Where last week was about treatments outrunning the biology meant to explain them, this week the settled biology itself keeps not holding. The clearest case is the ageing brain: two independent teams, publishing a week apart, both show that the immune cells in an old human brain are mostly not brain cells at all — they are blood monocytes that crossed in after about age 50 and took over, carrying a pro-inflammatory wiring pattern with them. The dogma that microglia are laid down before birth and last a lifetime was true in mice and simply wrong in people, and nobody caught it until someone looked at DNA methylation instead of gene expression.
A modified virus injected into the bloodstream built CAR T cells inside 16 people and appeared to reset the immune systems driving their multiple sclerosis — no bioreactor, no per-patient manufacturing line. And daraxonrasib, approved for pancreatic cancer only a week earlier, is already shrinking lung tumours: a drug against a target the field spent decades calling “undruggable” now looks like it might work across cancer types.
Researchers mutated nearly every base of bacteriophage ΦX174 — the first genome ever sequenced, studied for fifty years — and still can’t say why a quarter of the lethal mutations are lethal. The AI models built to predict exactly this struggled. Below the four focus stories, Also This Week pulls together a set that rhymes with them: three drugs that switch on only where the disease is, two ways of interrogating the genome’s control layer, two arguments for reading omics data more carefully, and a calculator for stacking heart-failure drugs safely.
For decades the textbook account of microglia — the brain’s resident immune cells, its housekeepers and first responders — has been that they arrive from the yolk sac during embryonic development and then maintain themselves locally for the rest of your life, with almost no input from the blood. That is well established in mice. Two papers published a week apart, by teams that were not in contact with each other, show it is wrong in humans.
The Stanford team, led by Julia Belk, used a technique called somatic-mutation lineage tracing. Every cell accumulates small DNA mutations over a lifetime; shared mutations mark cells that descend from a common ancestor, which makes them a gold-standard way to trace lineage in people. Because those clonal patterns only become common in the very old, the study focused on donors who were mostly over 90, and looked across several brain regions — occipital cortex, putamen, cerebellum. In all of them, a large fraction of the “microglia” turned out to descend from blood cells. This extends a 2023 finding the same group made in people with clonal haematopoiesis to the general ageing population, and it suggests the phenomenon is brain-wide rather than confined to one structure.
The UCSD team, led by Nathan Zemke, came at it from a completely different direction: single-nucleus multi-omics — gene expression, chromatin accessibility, DNA methylation and 3D genome organisation — on 40 human hippocampi spanning ages 20 to the 90s. The decisive layer was DNA methylation, which is rarely done at this resolution. Methylation marks record a cell’s lineage of origin, so even after a blood monocyte moves into the brain and switches on the standard microglia gene-expression program, its methylome still says “monocyte.” That is why every previous single-cell RNA study of the aged brain missed this: by transcription alone, these cells look exactly like microglia. The companion Nature paper, “Somatic mutations reveal the ontogeny of microglia in human ageing” (Belk et al.), and Zemke’s “Epigenetic and 3D genome reprogramming during the aging of the human hippocampus” land on the same picture.
Three things about the replacement cells matter. First, the timing: before 50 nearly everyone still has their original microglia; the shift is a feature of the second half of life, when neurodegenerative pathology is thought to begin. Second, the character: Zemke’s 3D-genome data show the incoming cells are folded so that active regulatory regions sit close to inflammatory genes like interleukin-15, and that folding matches circulating monocytes, not embryonic microglia. It is “environment-resistant” — the cells keep the monocyte architecture, and its inflammatory output, even after settling in the brain. Third, how they get in: astrocytes and blood–brain-barrier endothelial cells decline with age, which may open the door, though that link is still only correlative. Mouse work adds a twist — if you deplete existing microglia, monocytes rush in to fill the niche even with an intact barrier, hinting that the original cells may simply wear out over a human lifespan and trigger the influx.
The most surprising result is about CHIP — clonal haematopoiesis, the age-related expansion of mutation-bearing blood-stem-cell clones, normally flagged as a risk factor for cardiovascular disease and blood cancers. Here it looked protective: people with larger CHIP clones had substantially lower Alzheimer’s risk and less amyloid and tau in the brain. The working hypothesis is that these expanded monocyte clones are simply better at clearing amyloid and tau by phagocytosis once they become brain-resident. That inverts the usual framing and raises the idea of deliberately engineering peripheral monocytes — a “brain immunotherapy of the blood.”
This is a genuinely new route into the ageing brain. If a pro-inflammatory blood population is displacing the resident immune cells on a schedule that overlaps with when dementia risk climbs, then delaying that influx — or seeding it with protective, engineered monocytes, or using those cells as delivery vehicles — becomes a target class that did not exist a month ago. It also lands a hard caveat on the field: this process is human-specific, absent in mice and non-human primates, so a lot of preclinical neurodegeneration work has been modelling a brain that ages differently from ours. It continues the thread from CW33’s coverage of why memory changes after 50.
Sixteen people with multiple sclerosis or other autoimmune conditions — some involving muscle weakness and inflammation, some involving immune attack on the brain, spinal cord and eyes — received a single injection of a modified virus into the bloodstream and were followed for about six months. The idea was to make the treatment do its own manufacturing.
Conventional CAR T-cell therapy, now standard for several blood cancers and in trials for lupus, works by collecting a patient’s T cells, engineering them in a lab over weeks to display a chimeric antigen receptor, and reinfusing them. It is powerful and extraordinarily expensive, and every dose is a bespoke production run. This trial, run by a team including neurologist Dai-Shi Tian at Huazhong University of Science and Technology in Wuhan, used a lentivirus — engineered by the biotech company Shenzhen Genocury — to carry the CAR genetic instructions directly to T cells inside the body. No cells leave the patient. A lentivirus-based in vivo CAR T approach was shown to work against blood cancer last year; this is the proof-of-concept extension to autoimmune disease.
The CAR directs the T cells to attack the B cells producing autoantibodies against the body’s own tissue. Over time the participants generated more CAR T cells, and B-cell numbers and autoantibody levels fell. Crucially, when B cells recovered, the replacements did not make the harmful autoantibodies — the sign the researchers read as an immune “reset” rather than a temporary knock-down. The multiple-sclerosis participants showed improved motor and cognitive function and less fatigue; those with muscle-involving conditions gained strength and had less inflammation. Outside commentators called it “a very exciting proof-of-concept.” Tian is more measured: “The responses are promising signals, but they are not yet definitive evidence of efficacy or permanent restoration of immune tolerance.”
The bottleneck holding cell therapy back is not usually the biology — it is the per-patient bioreactor. Last week’s analysis of the Moderna melanoma vaccine made the same point about personalised mRNA. An in vivo approach, if it holds up, removes that step and could make a CAR T-style immune reset something you deliver from a vial. This is 16 people, open-label, six months, no control arm — a long way from that — but it is the clearest sign yet that the manufacturing problem might be designed around rather than scaled through.
A week after the FDA approved daraxonrasib (brand name Rasonque) for advanced pancreatic cancer — more than six months ahead of schedule — the same drug is showing activity against non-small-cell lung cancer. In a small trial of previously treated patients, each carrying a mutation in a RAS-family gene, tumours shrank in more than 30%. “We consider the results incredibly promising,” said Kathryn Arbour of Memorial Sloan Kettering, an author of the study.
RAS proteins are mutated in many of the deadliest cancers — pancreatic, lung, colorectal — and were long considered undruggable, because their surface offered nothing obvious for a small molecule to grab. The first breakthroughs, the KRAS-G12C inhibitors, target one specific mutant form of one RAS gene. Daraxonrasib, from Revolution Medicines, works differently: it acts as a molecular glue, first binding a common cellular protein called cyclophilin A, then using that complex to grip both normal and mutant RAS in their active state and block them from signalling to their downstream partners. Hitting the pathway broadly, rather than one allele, is what makes a tumour-type-agnostic strategy conceivable.
The pancreatic-cancer data that drove the approval were striking: in about 500 people with advanced disease, daraxonrasib nearly doubled median survival, to 13.2 months from 6.7 with chemotherapy. The lung-cancer results are earlier and come with a warning — some tumours that responded began growing again, so resistance develops, and its mechanisms are still being characterised. A randomised trial pitting daraxonrasib against standard chemotherapy in non-small-cell lung cancer is under way, and regulators will almost certainly want that readout before any lung-cancer approval. “Everyone is now asking the question: what next?” said Channing Der of the University of North Carolina.
This is the first real signal that broad RAS inhibition generalises beyond the tumour type it was approved for — a plausible route to one drug against one of the most common oncogenic drivers across cancers. The response rates are moderate and resistance is quick, so the randomised lung-cancer trial is what will actually settle it. Daraxonrasib is not yet tracked on the site’s clinical-trials dashboard.
Bacteriophage ΦX174 — “phi X” — is about as well-characterised as a genome gets. Its 5,386-nucleotide circular single-stranded DNA, encoding 11 proteins, was the first genome ever sequenced, in the 1970s; the first to be chemically synthesised, in the 2000s; and the template for the first AI-designed viral genomes. A team led by Ben Lehner at the Wellcome Sanger Institute, with Huijin Wei in Barcelona and Xianghua Li at King’s College London, has now done another first with it: they made nearly every possible single change — more than 44,000 variants, covering every one-nucleotide substitution and every single-amino-acid substitution in every protein — and measured what each one does.
The measurement was a competition. Pools of mutant phage were grown with E. coli for 80 minutes, long enough for two or three infection cycles; variants that multiplied were winners, variants with crippling mutations died out, and DNA sequencing of the survivors read out each mutation’s effect on fitness. Around half of the single-nucleotide changes and 60% of the amino-acid changes were harmful — a much higher fraction than expected for a genome generally assumed to be fully optimised for the lab. A handful of mutations actually improved the virus’s fitness.
The part that got attention is what the team could and couldn’t explain. Of the harmful amino-acid mutations, roughly half disrupted interactions with other proteins (phi X’s own and its host’s), and about a quarter hit residues buried in a protein’s structural core. The remaining quarter had no identifiable cause. “There’s hundreds of mutations in here where we haven’t got a clue what they’re doing,” Lehner said. And the cutting-edge biological AI models built to predict which mutations matter — the same class of tool being promoted for interpreting human genetic variants — struggled on phi X, the friendliest possible test case.
It is a reality check on genomic AI. If today’s models can’t reliably predict mutation effects in the most exhaustively studied genome on Earth, confident claims about predicting pathogenicity of human variants deserve more scepticism — and the fix the authors point to is not a cleverer architecture but a great deal more experimental data. The result pairs directly with this week’s pan-histone mutagenesis work: the field’s answer to “what does this change do?” is increasingly to mutate everything and measure, rather than to predict.
Three papers this week share a design principle: instead of homing to a receptor that healthy and diseased tissue both display, they exploit a condition that only the diseased tissue creates. NOCAGE keeps an engineered protein inert by chemically caging an essential glutamate residue; the cage is cleaved only where nitric oxide is elevated, which is what happens at a site of inflammation — demonstrated for inflammation-localised enzyme activity, viral gene delivery and in vivo biosensing (the catch: the protein needs a cageable glutamate in the right place). McR-TACs tackle a limitation of lysosome-targeting degraders, which normally need a specific uptake receptor and are consumed in the process: a polyzwitterion–ligand chimera drives bulk uptake by macropinocytosis instead — a route tumour cells use heavily — degrading surface and secreted proteins receptor-independently, with a carrier that recycles. And the phase 1 GlutaPanc trial (16 patients) adds clinical-grade l-glutamine to first-line gemcitabine plus nab-paclitaxel in advanced pancreatic cancer, the first clinical test of l-glutamine’s preclinical antitumour activity; an accompanying briefing notes survival and response numbers above historical benchmarks, with the mechanism still a hypothesis. Read together, they sketch an alternative to receptor targeting for tissues — inflamed joints, tumours — defined more by their metabolic state than by a clean surface marker.
Two Nature Genetics pieces on the regulatory genome. A News & Views on work by Price and colleagues describes prime editing scaled up to mutate any lysine residue, or combination of them, across all copies of the histone H3 genes at once — moving histone-tail genetics from one-site-at-a-time edits to genome-wide combinatorial studies of what individual chromatin marks actually do. It is the chromatin counterpart to the phi X saturation-mutagenesis approach. Separately, a review of long-range enhancer–promoter regulation works through the puzzle of why the most critical developmental genes are so often controlled by enhancers hundreds of kilobases away: loop extrusion, phase-separated condensates, tethering elements, and a newly proposed “facilitator” class of elements — and why all of this matters for interpreting non-coding disease variants and for engineering gene expression.
Two methods arguments about not over-trusting a model fed messy or thinly characterised data. A semi-supervised method combining protein language model embeddings with Gene Ontology annotations pinpoints which individual residues carry a protein’s function, connecting whole-protein function prediction to site-level mechanism. And a recommendation in Nature Biotechnology argues that multi-omics studies should co-profile standardised reference materials alongside their samples and report sample-to-reference ratios, so that datasets feeding AI/ML models are actually comparable across labs and platforms — a direct answer to the batch effects that deep-learning methods otherwise learn as if they were biology.
A cross-trial analysis in Nature Medicine pools 38,753 participants from nine heart-failure trials to estimate how much comprehensive guideline-directed therapy — and specific combinations of an ARNI, a beta-blocker, a mineralocorticoid-receptor antagonist and an SGLT2 inhibitor — shifts blood pressure, kidney function and serum potassium in the first weeks of treatment. The goal is practical: current guidance is to get patients on all four drug classes quickly, but clinicians often add them one at a time over months for fear of dropping blood pressure or pushing up potassium. Quantifying the expected short-term trajectory of each combination is meant to make starting several at once a more calculable decision.