The throughline across both journals this week is AI moving from a topic science covers to a tool science designs with. AlphaFold3 is now guiding precision base-editor design, generative models are redesigning enzymes into better evolutionary starting points, and separately inventing non-natural CRISPR-like nucleases from scratch. Alongside that, the field is publicly wrestling with what AI does to the publishing process itself and to global research competitiveness — China's Kimi K3, the UK's rise as an AI-safety hub, and tightening US restrictions on China collaboration all surfaced independently in both outlets.
Materials science clustered hard around energy and next-generation computing hardware. Two unrelated groups pushed perovskite solar cells toward real manufacturability from different angles — interfacial chemistry versus scalable printing — while topological photonics, Weyl-semimetal interconnects, and single-electron memory each chip away at physical limits of optics and electronics.
Tertiary lymphoid structures got mechanistic treatments in both journals simultaneously — what builds them, and what they contain. Cancer research spanned scales, from a mouse model showing genetic ancestry steers tumour evolution to a genuinely strange transmissible cancer spreading between wild catfish. Aging research identified a specific, reversible immune-clearance failure behind organ decline.
The earliest known snakes turn out to have been ecologically diverse from the start, cold radioactive molecules are now laser-spectroscopy-ready on a tabletop, and a massive health-data study quantifies just how much the COVID-19 pandemic set back life-expectancy gains across Asia. Read individually these are unrelated papers; read together, they're a snapshot of AI-as-design-tool, materials-for-energy-transition, and immune-system mechanics all accelerating in parallel right now.
The problem: Scientists "evolve" proteins in the lab — make random changes, keep the ones that work better, repeat — to build useful enzymes. But natural proteins are fragile starting points: mutate them and they usually just break.
What they did: This team used an AI model (ProteinMPNN) to first redesign three toxin-derived enzymes to be sturdier without changing their job, then ran the usual lab-evolution process on both the AI versions and the natural originals, side by side.
What they found: The AI-started versions consistently evolved into better enzymes and reached improvements the natural ones simply couldn't get to. In a therapeutic test — evolving an enzyme to precisely cut a protein linked to a neurodegenerative disease — the AI-started version ended up over 79× more specific for its target than the best version evolved from the natural enzyme.
Method: ProteinMPNN-redesigned BoNT proteases (improved stability, retained catalytic efficiency) served as PACE (phage-assisted continuous evolution) starting points, run side-by-side against WT across four campaigns.
Result: Redesign conferred mutational robustness unlocking otherwise-inaccessible high-fitness sequence space; redesign-evolved mutations were non-functional in WT backgrounds, indicating epistatic dependence on the redesigned scaffold. Adaptation rate increased where redesign raised local fitness. Retargeting BoNT/E to cleave ataxin-2: redesign-evolved variants achieved >79× greater selected specificity vs. the best WT-evolved variant, with higher catalytic efficiency/stability and minimized native-substrate cleavage.
The problem: Gene therapy often needs to insert a working copy of a gene at a precise spot in a cell's DNA — but the standard CRISPR approach cuts both strands of the DNA double helix, which is powerful but risky: the cell's repair process can introduce unwanted extra mutations, scramble chunks of chromosome, or edit the wrong location.
What they did: This team built a new editing system ("KNIT editing") that inserts large chunks of DNA — up to 10,000+ genetic letters — without ever fully cutting both strands, just nicking one.
What they found: It worked up to 89% of the time, with far fewer accidental mutations, and they used it to build CAR-T cells (an engineered immune cell therapy for cancer) without the double-strand breaks that current methods rely on — and those engineered cells still fought tumors effectively in mice.
Method: CRISPR kilobase-scale nickase-targeting (KNIT) editing couples a Cas9 nickase with a DNA-donor-recruiting system for programmable kb-scale insertion (0.7–>10 kb) without double-strand breaks, across genomic loci and cell types.
Result: Up to 89% efficiency with markedly reduced indels, translocations, and off-target editing; supports repeated/multiloci insertion with minimal translocations. KNIT editor 2 improves efficiency via single transfection. Restored normal gene expression in pathological-mutation cells via safe-harbour or native-locus insertion; enabled non-viral, double-strand-break-free CAR-T engineering with clinically relevant efficiency and effective in vitro/in vivo antitumour activity.
The problem: Base editors are gene-editing tools that can rewrite a single genetic "letter" without cutting DNA at all — but they sometimes edit the wrong letter elsewhere in the genome (off-target effects), and fixing that has traditionally meant slow, expensive trial and error.
What they did: This team had an idea: feed the DNA sequences where off-target edits happen into AlphaFold3 (the AI that predicts 3D protein shapes) and see if it can tell you which exact points of contact between the editing protein and the DNA are responsible.
What they found: It worked — the AI's "contact probability" score was a better signal than looking at 3D structure alone, and using it let them identify and mutate the specific spots causing the mismatch. Their best redesigned editor beat existing high-precision editors, and the same approach generalized to a different type of base editor entirely.
Method: ContactSeek: an AI framework using AlphaFold3-predicted contact probability (found more sensitive than predicted 3D structure) to map genome-wide off-targets of Cas9–TadA adenine base editors, correlating contact probability with sequencing-based off-target signal to rank consensus contact regions and pinpoint specificity-determining Cas/TadA8e residues.
Result: Best variant (two Cas9 + TadA8e mutations) outperformed known high-fidelity adenine base editors per targeted amplicon sequencing, genome-wide profiling, R-loop assay, and RNA-seq; framework generalized to Cas12a-based cytosine base editors. Establishes an AF3-driven paradigm integrating structural and functional dimensions for genome-editor specificity engineering.
The problem: CRISPR-style gene-editing tools are all derived from natural proteins that evolution happened to produce — which means their properties are limited to whatever evolution stumbled onto, not necessarily what's most useful for engineering.
What they did: This team combined an AI structure-prediction model with knowledge of which parts of the protein evolution tends to conserve, to design entirely new, non-natural variants of a minimal CRISPR-like enzyme (TnpB).
What they found: Many of their AI-designed variants worked as well as or better than the natural original across bacteria, plants, and human cells, and they even solved the 3D structure of the most different variant to confirm it folds correctly and functions as designed — expanding the design space for gene-editing tools beyond what nature happened to evolve.
Method: Couples a structure-guided inverse-folding model with evolution-informed residue constraints to generate active, divergent variants of TnpB (a minimal CRISPR-Cas12-like nuclease), termed SynTnpBs.
Result: High-throughput screening of AI-generated variants yielded editors retaining or exceeding wild-type activity in bacterial, plant, and human cells. Cryo-EM structure determination of the most divergent variant revealed stabilizing contacts across RNA–DNA interface conformations, validating the design approach. Establishes a strategy for engineering non-natural RNA-guided nucleases and conformationally active nucleic-acid binders, expanding designable protein space beyond natural evolutionary sampling.
The problem: There's a trick in photonics for making light travel through a circuit in one direction only, immune to bumps and defects — useful for building robust optical chips. The catch: normally you need a big "insulating" buffer material around the narrow channel that actually carries the light, so most of the chip's area is wasted doing nothing.
What they did: This team designed a new structure where there's no dead insulating bulk at all — the whole area is made of four cleverly arranged active channels that each act as both a light-guiding lane and a wall for the lane next to it.
What they found: The result is multiple lanes of one-way light packed tightly together with zero wasted space, while keeping the "shrug off defects and sharp bends" robustness that makes this approach valuable in the first place. It's a blueprint for much denser optical chips.
Method: Introduces an insulator-free topological waveguide architecture eliminating the robustness/footprint trade-off inherent to conventional topological-insulator-domain interfaces. Combines time-reversal and inversion-symmetry breaking in gyromagnetic honeycomb photonic crystals to realize four inequivalent photonic valley half-semimetals (PVHSMs) at distinct trivial/Chern transition boundaries, arranged in parallel cyclic configuration so each domain is simultaneously a valley-selective waveguide and topological barrier for the adjacent valley.
Result: Yields densely packed, large-area one-way modes with alternating unidirectionality across domains, 100% spatial utilization, and robustness under arbitrary sharp bends/shape variation — a design strategy for ultracompact topological photonic circuits.
The problem: MXenes are a family of ultra-thin 2D materials (think: cousins of graphene) that materials scientists have been using as a flexible platform to engineer new electronic properties. One goal is combining semiconductor behavior with magnetism in the same material — useful for "spintronic" devices that use an electron's spin, not just its charge, to store or process information. Making MXenes out of lanthanide elements (which have the right magnetic properties) with the usual "etch away material" method doesn't work well, because lanthanides dissolve in the etching chemicals.
What they did: This team instead built the material up from smaller layered pieces rather than etching it down.
What they found: They successfully made a new class of lanthanide MXenes that are simultaneously semiconducting and magnetic — a combination that opens the door to new spintronic devices.
Method: Reports a general bottom-up route to lanthanide MXenes (Ln₂CT₂; Ln = Gd, Tb, Dy, Ho, Er, Lu; T = Cl, Br) using layered halides as van der Waals building blocks, circumventing top-down HF-etching limitations from Ln solubility.
Result: Multilayer Ln₂CT₂ shows composition-tunable optical absorption onsets (1.26–1.71 eV), room-temperature resistivity 0.329–36.1 Ω·cm with negative temperature coefficient, and low-T ferromagnetic hysteresis at 2 K with positive Curie–Weiss temperatures (6–59 K). DFT: d-electron states near Ef are diminished in bare Ln₂C and further depleted by surface terminals (opening band gaps), while localized 4f electrons far from Ef drive the observed ferromagnetic spin splitting — a viable 2D magnetic-semiconductor candidate for spintronics.
The problem: Perovskite solar cells are one of the most promising next-gen solar technologies — cheap, efficient, but historically not very durable. A key weak point is the ultra-thin molecular glue layer that anchors the light-absorbing material to the electrode; under heat and light stress it tends to peel off, degrading the cell.
What they did: This team redesigned that glue molecule so that its internal electronic structure pulls extra negative charge toward the anchoring point, making the bond to the electrode significantly stronger and much more resistant to peeling.
What they found: The resulting cells kept over 93% of their performance after 1,080 hours of heat+light stress testing, retained 98% after 720 extreme hot/cold cycles, and reached a certified 27.69% efficiency — near the top of what's been achieved for this technology — while working on both rigid and flexible surfaces.
Method: Designed a donor-acceptor-donor (D-A-D) resonant self-assembled monolayer (SAM) anchoring perovskite to ITO, where electronic resonance increases negative charge density at the phosphonic-acid acceptor group, strengthening the SAM–ITO bond and suppressing desorption under operational stress.
Result: Devices retained negligible decay under MPPT at 85±5°C for 1,080 h, >93% after 1,080 h under MH-lamp illumination (100 mW cm−², 4.4% UV) at 85±5°C, and >98% after 720 thermal cycles (−40 to 85°C). Certified PCE: 27.69% (0.063 cm²), 23.63% (15.64 cm² aperture), 26.64% on flexible substrates — demonstrating cross-substrate universality via resonance-induced charge delocalization improving carrier transport.
The problem: A different angle on the same perovskite-solar-cell durability problem as the Nature paper above: to manufacture these cells at scale (think: printing them like ink onto large sheets), the material has to form a smooth, uniform crystal layer as it dries in normal room air — but the perovskite material is chemically unstable and tends to degrade unevenly during that drying process, hurting performance.
What they did: This team developed a coating technique that dynamically manages moisture during the printing process, preventing that uneven early degradation.
What they found: The resulting cells hit 26.7% efficiency in lab tests (independently certified at 26.1%), scaled successfully to full-size (100 cm²) panels at ~20% efficiency, and kept over 90% of their performance after 1,500 hours of intense heat-and-power stress testing — a serious step toward these cells being manufacturable, not just a lab curiosity.
Method: Demonstrates spatiotemporally homogeneous crystallization of α-phase FAPbI₃ via a phase-locking strategy establishing a dynamically evolving, moisture-buffering intergranular network during large-area blade coating, preventing premature humidity-driven degradation and eliminating directional inhomogeneity that otherwise arises from spatially heterogeneous degradation of metastable perovskites during prolonged ambient coating.
Result: Blade-coated devices: 26.7% PCE (26.1% certified); rigid/flexible 100 cm² modules: 21.5%/19.5%. Encapsulated devices retained >90% initial PCE after 1,500 h at 85°C MPPT in ambient air — improved morphological homogeneity mitigated localized degradation and suppressed self-amplifying aging pathways.
The problem: As computer chips pack transistors ever closer together, the tiny copper wires connecting them become a bottleneck — at very small sizes, copper's electrical resistance actually goes up, wasting power and generating heat.
What they did: This team grew ultra-thin wires (down to 40 nanometers, about 2,000× thinner than a human hair) from a different material, niobium arsenide, which has an unusual quantum property: its electrical current likes to travel along the wire's surface rather than through its bulk.
What they found: Counter-intuitively, as they made the wires thinner, resistance went down rather than up (about 70% lower than the bulk material) — the opposite of how copper behaves at these scales. Combined with good heat handling and stability, this points to a real alternative to copper for future, smaller chip interconnects.
Method: Synthesized single-crystalline Weyl semimetal NbAs nanowires via thermomechanical nanomolding, diameters down to 40 nm. Resistivity decreases with decreasing diameter; 40 nm-diameter nanowires: 10.5 ± 1.9 μΩ·cm at room temperature, ~70% below bulk.
Result: Calculations attribute this to surface-dominant conduction with long carrier lifetime at finite temperature (topological surface-state transport counteracting the classical size-effect resistivity increase seen in conventional metals like Cu at comparable dimensions). Nanowires/bulk crystals also show high breakdown current density, stability, and thermal conductivity — positions NbAs nanowires as a candidate to surpass copper-interconnect scaling limitations.
The question: When does a tiny cluster of atoms melt or freeze? For bulk materials the answer is simple (a fixed temperature), but at the nanoscale, the geometry of the cluster itself starts to matter a lot.
What they did: Using a specialized electron microscope, this team watched a single, extremely small bismuth cluster — confined in a tunable nano-gap — cycle back and forth between a disordered blob, an ordered crystal wire, and a liquid droplet, entirely reversibly, live and in real time.
What they found: The deciding factor for which state the cluster takes isn't how much material is there (its volume), but its shape (aspect ratio) — the tug-of-war between surface effects and interface energy. The confinement itself also forces the crystal to grow in a specific orientation it wouldn't pick on its own, giving materials scientists a real, mechanistic handle on designing nanomaterials by engineering their confinement.
Method: In situ TEM directly controls and observes a single critical-sized bismuth nanocluster within a tunable nanoscale gap through a reversible cycle: quasi-amorphous nanodisc → crystalline nanowire → liquid nanodroplet.
Result: Aspect ratio, not volume, is the primary descriptor governing these transitions, set by the interplay of intrinsic surface anisotropy and interfacial energetics. Confinement imposes texture, forcing the nanowire into a preferred orientation absent in unconfined nanoparticles. Establishes a mechanistic foundation for geometry-driven phase/orientation selection, enabling rational nanomaterial design through engineered confinement.
The theoretical limit of data storage is using a single electron to hold one bit — today's flash memory chips use many thousands of electrons per storage cell. But shrinking a memory device down to where a single electron can meaningfully change its behavior runs into a nasty side effect: stray electrical interference ("fringe capacitance") that gets worse, not better, as you shrink the device, which has made single-electron memory hard to actually observe experimentally. This team built a device with a clever flat, side-by-side layout that specifically suppresses that interference, and successfully showed a clear, stable, non-volatile (doesn't reset when power is removed) memory effect from the presence or absence of just one electron. They also observed a distinctive new quantum effect they describe as a quantum state being "cut off" by a change in the material's electronic density of states — a genuinely new quantum memory mechanism.
Reports a two-dimensional single-electron memory device using a coplanar drain-channel-source structure that suppresses fringe capacitance effects (which otherwise amplify as devices scale down, historically preventing experimental observation of single-electron memory). Device exhibits a nonvolatile threshold voltage shift of 0.5 V upon the addition/removal of a single electron. Two quantum behaviours verified with respect to programming voltage; additionally predicts and observes a distinctive quantum memory effect described as a quantum state being cut off by "density of states scissors" — a novel mechanism relevant to the ultimate (single-electron) limit of nonvolatile information storage.
The question: Chemotherapy and radiation save children's lives, but they're also known to damage DNA as a side effect — this study measured exactly how much, using genome sequencing of relapsed childhood tumors where the exact drugs and doses were tracked.
What they found: Treatment-caused DNA damage wasn't a minor factor: post-treatment cancers had roughly triple the number of distinct "mutation signatures" and double the total mutation load compared to tumors that never saw treatment. Platinum-based chemo drugs were the worst offenders by far, and remarkably, over a third of patients treated with them showed detectable platinum-linked mutations within just one year of treatment.
Why it matters: This gives doctors genomic evidence to potentially scale back chemo doses in some cases and to watch for early warning signs of treatment-driven relapse.
Method: Mutational-signature analysis of whole-genome-sequenced relapsed childhood tumours from a multi-institutional cohort with uniformly collected therapy dose/exposure data.
Result: Chemo/radiotherapy were the only exogenous mutagens identified and often the dominant DNA-alteration source; post-therapy tumours carried ~3× the private mutational signatures and ~2× total somatic mutation burden vs. treatment-naive tumours. Platinum-based therapies drove the highest variant counts (associated signature count more than doubled); using exposure-date tracking, >1/3 of platinum-treated tumours showed detectable platinum signatures within one year, establishing a minimum emergence threshold. Provides genomic evidence for treatment de-escalation and pre-expansion tracking of resistant clones.
The problem: Crimean–Congo haemorrhagic fever is a tick-borne viral disease that can be fatal in humans, and there's currently no approved vaccine or drug for it — the WHO lists it as a priority pathogen.
What they did: Two independent teams solved the 3D structure of the virus's core replication machine (its RNA polymerase, the enzyme it uses to copy its own genetic material), which is enormous even by viral standards. Having the actual 3D shape lets researchers see exactly how experimental drugs bind and block it.
What they found: A nucleoside-analog drug related to the hepatitis C treatment sofosbuvir turned out to be a strong, specific inhibitor — structural groundwork needed to rationally design real antivirals against a disease that currently has none.
Method: Two independent cryo-EM structural studies of CCHFV-L, the ~4000-residue Nairoviridae RNA-dependent RNA polymerase (among the largest known viral polymerases). Captured elongating L–RNA complexes (one at 3.0-Å resolution), revealing an enlarged architecture with substantial insertions in the endonuclease, RdRP, and cap-binding domains extending RNA-binding paths on both sides of the active site.
Result: Define mechanisms of the baloxavir-derived candidate WXSH0208 (endonuclease inhibition) and ribose-2′-modified nucleoside analogues structurally identical to the sofosbuvir class (post-translocation chain termination), the latter showing nanomolar cellular potency and efficient, specific RdRP inhibition in minigenome/competition assays — structurally guiding both nucleoside and non-nucleoside CCHFV antiviral design.
The question: Tertiary lymphoid structures (TLSs) are makeshift immune-system hubs that sometimes form directly inside tumors, and their presence is linked to better responses to cancer immunotherapy — but why wasn't well understood.
What they found: Studying kidney cancer, this team found that tumors containing TLSs had more immune cells that specifically recognize the tumor, and a subset of those tumor-fighting cells living inside the TLS had a "stem-like" quality that makes them better long-term fighters rather than burned-out and exhausted.
The catch: In TLS-containing tumors, immune-suppressing cells at the tumor's edge were co-located with exhausted tumor-fighting cells, suggesting a defense mechanism the tumor uses to blunt the attack. The findings suggest TLSs are a reservoir of high-quality cancer-fighting T cells that future immunotherapies could specifically target and boost.
Method: Across 24 treatment-naive renal cell carcinoma (RCC) tumours, TLS-containing tumours showed heavier CD8+ exhausted T-cell infiltration with reduced terminal-exhaustion transcriptional programs vs. TLS-negative tumours. Specificity screening of 554 intratumoural T-cell clonotypes (6 RCC tumours) identified 82 TCRs reactive against tumour cells/RCC antigens.
Result: 12% of tumour-specific clonotypes were TLS-enriched, expressing elevated stem-like progenitor exhaustion programs associated with favourable anti-tumour immunity. In 60 independent RCC tumours, TLS-containing tumour margins showed immunosuppressive-phenotype macrophages colocalized with exhausted putative tumour-reactive T cells, supporting this as an immune-evasion counterbalance to T-cell pressure. TLSs are reservoirs of stem-like tumour-specific T cells exploitable by T-cell immunotherapies.
The problem: Two people with the "same" cancer can have wildly different disease courses, and part of that is thought to come from differences in their genetic background — but this is very hard to study directly in humans, where genetics, environment, and exposures are all tangled together.
What they did: This team got around that by re-running early tumor development hundreds of times in genetically diverse but tightly controlled inbred mouse strains, with everything else (sex, environment, carcinogen exposure) held constant.
What they found: Even genetic differences comparable to the variation between different human ancestry groups were enough to dramatically change how a tumor evolves — which mutations it picks up, whether it duplicates its whole genome, and how fast it grows. Genetic ancestry isn't just a cancer-risk factor; it actively steers how a cancer, once started, evolves.
Method: Replayed early tumour evolution hundreds of times across diverged inbred mouse strains with controlled sex/environment/carcinogenic exposure, generating matched histology, WGS, and transcriptomes, exploiting nested strain–litter–animal–tumour hierarchy to capture genetic variation comparable to human population-level diversity.
Result: Found epistatic interactions between genetic background and acquired somatic mutations producing population-specific disease progression: driver-mutation choice, whole-genome-duplication incidence, and subclonal selection dynamics all varied with strain background, mirroring both cancer susceptibility and growth rate. Modest genetic divergence (human-ancestry-comparable) measurably altered selection pressures during early cancer development, shaping both risk and evolutionary trajectory.
The mystery: Since 2012, catfish in a lake spanning Vermont and Quebec have been showing up with unusually high rates of skin cancer (melanoma). The obvious guess would be a pollutant in the water, but this team tested a much stranger hypothesis: that the cancer itself was contagious — actual cancer cells spreading from fish to fish, like a parasite.
What they found: Genome sequencing confirmed it: tumors from different, unrelated fish were genetically more similar to each other than to their own hosts, sharing hundreds of thousands of genetic variants that healthy fish don't have. This makes it only the fourth known naturally occurring transmissible cancer in the animal kingdom, after specific cancers seen in dogs, Tasmanian devils, and some shellfish.
Method: Whole-genome sequencing of tumour and matched non-tumour host tissue from brown bullhead catfish (Ameiurus nebulosus, elevated melanoma incidence since 2012 in a Vermont/Quebec lake).
Result: Tumour mitochondrial and nuclear genomes are more closely related to each other than to their respective hosts or unaffected fish. Hundreds of thousands of shared genetic variants across tumour samples are absent from host fish, far exceeding variant-sharing levels in conventional (non-transmissible) cancers. Establishes this melanoma as the fourth documented naturally occurring clonally transmissible cancer in animals, after canine transmissible venereal tumour, Tasmanian devil facial tumour disease, and several bivalve lineages.
The question: This is a companion story to the Nature kidney-cancer TLS paper above, from a different angle: what actually builds and maintains tertiary lymphoid structures (TLS) in tumors in the first place? Using a lung cancer model, this team found that a specific type of immune cell — mature dendritic cells — are the key organizers.
What they found: Early on, these cells need to sense inflammation signals, mature, and travel to nearby lymph nodes to recruit T cells and kick off TLS formation. But once the tumor is established, the TLS keeps going on its own, powered by dendritic cells that have set up shop directly inside the tumor and continuously "present" tumor material to keep the local immune response going — identifying dendritic cells as the linchpin that could be specifically targeted to strengthen anti-tumor TLS immunity.
Method: Spatial transcriptomics and multiplex imaging across human tumours show CCR7+ mature dendritic cells (DCs) accumulate in TLSs. In a mouse NSCLC model forming mature TLSs, early TLS development requires IFN-γ-driven cDC1 maturation, migration to tumour-draining lymph nodes, and T-cell recruitment.
Result: As tumours progress, TLS persistence becomes independent of tdLN T-cell egress, coinciding with cDC1 accumulation in intratumoural CCL19+ stromal hubs, where cDC1 MHC-I/MHC-II antigen presentation plus CD40 signalling sustain TLS, TFH pool, germinal centers, and tumour-specific IgG. Identifies local mature cDC1s as key TLS orchestrators and candidate immunotherapy targets.
As we age, our organs gradually decline, and a lot of that decline is linked to chronic, low-grade inflammation. This study identifies a specific mechanism behind it: aging impairs the ability of "housekeeping" immune cells (tissue-resident macrophages) to clean up worn-out white blood cells (senescent neutrophils), and this cleanup failure is controlled by a specific signaling receptor (EP2). When researchers dialed down that receptor's signaling in aged mice, the aged mice's mitochondria (the cell's energy factories) stayed youthful, and the mice were protected against cognitive decline, frailty, muscle loss, weight gain, heart problems, and body-wide inflammation. The same faulty signaling and buildup of worn-out neutrophils showed up in aged and diseased human tissue too, and a drug blocking the receptor restored proper cleanup — suggesting this specific process is a reversible, drug-targetable driver of aging, not just a symptom of it.
Identifies tissue-resident macrophages (TRMs) as central coordinators of age-related organ decline via impaired clearance (efferocytosis) of senescent neutrophils, regulated by immunomodulatory PGE2 receptor EP2. Reducing TRM EP2 signalling in aged mice preserved youthful mitochondrial fitness and prevented cognitive decline, frailty, sarcopenia, adiposity, cardiac impairment, and systemic inflammation. Plasma proteomics implicates the liver as a major source of age-associated immune change; reduced TRM EP2 signalling rescued neutrophil efferocytosis and prevented paracrine stress in neighbouring cells. Elevated TRM EP2 expression and senescent-neutrophil accumulation observed in aged/diseased human tissue; pharmacologic EP2 inhibition restored youthful neutrophil clearance — establishes impaired TRM efferocytosis as a reversible driver of organ aging.
The setup: DNA in a cell's nucleus isn't just a tangled string — it's physically organized into 3D neighborhoods, with some genes parked near the nuclear envelope (a repressive "quiet" zone) and others near active "speckle" hubs in the middle.
What they found: Using human brain tissue from mid-development, this team found that as brain cells (neurons) mature, hundreds of important developmental genes physically relocate from the quiet envelope zone to the active speckle zone — and this move, not just the chemical tags on the DNA itself, is what switches many of these genes from off to fully on (over 8-fold more active). It shows that where a gene physically sits inside the nucleus is itself part of how the cell decides whether to use it, not just a side effect.
Method: High-resolution maps of lamina/speckle genomic interactions in neurogenic-lineage cells from mid-gestational human cortex reveal extensive remodeling of subnuclear compartmentalization during cortical neurogenesis, relocating hundreds of neuronal genes — including H3K27me3/H3K4me3-bivalent neurodevelopmental genes — from lamina to speckles.
Result: Lamina-resident bivalent genes show exceptionally low expression; relocation to speckles resolves bivalency to H3K4me3 monovalency and increases transcription >8-fold. Proximity to nuclear periphery, not H3K27me3 presence per se, maintains the poised state; the lamina's repressive environment correlates with spatial segregation of transcriptional elongation machinery from the periphery — spatial location is necessary information for epigenomic regulation, not merely correlated with it.
The question: How does an embryo build a nervous system with thousands of precisely different neuron types, all wired correctly? This team built an extremely detailed genetic atlas of every cell in the developing fruit fly nerve cord (the insect equivalent of a spinal cord) and matched it to the adult wiring diagram.
What they found: Neurons born earlier in development (in the embryo) become more molecularly distinct from each other faster than neurons born later (as larvae). More strikingly, they found 17 genes that, across essentially all neuron types regardless of lineage, act together as a shared code that records exactly when in development each neuron was born — a kind of universal timestamp system. They also found that a specific pattern of female-only cell death and gene activity is a major driver of how the sexes' nervous systems end up different.
Method: High-resolution developmental transcriptional atlas of the Drosophila melanogaster nerve cord (38× aggregate coverage relative to reference connectome) enabling robust adult-connectome alignment.
Result: Three principles: (1) neurogenesis timing shapes diversification rate — embryonic-born neurons diverge faster than larval-born, mirrored in the adult connectome; (2) 17 transcription factors shared across all lineages constitute a global molecular birth-order code; (3) sex-specific transcriptional profiles mapped to the connectome identify female-specific apoptosis and transcriptional divergence as key global drivers of sex specification. Opens mechanistic dissection of molecular-identity axes underlying neural circuit development/evolution.
The setup: Neurons aren't simple on/off switches — their branching dendrites (the tree-like input branches) can do local computation of their own before signals reach the cell body. This team used voltage imaging (a technique to directly watch electrical activity) across the branches of hippocampal CA3 neurons — cells central to spatial navigation and memory — while mice navigated a virtual-reality maze.
What they found: Individual dendritic branches act like semi-independent computing units, sometimes syncing up with the main cell body's output and sometimes operating on their own, depending on what the animal is doing. These branches encode information about space, reward, and context, and even during "sharp-wave ripples" (bursts linked to memory replay), branches tuned to the same location kept their coordination. In effect, a single neuron's dendrites can simultaneously represent past, present, and possible future information.
Method: Voltage imaging across dendrites and somata of CA3 pyramidal neurons during virtual-reality-guided navigation in mice reveals the dendritic arbor comprises multiple independent computational units that dynamically couple to or dissociate from somatic activity depending on behavioural state.
Result: Dendritic activity shapes subcellular representations of space, reward, and context via conditional somatic coupling; spatially cotuned dendrites retain coordination during sharp-wave ripples. Demonstrates that past, present, and future representations coexist within a single CA3 pyramidal neuron's dendritic arbor, collectively shaping behaviourally relevant neuronal coding — extends dendritic computation's role beyond cortex into hippocampal spatial processing.
The question: How snakes evolved their body plan and lifestyle from lizard-like ancestors has been hard to answer, mostly because early snake fossils are rare and hard to interpret.
What they did: This paper describes a beautifully preserved new fossil snake from Cretaceous-era Brazil (~100+ million years old), one of the earliest branches on the snake family tree. Using high-resolution CT scanning, the researchers reconstructed its brain, inner ear, and nerve anatomy in unprecedented detail.
What they found: Its brain shape was distinct from both other early snakes and living snakes today, implying its senses worked differently. Combined with skeletal clues, the evidence suggests this particular ancient snake was a burrowing animal, while a different early snake species (Dinilysia) was not — meaning early snakes had already split into quite different lifestyles very early on, contradicting the idea that all early snakes shared one simple ancestral ecology.
Method: Describes Tametara mirim gen. et sp. nov., an exceptionally preserved Cretaceous stem snake from Brazil among the earliest-diverging lineages. High-resolution micro-CT enabled reconstruction of cranial nerves, inner ear, and brain endocast anatomy — the most integrated stem-snake neuroanatomical reconstruction to date.
Result: Quantitative/qualitative endocast analyses show brain morphology distinct from both other stem snakes and extant snakes, indicating substantial early neuroanatomical (and likely sensory) disparity. Independent telencephalon-shape and bone-microstructure evidence converge on a fossorial lifestyle for Tametara vs. non-fossorial for stem snake Dinilysia — indicating major ecological transitions occurred early in snake evolution and that known stem taxa do not represent a single ancestral condition for crown snakes.
The gap: Big genome databases have transformed marine microbiology, but they've historically missed some of the ocean's most abundant microbes because certain bacterial groups are hard to sequence well with standard methods.
What they did: This team built a new genome database specifically from Great Barrier Reef seawater, using a sequencing approach (Nanopore, alongside standard Illumina) better suited to capturing the tricky, underrepresented groups that standard short-read methods miss.
What they found: The result is a resource of over 5,000 bacterial genomes plus microscopic algae and more than 800,000 viral genomes, including a newly described group of ocean viruses. As a practical demonstration, they showed this database can identify specific microbial "indicator species" that predict whether reef-management actions, like protected zones, are actually working.
Method: Presents the Great Barrier Reef Microbial Genomes Database (GBR-MGD): 5,283 prokaryotic genomes from GBR seawater via combined Nanopore/Illumina sequencing, including high-quality genomes of dominant-but-underrepresented lineages (e.g., Pelagibacter, Prochlorococcus) that standard short-read assemblies miss due to strain heterogeneity and low-GC-percentage sequencing bias.
Result: Also includes 20 chromosome-level picoeukaryote genomes and 808,585 viral genomes, including a newly described marine Crassvirales clade. Demonstrates utility for identifying indicator taxa that reliably predict effects of reef-management practices (e.g., marine protected zone establishment).
The setup: Transform faults are the type of fault where two of Earth's tectonic plates slide past each other sideways (like California's San Andreas, but this study is on an underwater one). They've traditionally been thought of as "dry," pure-shear boundaries without much fluid involved.
What they found: This team discovered a faint, continuous seismic hum ("tremor") at an oceanic transform fault that rhythmically strengthens and weakens in sync with the tides — a clear sign that fluids and gases trapped underground are involved.
The mechanism: They propose the fault acts like a valve: it seals up, trapping volcanic fluids and gases which builds pressure and tidal sensitivity, until an earthquake cracks it open, draining the fluids, silencing the tremor, and triggering small quakes — before it heals back up and the cycle restarts. This overturns the "dry fault" picture.
Finding: Discovered tidally modulated tremor at the Gofar transform fault (East Pacific Rise). Tremor amplitude correlates with semidiurnal tides during sparse-seismicity, low Vp/Vs periods; correlation weakens following earthquake swarms accompanied by high Vp/Vs.
Model: Proposes a valve-like sealing-drainage dynamic: sealing traps magmatic volatiles and boosts tidal sensitivity, sustaining tremor until rupture opens high-porosity/permeability pathways, silencing tremor, triggering microseismicity, and resetting via hydrothermal resealing. Indicates transform faults are permeable and tide-critical, with energy release oscillating between tremor and rupture — revises the conservative shear-dominated-boundary model.
One of the defining innovations in mammal evolution is the "tribosphenic molar" — a tooth shape that can both shear (like a knife, for slicing meat) and crush (like a mortar, for grinding) in one structure, generally considered a key reason mammals diversified successfully. This study asks: can a tooth actually be great at both jobs at once, or is there an inherent trade-off? Looking across predatory mammal lineages, they found shearing performance is a strict, narrow target — there's basically one optimal way to build a slicing tooth — while crushing performance can be achieved many different ways. The upshot: fewer than 1% of predator species ever evolved teeth that were simultaneously optimized for both functions. So rather than mammals commonly having "do-it-all" teeth, most lineages specialized in one direction or the other, which shaped how mammal diversity actually unfolded over time.
Examines whether trade-offs between the dual shearing/crushing functions of the tribosphenic carnassial molar influence macroevolutionary outcomes across predatory mammal lineages. Predatory mammals evolved dichotomized tribosphenic performance: slicing function is constrained to a narrow set of optimal phenotypes, while crushing exhibits redundant/degenerate solutions. <1% of predator lineages evolved teeth optimized for both shearing and crushing simultaneously. The fundamental functional trade-off promoted divergent macroevolutionary specialization rather than sustained functional duality — illustrates how a key evolutionary innovation can drive early success while constraining subsequent diversification.
The gap: Despite over 6,000 confirmed exoplanets, nobody has ever confirmed a moon orbiting one of them — every candidate exomoon so far has been controversial and unconfirmed.
What they did: This team looked instead at a brown dwarf (an object between a planet and a star in mass) that itself orbits a star, and used the same "wobble" technique that found the first exoplanet to look for something orbiting the brown dwarf.
What they found: A periodic signal consistent with an object roughly 0.9 times the mass of Jupiter, orbiting the brown dwarf every ~170 days. Because the host is a brown dwarf and not a planet, it's technically an "exosatellite" rather than a confirmed exomoon under current (loose) definitions — but it's the strongest evidence yet that this detection method can find these objects, and points toward the real thing eventually being confirmed.
Method: Applied radial-velocity analysis to VLT/CRIRES+ spectra of directly imaged brown dwarf companion CD-35 2722 B, detecting a periodic signal consistent with an orbiting satellite — the first application of this technique to a substellar (brown dwarf) companion rather than a star.
Result: Best-fit model: minimum mass ~0.9 MJup, orbital period ~170 days. Terminological note: "exomoon" lacks a formal definition extending to satellites of substellar (non-planetary) primaries, hence "exosatellite"; advancing instrument sensitivity should extend this method to lower-mass targets, the path toward an uncontroversial exomoon detection.
Molecules built around heavy, radioactive atomic nuclei are extremely useful for fundamental physics — they're unusually sensitive to subtle effects that could reveal new physics beyond our current understanding of particles and forces. The catch is that radioactive nuclei are scarce and hard to work with, and you generally need molecules cooled to near-zero temperatures for precision measurements to work at all. This team built a tabletop setup that makes and laser-cools three different molecules containing radioactive radium, at high enough precision to support serious physics experiments, all in a lab-bench-sized apparatus rather than requiring a massive facility. This establishes a practical, reusable toolkit that other labs can now apply to a wide range of exotic radioactive species for precision physics and quantum-sensing experiments.
Demonstrates gas-phase synthesis, cryogenic buffer-gas cooling, and high-resolution laser spectroscopy of radium monohydroxide, monodeuteroxide, and monofluoride (²²⁶RaOH, ²²⁶RaOD, ²²⁶RaF) in a tabletop apparatus, combining trace radioactive target production, optically driven chemistry in cryogenic buffer gas, and low-background spectroscopic detection. Molecules are cooled in the lab frame to starting conditions matching current molecular precision-measurement and quantum-information experiments. Establishes a generalizable capability for molecular quantum sensing of exotic (radioactive) nuclei, applicable across a wide range of species despite limited nuclide availability.
The gap: Asia holds about 60% of the world's population, but most long-term studies of life expectancy have focused on Western, high-income countries. This study fills that gap: across 34 Asian countries/territories from 1990–2023, life expectancy rose everywhere, with the biggest gains in South Asia and the smallest in already-wealthy Asia-Pacific nations.
What drove it: Fewer heart-disease deaths mattered most in East/Central Asia, while fewer deaths from diarrheal disease and tuberculosis mattered most in South/Southeast Asia. Then COVID hit: life expectancy fell across several regions starting in 2019–2023, with almost two years lost in the pandemic's first year in some places.
Why it matters: A one-size-fits-all health policy won't work — the right interventions differ sharply by region.
Method: Comprehensive life-expectancy, cause-specific mortality, and risk-factor analysis across 34 Asian countries/territories (1990–2023) using Global Burden of Disease Study 2023 data.
Result: Life expectancy rose in all units studied; largest annual gains in South Asia, smallest in high-income Asia-Pacific. Cardiovascular mortality decline was the primary driver in Central/East Asia and high-income Asia-Pacific; diarrhoeal disease and TB declines dominated in South/Southeast Asia. 2019–2023 saw declines in several regions, largely COVID-19-driven, with ~2-year loss in pandemic year one. Cause/risk-factor contributions varied by region, supporting proportional-universalism-based, regionally differentiated policy over uniform national approaches.
The gap: Borderline personality disorder (BPD) affects roughly 1-2% of people and causes intense emotional instability, unstable relationships, and impulsivity, but until now, genetics research on it has lagged far behind other psychiatric conditions like depression or schizophrenia.
What's new: The largest genetic study of BPD to date found 11 specific locations in the human genome linked to the condition — the first solid genetic foothold researchers have had, usually the first step toward treatments that target the actual cause rather than just managing symptoms.
Finding: Largest genome-wide association study (GWAS) of borderline personality disorder to date, identifying 11 genome-wide significant loci — the first robust genetic signal for BPD, historically understudied relative to other psychiatric GWAS efforts (depression, schizophrenia) despite comparable prevalence and heritability estimates.
Most genetic variants linked to diseases and traits by large genome studies (GWAS) are thought to work indirectly — by tweaking how much a nearby gene gets expressed, rather than changing a protein directly. To actually connect a genetic variant to a disease mechanism, you need to know which variants affect which genes' expression, in which tissues — that's what this study maps, at a huge scale (over 14,000 RNA samples across six tissue/cell types, including blood and lung). They then used this new expression map to interpret 164 different GWAS traits from the UK Biobank, and successfully explained the likely mechanism behind over 10,000 disease-linked genetic signals. Notably, most of what they found were "secondary" effects that simpler, smaller studies would have missed entirely — suggesting there's still a lot more to find with bigger studies like this one.
Characterized 14,324 RNA-sequencing samples from the Trans-Omics for Precision Medicine (TOPMed) program, performing expression and splicing quantitative trait locus (e/sQTL) analyses across six tissues/cell types including whole blood (n=6,454) and lung (n=1,291). Detected tens of thousands of secondary cis-e/sQTLs, demonstrating secondary cis-e/sQTL discovery remains unsaturated even at this scale. Fine-mapped UK Biobank GWAS signals from 164 traits, identifying e/sQTL colocalizations for 10,611 GWAS signals, including 7,096 colocalizing specifically with secondary e/sQTLs — indicates larger e/sQTL analyses will continue to uncover additional secondary signals, with direct benefit to GWAS mechanistic interpretation.
The claim: Chinese AI lab Moonshot AI released a new model, "Kimi K3," which the company claims can match or beat the best current US models on various benchmarks.
The caveat: Scientists quoted in this piece are cautiously weighing in on whether that claim holds up under scrutiny — and flag that even if the performance claims are accurate, the model's very large size could make it impractical for many organizations to actually run, limiting its real-world impact despite the benchmark numbers.
Finding: News analysis of Moonshot AI's "Kimi K3" release, evaluating claims of parity with or superiority to leading US frontier models on benchmark performance. Key caveat raised by researchers: large parameter count may limit deployment/inference-cost feasibility relative to benchmark gains, constraining practical adoption independent of raw capability claims.
The argument: While Silicon Valley remains the center of gravity for building AI systems, this piece argues the UK is quietly becoming the center of gravity for AI safety work specifically — the research and policy effort focused on making sure powerful AI systems behave safely and don't cause harm.
Why it matters: The UK's growing ecosystem of AI-safety institutes, researchers, and policy initiatives is giving it outsized influence over how AI safety norms and regulations develop globally, separate from where the AI models themselves are actually built.
Finding: Analysis piece on the UK's growing influence in AI-safety research/policy ecosystem, positioned as increasingly distinct from and influential relative to Silicon Valley's AI-development center of gravity — institutional and policy dynamics of the global AI-safety governance landscape.
The story: Two pieces covering the same underlying story from different angles. US lawmakers from both parties are pushing federal science agencies to restrict research collaboration with Chinese institutions, worried about research theft and exploitation of the open scientific system — though Republicans and Democrats don't agree on exactly how to fix it.
The escalation: Separately, the National Science Foundation has gone further than a "manage the risk" approach and moved to an outright ban on most collaboration with Chinese research institutions — a significant escalation in US science policy with real consequences for international collaboration on everything from basic physics to public health research.
Nature: Bipartisan Congressional pressure on federal science agencies to restrict China research collaboration, with Republicans and Democrats diverging on policy mechanism despite shared concern over research theft/exploitation.
Science: NSF has abandoned a risk-mitigation framework in favor of an outright prohibition on collaboration with most Chinese research institutions — a substantive policy escalation from managed-risk to blanket restriction, with implications for cross-border scientific collaboration norms.
The argument: The AI industry's pitch has always been "faster, better, cheaper" — and this essay argues that scientific publishing is now the counter-example. As AI-generated text and even AI-assisted research papers enter the pipeline, publishers now have to add rigorous human fact-checking specifically to catch AI-introduced errors, which is creating bottlenecks rather than removing them.
The parallel: The piece draws a historical parallel to early-20th-century "scientific management" (Taylorism), where efficiency drives ended up demanding more surveillance and more human effort while the benefits flowed mostly to those at the top — arguing AI in publishing risks repeating that pattern instead of delivering the promised abundance.
Argument: Opinion piece arguing scientific publishing is experiencing a Taylorist dynamic from AI integration: rigorous human verification of AI-generated/assisted manuscripts, necessary to preserve record integrity, is creating review bottlenecks that increase cost and slow throughput — contradicting the "faster, better, cheaper" framing popularized in tech/aerospace management discourse (attributed to NASA administrator Daniel Goldin).
Context: Draws historical parallel to Frederick Winslow Taylor's Principles of Scientific Management, where productivity gains from surveillance-driven efficiency concentrated benefits upward while increasing worker burden.