves insulin sensitivity, tying the genetic association to a real, testable mechanism rather than leaving it as a statistical correlation.

This isn't just a genetic correlation — the mouse and cell experiments show FNIP1 disruption causes the metabolic benefit, not just tracks alongside it. That moves FNIP1 from "associated with" to "a validated target for" metabolic disease, which is a meaningfully higher bar of evidence than most genome-wide association findings clear.

Cardiometabolic disease affects hundreds of millions of people. A target this well-validated — human genetics at population scale, plus a demonstrated mechanism — is a rare, high-confidence starting point for a real drug program, not just an association to be followed up on someday.

A correspondence piece arguing that Europe's heatwave-response planning needs to put health-system capacity — not just general emergency response — at its center, since climate, disease, and conflict pressures increasingly compound across borders together rather than arriving as isolated crises.

The framing, not a new data point: it argues current planning treats heatwaves as a weather emergency rather than a health-system load event, which understates what's actually needed on the ground.

If hospitals and clinics aren't explicitly planned around as the actual bottleneck during a heatwave, response plans risk looking adequate on paper while failing where it matters most — at the point of care.

A News Feature on researchers hunting for ways to enhance the body's own DNA repair machinery to extend healthy lifespan, inspired by long-lived animal species and human centenarians who appear to repair DNA damage unusually well.

The specific angle is cross-species: rather than starting from human aging biology alone, researchers are working backward from animals that already solved this problem over evolutionary time.

If DNA repair efficiency really is a common thread across long-lived species and long-lived humans, it becomes a plausible, biologically-grounded target for longevity interventions — rather than one of the many aging theories that don't generalize across species.

Researchers designed covalent caspase inhibitors that exploit pores formed by the protein GSDMD (gasdermin D) to selectively enter and rescue cells from pyroptosis — a violent, inflammatory form of cell death — without blocking normal, non-inflammatory cell death (apoptosis). In a mouse model of endotoxic shock, the approach curbed release of the inflammatory signals IL-1β and IL-18.

The selectivity is the key design feature: it targets the specific pore that forms during pyroptosis to get the drug inside only the cells undergoing this inflammatory death, rather than broadly suppressing cell death pathways the body still needs.

Pyroptosis-driven inflammation is implicated in a range of severe inflammatory diseases; a way to shut it down selectively, without collateral damage to normal cell death, is a genuinely new route to treating them.

A drug combination restored blood stem-cell function in mouse models of sickle-cell disease by reversing premature senescence (aging) of hematopoietic (blood-forming) stem cells — a mechanism also observed directly in samples from human sickle-cell patients.

Reframes sickle-cell disease partly as a stem-cell-aging problem, not purely a hemoglobin-structure problem, and shows that aging-reversal drugs (not disease-specific ones) can meaningfully help.

If premature stem-cell aging is a real contributor to sickle-cell severity in humans, it opens a treatment angle distinct from existing gene-therapy and hemoglobin-targeted approaches, and potentially applicable to other stem-cell-depleting conditions.

A News & Views piece on research showing that privacy attacks on machine-learning models trained on medical records can reveal whether a specific patient's data was used in training — and that patients who differ from the majority population are the most vulnerable to this kind of re-identification.

The inequity angle is the real finding: privacy risk from medical AI isn't evenly distributed, it concentrates on people who are already underrepresented in the training data.

This means standard, population-average privacy safeguards can systematically under-protect exactly the patients most vulnerable to re-identification — a real gap for anyone deploying medical AI at scale to account for.

A career-feature interview with behavioural neuroscientist Shaira Berg, who studies PTSD and OCD using VR experiments. She found that immersion strengthens habit-based threat responses while goal-directed responses stay constant, and pairs this research with filmmaking and theatre to communicate what trauma actually feels like from the inside.

The finding itself — immersive VR selectively strengthens habitual (not goal-directed) threat responses — is a specific, testable dissociation between two different systems the brain uses to respond to danger.

If habit-based threat responses are what immersion specifically strengthens, that's a plausible mechanistic clue for why real trauma (which is deeply immersive) so often produces automatic, hard-to-override reactions rather than conscious, goal-directed ones.

A new pneumococcal vaccine candidate targets proteins shared broadly across Streptococcus pneumoniae strains, rather than the polysaccharide capsule current vaccines rely on — potentially offering wider protection, including against strains existing vaccines don't cover.

Targeting shared proteins instead of the strain-specific capsule is the core design shift — it trades some of the precision of capsule-based vaccines for much broader strain coverage.

Current pneumococcal vaccines only protect against the specific strains included in their formulation, and new strains keep emerging to fill the gap left by vaccinated ones ("serotype replacement"). A truly strain-independent vaccine would sidestep that whole problem.

Researchers use two very different methods to estimate heatwave deaths: statistical excess-mortality modeling (comparing total deaths in a period to a historical baseline) versus attributing individual deaths directly to documented heat exposure. The piece explains why these approaches produce very different tallies as heatwaves intensify across Asia and Europe.

The methodological gap itself is the story — direct-attribution counts are typically far lower than excess-mortality estimates, because most heat deaths never get formally coded as heat-related on a death certificate.

Which method gets used shapes public perception and policy response — a death toll that looks "modest" under direct attribution can look like a mass-casualty event under excess-mortality modeling, and the true number is almost certainly closer to the latter.

The RNA-binding protein ZFP36L2, mutated in 5–10% of colorectal cancers, acts as a molecular switch that lets cells dedifferentiate into a stem-cell-like state during wound healing and metastasis. Losing ZFP36L2 blocks the usual pattern of metastatic spread, but instead promotes alternative, harder-to-treat forms of cancer cell plasticity.

The double-edged result is the notable part: removing this "switch" doesn't simply stop cancer spread — it redirects it into a different, less tractable pattern, which complicates the naive idea that blocking the switch is straightforwardly beneficial.

Any therapy targeting ZFP36L2 would need to reckon with this tradeoff up front, rather than assuming disabling the switch is a clean win — a cautionary, useful finding for drug development in this space.

Amygdala astrocytes (a type of brain support cell) develop shortened primary cilia — small antenna-like sensory structures — during stress, and restoring cilia-related signalling via the receptor S1PR1 reversed stress-related molecular and behavioural changes in mice.

Astrocyte cilia specifically, rather than the neurons themselves, turn out to be a controllable lever over stress behavior — support cells acting as active participants in brain states, not passive scaffolding.

Because restoring cilia signalling reversed the stress-related changes, this points to a specific, druggable target (the S1PR1 receptor pathway) for stress-related brain disorders, rather than a purely descriptive finding.

Using an ex vivo platform combining electrical stimulation with single-cell genomics on human brain tissue removed during surgery, researchers showed that cortical stimulation strengthens specific neural cell assemblies and triggers cell-type-specific gene-expression programs.

Combining live stimulation with single-cell genomics on real human tissue (not just animal models) is the methodological advance — it directly links an electrical intervention to specific gene programs in specific cell types.

Offers a mechanistic clue for how neuromodulation therapies (already used clinically for conditions like Parkinson's and depression) actually restore cognitive function at the molecular level, not just that they work.

The international Human Cancer Models Initiative generated 665 next-generation patient-derived cancer models (including organoids) from 2,780 donors across 25 cancer types, finding high genetic and epigenetic fidelity to the original tumours and creating a large public resource for cancer research and drug discovery.

The scale and the fidelity check together are what's new — 665 models across 25 cancer types, verified to genuinely resemble the tumours they came from rather than drifting during lab culture, which is a common failure mode for cell models.

A public resource this large, with fidelity actually verified rather than assumed, lowers the barrier for any lab studying a given cancer type to work with realistic models instead of building their own from scratch.

Astrocytes expressing CD40 and MHC-II directly interact with CD4+ T cells to promote autoimmune attacks on the central nervous system in a multiple sclerosis mouse model. The interaction triggers lipid-droplet accumulation inside astrocytes that drives inflammatory antigen-presentation signalling — a newly identified mechanism behind MS-like autoimmunity.

Astrocytes actively presenting antigens to T cells — a role usually associated with dedicated immune cells, not brain support cells — is the surprising part.

If this lipid-droplet-driven signalling pathway is confirmed in human MS, it's a specific new mechanistic target — distinct from existing MS therapies, which mostly target T cells or B cells directly rather than the astrocytes helping to activate them.

Analysing multi-omic data from over 1,100 hospitalized COVID-19 patients, researchers found that herpesviruses and anelloviruses frequently reactivate during acute illness — even in people with otherwise normal immune function — and that this reactivation tracks with disease severity and is specifically linked to long COVID.

That reactivation happens even in patients with normal immune function is the surprising part — this isn't just immunocompromised patients losing control of dormant viruses, it's a broader phenomenon tied to the severity of the COVID infection itself.

If viral reactivation is a real contributor to long COVID rather than just a correlated side effect, it opens a specific, testable mechanism — and a specific treatment angle (antivirals targeting the reactivated viruses) — for a condition that's otherwise been mechanistically murky.

A synthetic, pH-responsive peptide triggers a novel form of immunogenic tumour cell death by rupturing lysosomal membranes and then plasma membranes in sequence, boosting anti-tumour immune responses and substantially improving the effectiveness of immune checkpoint blockade therapy in mice.

The sequenced, two-step membrane rupture (lysosome first, then plasma membrane) is a specifically engineered mechanism, designed to trigger the kind of cell death that alerts the immune system, rather than a "quiet" death that goes unnoticed.

Checkpoint blockade therapy only works well when the immune system is already primed to notice a tumour; a peptide that reliably triggers immune-visible cell death could meaningfully widen which tumours respond to checkpoint drugs.

Researchers built and CRISPR-screened 256 patient-derived tumour organoids spanning five cancer types, mapping gene dependencies and finding that different KRAS mutation variants respond differently to EGFR-RAS-MAPK pathway drugs in colorectal cancer — a resource meant to advance precision oncology.

The KRAS-variant-specific drug response is the sharpest concrete result: not all KRAS mutations behave the same way against the same drug class, which argues against treating "KRAS-mutant" as one uniform category for treatment decisions.

If specific KRAS variants predict specific drug responses, that's directly actionable for treatment selection — a step toward genuinely personalized colorectal cancer therapy rather than one-size-fits-all KRAS-targeted treatment.

By selectively blocking opioid receptors on a specific class of neurons in the brain's reward centre (the nucleus accumbens), researchers eliminated morphine's rewarding, addiction-related learning effects in mice while leaving its pain-relieving effects intact.

The separation itself is the finding: pain relief and addictive reward-learning turn out to run through at least partially distinct neural circuits, meaning one can be disabled without the other.

This is exactly the kind of mechanistic separation a safer opioid painkiller would need to exploit — a specific, testable strategy toward pain medications that don't carry the same addiction risk as current opioids.

A gene variant inherited from Neanderthals is linked to greater height and muscle mass in some modern humans, and is most common among people of South and East Asian ancestry.

Adds to a growing list of specific, functionally meaningful traits traced to Neanderthal-inherited DNA still present in modern human genomes, rather than treating that ancestry as evolutionarily inert.

Reinforces that ancient interbreeding left behind DNA with real, measurable physiological consequences today — not just genetic trivia, but variants actively shaping body composition in living people.

Researchers found that the multidrug-resistant fungus Candida auris exposes chitin (a component of its cell wall) on its surface to deliberately trigger interferon-gamma immune signalling — a mechanism that appears to help the fungus persist specifically in hair follicles on skin, a major reservoir behind dangerous outbreaks.

The fungus isn't hiding from the immune system, it's actively provoking a specific immune response that turns out to favor its own persistence — a counterintuitive strategy where triggering inflammation is the pathogen's advantage, not its downfall.

Candida auris outbreaks are a serious, growing hospital-acquired infection problem partly because colonized skin (especially hair follicles) is hard to fully decolonize; understanding the specific immune mechanism it exploits is a concrete step toward better decolonization strategies.

A report finds that an Ebola outbreak in a Congolese mining town went undetected for months amid local fear and mistrust, with hundreds of deaths occurring before the outbreak was officially declared on 15 May.

The months-long detection gap, driven specifically by community fear and mistrust of health authorities rather than a lack of testing capacity, is the core finding.

The failure here wasn't a scientific or diagnostic one, it was a trust and surveillance failure — local fear of health authorities delayed detection long enough for hundreds of deaths to occur first.

A once-weekly oral HIV prevention pill is advancing toward approval, even as a major vaccine trial raised new questions after potent antibodies researchers hoped a vaccine could elicit failed to actually prevent HIV infection.

The vaccine setback is the sharper story: antibodies that looked potent enough in the lab still didn't translate into real-world protection, which complicates a leading strategy for HIV vaccine design.

Progress on prevention (the weekly pill) and progress on a vaccine are running on very different timelines right now — worth not conflating "HIV prevention is advancing" with "an HIV vaccine is close," since those are currently two separate stories.

New regulations restricting risky "gain-of-function" pathogen research are drawing mixed reactions from scientists — some praising tighter biosafety oversight, others warning the rules could hinder legitimate research.

A governance story, not a scientific one — the actual content is about where the new rules draw the line between "dangerous enough to restrict" and "legitimate research that shouldn't be caught in the net."

How this line gets drawn has real downstream effects on which pathogen research actually gets done, by whom, and how fast — a policy decision with direct consequences for pandemic-preparedness research specifically.

NextLongIso unifies long-read RNA sequencing analysis (from PacBio/Oxford Nanopore platforms) — alternative splicing, isoform switching, transcript boundaries, and transposable-element transcription — into one reproducible workflow, replacing a patchwork of single-purpose tools researchers previously had to stitch together themselves.

The unification itself, built on the widely-used Nextflow workflow framework for reproducibility across different computing environments.

Long-read RNA-seq is increasingly the standard for studying isoform-level gene regulation; a single reproducible pipeline lowers the barrier for labs without dedicated bioinformatics staff to actually use it well.

AniAnn's is an alignment-free algorithm that uses average nucleotide identity between repeat units to quickly and accurately annotate satellite DNA and other tandem repeat arrays across plant and animal genomes, improving accuracy while requiring only a fraction of the runtime of previous alignment-based methods.

Skipping sequence alignment entirely — historically the slow step in repeat annotation — is the specific speed-up mechanism.

Tandem repeats, including centromeric DNA, are notoriously hard to annotate well; a genuinely faster, more accurate method is directly useful infrastructure for genome assembly projects.

DeepGeSeq is a deep-learning library that simplifies building and applying genomic sequence models — for tasks like sequence-activity and variant-effect prediction — via a simple configuration file, validated on use cases including single-cell ATAC-seq clustering and MPRA-based regulatory-element analysis.

Lowering the technical bar to configuration-file-only, rather than requiring custom model code, is the practical contribution.

Genomic deep learning has a real expertise barrier; tools like this widen who can actually apply these models to their own data without becoming a deep-learning specialist first.

A new image-guided computational method addresses the resolution gap in spatial omics technologies by using multi-modal image data to computationally recover finer biological microstructures than the physical capture platform alone can resolve, while handling the inherent data sparsity and noise of these measurements.

Using image data to computationally boost resolution beyond the physical instrument's native limit, rather than waiting on hardware improvements.

Spatial omics resolution is a genuine bottleneck across the field; a computational workaround extends the useful life and reach of existing spatial platforms without new instrumentation.

A label-efficient machine learning method improves biomarker prediction from pathology whole-slide images by augmenting frozen, pretrained features, aiming to make computational pathology practical even when labeled training data is scarce and costly to obtain.

Working with "frozen" (unmodified) pretrained features rather than fine-tuning them is what keeps the method usable with very few labeled examples.

Labeled pathology data is expensive and slow to produce, since it requires expert pathologist annotation; a genuinely few-shot-capable method removes a real practical barrier to deploying computational pathology in smaller labs and hospitals.

Studying T-cell receptor to peptide binding prediction, researchers show that naively combining multiple data modalities (like sequence embeddings and structural predictions) can actually hurt model performance when the auxiliary data is noisy or incomplete, and propose contrastive alignment as a way to stabilize multimodal fusion.

The negative result is genuinely useful here: demonstrating that "more data modalities" can actively hurt performance pushes back on a common, often unexamined assumption in the field.

TCR-peptide binding prediction underlies immunotherapy and vaccine design work; a documented failure mode (plus a fix) saves other groups from repeating the same mistake with noisy auxiliary data.

Modified baker's yeast can now produce the plant-derived precursor for the chemotherapy drugs etoposide and teniposide, offering a scalable alternative to sourcing it from the endangered Himalayan mayapple plant and its relatives.

Getting yeast to produce this specific, structurally complex plant compound at scale — a genuine synthetic-biology engineering achievement, not a small tweak to existing yeast chemistry.

Solves two problems simultaneously: a more reliable, scalable drug-precursor supply for widely-used chemotherapy drugs, and removing pressure on an endangered wild plant population that current supply chains depend on.

A large systematic effort ("Codebook") profiled the DNA-binding specificity of 332 poorly characterized human transcription factors across more than 4,000 experiments, identifying binding motifs for 177 of them and revealing tens of thousands of previously unknown transcription-factor binding sites, concentrated in gene promoter regions.

The scale is the story: 332 transcription factors that had gone essentially unstudied, now systematically characterized in one coordinated effort rather than piecemeal over years.

Transcription factors are the proteins that turn genes on and off; tens of thousands of newly identified binding sites is a substantial expansion of the map researchers use to understand gene regulation in both health and disease.

A new 3D-printing technique called SHIFT transforms hydrogel particles into long, aligned microfibres during extrusion, creating structural anisotropy (direction-dependent properties) that produced unusually long muscle fibres in vitro and improved regeneration in a mouse model of severe muscle loss.

Achieving fibre alignment during the printing process itself, rather than as a separate post-processing step, is what lets the technique produce unusually long, well-organized muscle fibres.

Muscle tissue's function depends heavily on fibre alignment, which most bioprinting techniques struggle to replicate; a printing method that gets this right natively is a meaningful step toward functional muscle regeneration therapies.

VirTues is a new foundation model for spatial proteomics that learns unified representations of proteins, cells, and tissue niches across different imaging panels and datasets. In triple-negative breast cancer, it predicted response to immunotherapy and patient survival better than existing biomarkers derived from the same underlying data.

Working across different imaging panels and datasets without needing to be retrained for each one is the foundation-model advance — most existing spatial proteomics models are locked to a specific panel/dataset combination.

Outperforming existing biomarkers on immunotherapy response and survival prediction, using the same underlying data those biomarkers were built from, suggests there's real predictive signal in spatial data that current clinical biomarkers are leaving on the table.

Researchers developed a chemical screening platform that discovered M12, a "molecular glue" drug that only becomes active once inside a cell, via a glutathione-dependent chemical reaction, redirecting the DCAF11 protein-degradation machinery to destroy a previously undruggable target protein.

The drug is deliberately inert until a specific intracellular reaction activates it — a built-in mechanism for targeting activity to inside cells, and for expanding "undruggable" targets by hijacking the cell's own degradation machinery instead of binding the target directly.

Molecular glue degraders are one of the most active areas in drug development precisely because they can reach targets that don't have a good conventional binding pocket; this expands the toolkit with a genuinely new activation mechanism.

Researchers discovered that cells sense low haem (the oxygen-carrying component of haemoglobin) inside mitochondria via a signalling pathway that ultimately controls protein synthesis during red blood cell formation. Manipulating this pathway boosted fetal haemoglobin production — a therapeutic goal in diseases like sickle cell disease and thalassaemia.

Tracing haem-sensing specifically to a mitochondrial signalling pathway, and showing it's an "ancient" (evolutionarily conserved) regulatory program, gives a concrete mechanistic handle on something previously understood only at a phenomenon level.

Boosting fetal haemoglobin is already a validated therapeutic strategy for sickle cell disease and thalassaemia (it's how some existing gene therapies work); a new, mechanistically distinct pathway to achieve the same goal is a real additional treatment avenue.

Using genome language models, researchers generated complete, viable bacteriophage genomes from scratch with a chosen host-targeting range, based on the phage ΦX174 as a design template. 16 of the AI-designed phages were experimentally viable, and cryo-electron microscopy confirmed their physical structure matched the design.

This isn't just predicting or editing an existing genome — it's generating a complete, functional viral genome from a language model and having it actually work in the lab, with structural confirmation, not just a plausible-looking sequence on paper.

Real capability, real stakes: viable AI-designed phages could accelerate phage-therapy development against antibiotic-resistant bacteria, but the companion perspective piece is explicit that being able to computationally design a working viral genome at all raises serious biosafety and biosecurity questions that arrive at the same time as the capability itself.

A Research Briefing on a global survey of centromere sequences — the chromosome regions essential for accurately dividing DNA into daughter cells — across 65 diverse individuals, showing that centromeres evolve much faster than the rest of the genome, both between generations and over hundreds of thousands of years.

The rate of change is the finding: centromeres were assumed to be relatively conserved, functionally-critical regions, but this survey shows they're actually among the fastest-evolving parts of the human genome.

Centromere sequence has historically been one of the hardest parts of the genome to even assemble accurately; a real diversity map across individuals is foundational reference data other centromere and chromosome-segregation research can now build on.

Topol interviews Sarah Urbut about her Nature paper introducing ALADYNOULLI, an AI model that combines electronic health records and polygenic risk scores across three cohorts — 683,000 people, 348 diseases, up to 52 years of follow-up — to generate continuously-updating individual disease-risk trajectories.

The "GPS rerouting" framing is the real conceptual shift: instead of a one-time risk score calculated at a single point in time, the model continuously updates an individual's risk trajectory as new health data arrives — a fundamentally different way of thinking about disease risk.

A static risk score becomes stale the moment new relevant information exists; a continuously-updating trajectory model is a genuinely different product for both patients and clinicians — closer to ongoing monitoring than a single verdict.