Last issue was about capabilities that already work, with the open question being what happens once you deploy them. This week turns to the tools themselves: three stories about instruments that let researchers move one thing and watch what follows, and about the limits of the system that thing sits in.
Biological engineering is escaping the tug of war with the cell, either by removing the cell and working in a tube, or by letting millions of cells run random experiments and reading the winners out as DNA barcodes. The gut microbiome turns out to be a lever nobody had measured: it records ancient migrations, predicts who gets a fever after a vaccine, and makes a molecule that works on the same liver enzyme as statins. And neuroscience has just crowned its best lever, optogenetics, while the harder questions about how brains age remain open.
Below the three focus stories, Also This Week collects cancer results from ecDNA to a metal-framework vaccine, human tissue models, outbreaks and preparedness, computational tools, integrity worries and two voices outside the journals.
Michael Jewett, a bioengineer at Stanford, describes his working life as a tug of war with biology. He needs microbes to make a compound or follow instructions he has written into their DNA, but cells evolved to survive and reproduce, so they silence inserted genes, burn resources on their own needs, or die from the product itself. His answer, in a Nature feature this week, was to 'cut the rope': break the cells open and use the contents as a molecular factory. Cell-free protein expression is not new — it helped crack the genetic code in the 1960s — but it is having a moment, because it shortens the design-build-test cycle from days or weeks to hours and can make things that would poison a living host.
The price is real. Reagents for a litre-scale reaction can cost more than $4,000, most studies report under two grams of protein, and a reaction runs out of energy within hours. Jewett's lab tested 1,231 recipes for the 'stew' that feeds a bacterial lysate and found that 12 reagents are enough, cutting the cost by 95% to under $100 per gram. At the cleanest end sits the PURE system, rebuilt from the 36 non-ribosomal E. coli proteins needed to read DNA and make protein; the one part nobody can yet make in the tube is the ribosome itself, which is why the field's next milestone is a self-sustaining system that builds its own ribosome. Sensors show the payoff: Julius Lucks' freeze-dried ROSALIND kits detect copper and zinc in contaminated municipal water, and a machine-learning-guided cell-free screen produced a sensor that picks out lead at 5.7 parts per billion without being fooled by zinc.
The opposite strategy keeps the cell and lets it do the searching. In shotgun genetic engineering (SGE), Julie Trolle and colleagues pour a pool of small barcoded gene cassettes onto mammalian cells at high multiplicity, so each cell ends up with its own random combination of genes, promoters and localization tags. A selection (here, growth without an amino acid animals lost half a billion years ago) keeps the cells that work, and sequencing the barcodes reads out the winning combination. Designing an equally large multigene construct by hand would run into the problem that delivery gets worse the bigger the DNA is. In hamster cells SGE found a valine pathway that let them double every 1.1 days in valine-free medium, versus 3.8 days for a pathway the same group had designed rationally, and it delivered the first isoleucine pathway in a mammalian cell. Winning designs were enriched for enzymes sent to the mitochondria, a design variable rational planning had mostly ignored, and a classifier trained on the barcode data rediscovered the same rules. The accompanying News & Views calls it a way to find 'functional needles in the combinatorial haystack'.
Two further Nature Biotechnology papers this week share the logic of reading out biology by sequencing. Detectrons fuse a toehold RNA sensor to a bacterial retron so that when a target RNA appears, the cell writes a unique DNA barcode: the best design raised the on/off ratio 14.7-fold over the first one, and in a pooled E. coli culture the barcodes identified which phage had infected and how different LPS mutants resisted T7. And a Chinese team mined 1,139 bird genomes for 159 new R2 retrotransposons, then engineered one into an all-RNA gene inserter that placed a gene into about 80% of HEK293T cells under the best delivery conditions and over 60% of primary muscle cells and fibroblasts — but only at ribosomal DNA sites, and with expression fading in T and NK cells.
Biology's hardest engineering problem is not writing the design, it is knowing which design to write. Three different moves attack that from different sides: removing the host so it stops fighting you, letting a living population run millions of experiments at once, and reading the results out as DNA sequence. They trade off honestly — cell-free is fast but exhausts itself in hours, SGE needs a phenotype you can select for, and the R2 inserter works only at one genomic address — but together they make 'try everything and read the barcodes' a realistic way to build.
Three results this week treat gut bacteria as something that explains an outcome rather than merely accompanies it. The oldest is a story about human history. Carter and colleagues sequenced stool from the Tsimane, forager-farmers in the Bolivian Amazon, and compared it with Hadza hunter-gatherers in Tanzania, two groups whose ancestors separated tens of thousands of years ago. Of the 1,408 microbial species found in the Tsimane, 1,231 (87.4%) also occur in the Hadza, and 848 of those are rare or absent in industrialized populations. The strains tell the migration story: of 636 well-sampled species, 545 show no shared strain lineage in roughly the last 5,000 years, and gene flow between the two groups' bacteria appears to have stopped around 17,000 years ago, far earlier than the 4,600 years estimated between Asia and North America or the 1,100 between Europe and North America.
The commentary is careful about what that does and does not show. Seventeen thousand years is probably an underestimate of the human split, which is thought to be much older; mutation rates differ between species and bacteria swap DNA, which blurs the dating; and only two non-industrialized populations were studied in depth. The practical sting is the 848: many microbes with a long history alongside people are disappearing from industrialized populations, possibly through antibiotic exposure, diet and sanitation. The authors of the commentary stress that a long association does not prove a species is beneficial, and that putting lost microbes back would need functional evidence, safety testing and real partnership with the communities who provided the samples.
Whether any particular microbe does something useful is exactly what the other two results test. In Science, researchers followed 171 healthy adults receiving SARS-CoV-2 mRNA vaccines and found that fever risk tracked pre-vaccine low-grade intestinal inflammation and an abundance of flagellated Lachnospiraceae bacteria expressing flagellin. Microbiomes from 'fever-high' donors triggered stronger inflammation in human intestinal organoids and drove flagellin-dependent vaccine reactions in germ-free mice, and both the microbiome's flagellin traits and the mouse reactions could be shifted by diet. A companion Science perspective summarizes it as the microbiome before vaccination influencing later inflammatory reactions and fever.
Then there is the heart. A Nature study found that Bacteroides uniformis was depleted in people with heart disease, and that in mice engineered to develop atherosclerosis it lowered cholesterol and LDL and shrank arterial plaque (the news report gives figures between about a third and about a half), with fewer inflammatory cells in the artery wall. The molecule responsible is pentadecanoic acid, which inhibits a liver enzyme in cholesterol production that statins also target, prompting the liver to make more LDL receptors that pull cholesterol out of the blood. In more than 200 human donors, people with abnormally high blood lipids had lower pentadecanoic acid and fewer bacterial genes for making it. The independent experts quoted are enthusiastic but blunt: the mice received far more bacteria than a person could get from food or supplements, standard mouse models rarely reproduce the lesions that cause heart attacks, and pentadecanoic acid is already sold as a heart-health supplement even though human evidence is mixed.
A person's gut community now appears in three roles at once: a record of ancestry, a predictor of how your body reacts to a vaccine, and a source of a statin-like molecule. All three are still correlations or mouse experiments, which is the honest limit, but they point the same way — that part of what looks like individual variation in drug and vaccine response, and in heart risk, is microbial — and that the industrialized world has lost hundreds of species before anyone measured what they did.
The 2026 Nobel Prize in Physiology or Medicine went to optogenetics, the technique for turning specific neurons on or off with light. It is shared equally (12 million Swedish kronor) by Karl Deisseroth, Peter Hegemann and Georg Nagel. In the early 2000s Hegemann and Nagel found channelrhodopsin, a light-sensitive channel from the alga Chlamydomonas reinhardtii that opens in blue light and lets charge flow into cells. Deisseroth's group then put the gene into rodent neurons and showed that light could make them fire and signal to other cells — a switch that works in the brain of a living animal, on chosen cell types, with millisecond timing. Before that, researchers mostly stimulated tissue with electricity, which hits large populations of neurons indiscriminately.
What took so long? Michael Häusser told Nature's reporter the field had been 'waiting for this for a long time'; optogenetics won its first major award, the Brain Prize, in 2013, shared also with Edward Boyden, Gero Miesenböck and Ernst Bamberg, and the Nobel's three-person limit left those names off. His guess is that the committee waited for human medicine to catch up: trials have partly restored vision in people with the genetic blinding disease retinitis pigmentosa, four companies have trials running (one, Nanoscope Therapeutics, is seeking FDA approval), and the technique now works in non-human primates. Häusser called it an illustration of how an unexpected discovery in an alga moving towards light can reshape a field.
Optogenetics asks 'what happens if I move this one thing?' Two other brain stories this week ask the harder question of how to move something safely. A team led by Rasheen Powell at Boston Children's Hospital designed a peptide that breaks the bond between pain-signalling sodium channels and the anchor proteins holding them at the nerve cell surface (the paper in Cell Reports Medicine concerns the Nav1.8/ankyrin-3 interaction in pain-sensing neurons). In mice with nerve injury or inflammation it lowered abnormal firing and pain behaviour while leaving mobility intact, and in mice without chronic pain it did not blunt protective responses to harmful stimuli. The target proteins are raised in human nerve injury, and the authors see a blueprint for a non-addictive painkiller; it is, so far, a mouse result.
The other pair asks what age and hormones do to a brain. A study of 601 apes — chimpanzees, gorillas, orangutans and bonobos — found no significant drop in brain weight in older animals (63 were over 50), unlike humans, whose brains start to lose weight at around 45 to 50 even without disease; the authors suggest the marked decline is mainly a human feature, perhaps tied to our longer lifespan. And on the other side, Science reports that menopause may act as a stress test for the brain: a Nature Medicine study last week found blood proteins linked to brain function, inflammation and neurodegeneration raised in women in their 40s and 50s during menopause, and other cohorts are following women through the perimenopause to separate short-term 'brain fog' from long-term dementia risk. A fourth piece, Nature's feature on whether adult brains grow new neurons, shows how contested that old question still is: the evidence rests mostly on genetic markers, Arturo Alvarez-Buylla says he does not see it in the data, and Gerd Kempermann sees the field entering a 'consolidating phase'.
The Nobel rewards a tool that made brain cause-and-effect testable, and this week's brain news shows both how far that has reached and how little is settled. We can now switch one cell type on with light, and propose a peptide that quietly lowers one pain channel's signal — yet we still argue about whether adults make new neurons, why only human brains shrink so reliably, and what menopause does to the proteins in the blood. Tools are running ahead of the biology they were built to explain.
Four cancer results sit together this week. In Nature, blocking the repair enzyme Polθ made extrachromosomal DNA (ecDNA) circles carrying MYC vanish from colon cancer cells within a week, while blocking two other repair pathways did not: ecDNA breaks at stretches rich in repeated T and A letters, the protein FANCM guards them, and the damage repaired by microhomology-mediated end joining keeps the circles alive. A separate lung-cancer paper from Stanford is covered further down. On the vaccine side, a metal-organic-framework cancer vaccine, MOF-Vac, delivers peptide antigens and a STING agonist into the tumour; in mice a single injection nearly eradicated B16 melanoma (99.4% growth inhibition, complete regression in five of six), shrank an untreated distant tumour, and worked through macrophages rather than dendritic cells. And a pre-cancer paper links a brake-like T cell, the effector regulatory T cell, to high-grade airway lesions: blood levels of these cells rose with lesion grade, predicted faster recurrence in early lung cancer (hazard ratio about 3 to 4), and a PI3K-delta inhibitor cut tumour incidence in mice from 83% to 42%.
For context, Nature's piece on the past 50 years of cancer progress, drawing on an American Association for Cancer Research report, says 71% of US people with cancer now live at least five years, up from 49% in the late 1970s; lung cancer's five-year survival rose from 13% to 29%, and childhood lymphoid leukaemia from 57% to 90%, largely by tweaking doses and timings of existing drugs.
Researchers keep building human-cell systems to replace what animals cannot show. A Stanford group grew lung organoids from intact tissue fragments (from over 220 patients) that keep their own T cells; infected with SARS-CoV-2 they made virus-specific, tissue-resident T cell responses, and pre-treating them with a SARS-CoV-2 peptide pool lowered viral transcripts by 35.8% and plaque-forming virus by 44.7%, versus rises with an HIV peptide control — evidence for local protective immunity without lymph nodes. A placenta-on-a-chip of primary placental cells showed cadmium raising barrier permeability and cytokines at 5 micromolar, below the dose that kills cells, and cutting glucose transfer at 1 micromolar. An apallial mouse lacking most of its own cortex let transplanted human cortical organoids make up about 92% of cortical tissue at three months. Two Nature Biomedical Engineering reviews map where gut organoid-on-chip models and multi-organ chips for drug testing still fall short — mostly culture lifetime of days to weeks, missing immune cells and a lack of standardization.
A 28-year-old researcher died on 2 October at a Siberian plague-research institute in Irkutsk, possibly of pneumonic plague; about 200 colleagues were reportedly quarantined, and Russia revealed only on 7 October that she had been vaccinated and that thousands of tests were negative. Gregory Koblentz tells Nature the real risk of a wider epidemic is minimal — plague is treatable with antibiotics — and that the lesson is the cost of opaque government communication. The Bundibugyo Ebola outbreak declared on 15 May had reached 7,890 cases and about 3,799 deaths (around 48%) in the DRC by 23 September, with no licensed vaccine or cure, though two of four candidate vaccines are in trials. Early data point to an unusually early flu season, with H1N1 dominant and a mutation under watch. Longer-term, Nature profiles CEPI's work on 'Disease X', asks whether the world can still end AIDS by 2030 (41 million people live with HIV; the UN target is a 90% cut in new infections), and argues that universities need more agile science for crises, pointing to how COVID-19 structures, trials and vaccines were produced quickly.
A scoping review in Nature Medicine of 3,967 clinical-prediction-model papers found that only 482 (12.2%) shared code, and of 380 repositories checked, 80.5% had a README but only 3.9% had tests and 37.6% specified dependencies; papers citing TRIPOD+AI shared code nearly three times as often as those citing TRIPOD. In Bioinformatics this week: an amyloid predictor, VAEBAC, whose authors openly report that its latent space encodes sequence length (AUC 0.904, falling to 0.868 once length is removed), a modular decoder framework for spatial transcriptomics (Serval, with the new CoSiD decoder cutting blank-barcode detections from 4,801 to 1,020), a drug-target model that stays useful when only 10% of positives are labeled (ProDTI) and a tool, RAFA, that rebuilds full RNA atoms from sparse anchors. The common thread is candour about confounds, which is worth more than the headline score.
Nature staff discuss a network of organizations that has been growing its membership through deceptive practices and drawing in prominent researchers, and Nature has attached an Editorial Expression of Concern to a published paper reporting that maternal gut bacteria promote neurodevelopmental abnormalities in mouse offspring. Both are reminders that the credibility signals scientists use, society memberships and published papers, can be worn down.
Eric Topol's Ground Truths hosted the first interview with Michael Lewis about his new book BLOCKERS, calling its NIH chapter a murder mystery that differs from what was previously reported (the full transcript is clipped, but the summary here comes from Topol's own write-up). And Decoding Bio's BioByte 174 covers AlphaProtein Novo, a DeepMind, Caltech and Pittsburgh de novo enzyme-design pipeline tested on over 5,000 designs: for a reaction that normally yields a five-membered ring, its best design produced a six-membered piperidine with nearly 99% selectivity, and its plasticizer-degrading designs, with an 11% hit rate, were less active than natural enzymes but easier to express and tolerant of high temperature.