A recurring, evidence-scored roundup of the newest science on the Frontier index.
34 new breakthroughs went live on Frontier over the past two weeks, and biology led them with 10.
Every entry on the index is a peer-reviewed result traced to its primary source and scored from 0 to 100 on three axes: Evidence, Impact, and Novelty. This dispatch covers the results added between 7 and 14 September 2026, a fortnight weighted heavily toward the life and earth sciences but reaching into condensed matter, materials chemistry, planetary science, and machine learning. Scores this period ran from the high 30s to a top of 80, with the leading edge (six results at 70 or above) sitting in cell biology and organismal ecology rather than in the fields that usually own the headlines.
The single highest-scored result is MemBrain v2 (Frontier Score 80), a deep-learning framework for reading membranes in cryo-electron tomography. It is a useful reminder that the highest-evidence science of a given fortnight is often infrastructure: a tool many later discoveries will run on, not one dramatic claim. Just behind it sit a living-cell chromatin study at 78 and a cluster of three results at 71 spanning condensed-matter physics, human-evolution biomechanics, and the dark proteome. For the longer view, our running ranking of the biggest scientific breakthroughs of 2026 collects the year's highest-scored results across every field.
By field, biology led with 10 entries, followed by climate and environment and materials (6 each), neuroscience and AI (4 each), a catch-all category (2), and a single result each in space and physics. What follows is the top of the table in depth, then a full accounting of everything else that cleared the bar.
1. The top of the table, in depth
The dozen results below are the highest-scored additions of the period, in score order. Each carries its field tag, its Frontier Score, and a short note on the kind of evidence behind it, because a score means little without knowing whether it rests on a prospective clinical cohort or a single-system demonstration. The pattern this fortnight is clear: the strongest evidence clustered in mechanistic cell biology and in tools and methods that many labs will reuse.
MemBrain v2: an end-to-end tool for the analysis of membranes in cryo-electron tomography
Biology | Frontier Score 80 | Method paper
MemBrain v2 folds membrane analysis in cryo-electron tomography into one pipeline: segmentation, particle localization, and quantitative statistics. Its three parts divide the labor. MemBrain-seg segments membranes across variable tomographic conditions, MemBrain-pick localizes membrane-bound particles data-efficiently by pairing geometric constraints with deep learning, and MemBrain-stats computes the spatial metrics that describe how particles organize inside a membrane. Membrane analysis has been a long-standing bottleneck in the technique, held back by low signal-to-noise, missing-wedge artifacts, and tools that demanded heavy manual annotation and generalized poorly across datasets. Because this is a method paper, its worth rests on independent adoption and on how the framework holds up on membrane systems it was never trained on. The work is published in Nature Methods - the paper. It is the kind of result that quietly raises the ceiling for an entire imaging community. Read the full breakdown
Cohesin prevents local mixing of condensed euchromatic domains in living human cells
Biology | Frontier Score 78 | Living-cell mechanism
Using single-nucleosome tracking, super-resolution 3D structured-illumination microscopy, and euchromatin-specific labeling of histone H3.3, this study watched how the cohesin complex organizes active chromatin inside living human cells. Euchromatin turns out to form condensed domains held apart by cohesin-mediated loops, with the transcription machinery sitting near the domain surfaces and borders. Remove cohesin and the nucleosomes become more fluid without the domains losing overall compaction, so neighboring domains mix locally and transcriptional insulation breaks down. That refines the textbook picture in which active chromatin is largely open, and assigns cohesin an unexpected physical role in keeping domains distinct. Because the work is done in living cells and links a physical property (domain integrity) to a functional one (insulation), its evidence is unusually direct for chromatin biology. What to watch is how these imaging findings connect to gene-expression outcomes across cell types, and whether the physical role generalizes beyond the labeled euchromatin studied here. Read the full breakdown
Discovery of an intrinsic non-Hermitian phase transition in a bulk condensed-matter system
Physics | Frontier Score 71 | Single system, theory-backed
Non-Hermitian phase transitions, where the dynamics rather than the steady state change qualitatively at a critical point, had mostly been the province of engineered, driven systems. Here researchers find one intrinsically in an ordinary bulk solid. Optically exciting charge carriers in ferromagnetic europium monoxide, they see the relaxation dynamics shift with temperature, showing up in time-resolved reflection data as a change from biexponential real decay to single-exponential complex decay, with an accompanying theory that reproduces the behavior. The implication is that such transitions may arise generically in condensed matter, in interplay with a material's own magnetic order, rather than only in bespoke setups. The caveat is scope: this is one material probed through its relaxation dynamics, so the "generic" claim leans on the theory more than on many measured systems. The signature now needs to be found in other bulk materials, and its physical consequences beyond the decay dynamics remain open. Read the full breakdown
Heel-strike mechanics reveal evolutionary trade-offs in hominin bipedalism
Other | Frontier Score 71 | Comparative biomechanics
Using 3D marker-based kinematics and ground-reaction-force data from humans and chimpanzees walking on two legs and on four, this work measured exactly how the two species hit the ground. Chimpanzees used 2.4 to 8.6 times greater ranges of foot-strike angle than humans and often did not heel-strike at all, and in both species heel-striking came with high impact-peak forces and loading rates. Humans, in turn, burned 26 to 41% more metabolic energy when they contacted the ground with the distal foot before the heel. That trade-off (heel-striking lowers the energy cost of walking but raises potentially damaging impact loads) reframes a feature usually treated as simply efficient. It suggests early hominins with primitive lower-limb anatomy faced a real choice between hard impacts and uneconomical walking that likely constrained their terrestrial mobility. The evidence is living-species mechanics, so its bearing on extinct hominins is an inference awaiting tests against the fossil record. Read the full breakdown
TM184C is a GPCR-like regulator of intercellular exchange and autophagy
Biology | Frontier Score 71 | Structure-mined, cell-validated
Most protein annotation rides on sequence similarity, so proteins whose function lies beyond detectable homology stay dark. By mining millions of AlphaFold2 models, this team pulled two human families of "superdark" seven-transmembrane proteins (TM184 and PRRT) that are structurally homologous to G-protein-coupled receptors and perform hallmark GPCR activities such as beta-arrestin recruitment. Focusing on TM184C, the most broadly expressed member, they found it rides dynamic vesicles along microtubules, promotes tunnelling-nanotube-like bridges between cells that share organelles, and restrains autophagy by limiting LC3B lipidation. The role is conserved enough that the human protein rescues autophagic-body homeostasis in yeast lacking its homologue. The significance here is as much methodological as biological: structure-based mining can reach into the dark proteome that sequence search misses. The functional work is done in cell and cross-species systems, so a role in human physiology or disease is the next question, not a settled fact. Read the full breakdown
Scale-dependent effects of species richness and asynchrony regulate consumer-mediated nutrient dynamics
Climate | Frontier Score 70 | Observational, 25-year
Drawing on 146 time series spanning about 25 years (1999 to 2023) from six long-term monitoring programs across coral reefs, mangrove creeks, seagrass beds, and kelp forests, and representing roughly 1.5 million individual fishes, this study estimated how stable the nitrogen supplied by fish communities is over time. Across ecosystems, species richness strongly predicted the stability of these consumer-driven nutrient flows, matching diversity-stability theory. Within any single ecosystem, though, richness was not the main lever: species asynchrony (whether species' biomass swings are out of step) was the strongest and most consistent predictor, with synchrony destabilizing the system. That distinction tells conservation what to protect at which scale. The findings are associations from observational monitoring, not manipulations, so they identify predictors rather than prove cause. The scale dependence is the heart of the result, and the open question is whether asynchrony is a lever management can actually pull. Read the full breakdown
Formin-1 maintains cochlear microtubule architecture required for hearing in humans and mice
Biology | Frontier Score 69 | Single family plus mouse model
In an extended Palestinian kindred, this study traced autosomal-recessive congenital hearing loss to a homozygous FMN1 variant that triggers aberrant splicing, nonsense-mediated decay, and loss of detectable formin-1 protein. That supplies the first human phenotype for a gene, formin-1, that was the first formin ever identified yet had no described human disorder. The hearing loss is bilateral, moderate, and stable, alongside light hair and no other anomalies. A Fmn1 knockout mouse reproduced the deafness and revealed disorganized cochlear supporting cells, loss of the tightly bundled microtubule architecture, reduced auditory-nerve responses, and fewer nerve fibers, placing FMN1 among the more than 200 genes essential for mammalian hearing. The human evidence rests on a single family, with mechanism strongest in the mouse, so the causal chain from gene loss to cochlear disruption is best established in the animal. What to watch is additional families carrying FMN1 variants, and confirmation that the supporting-cell mechanism operates in the human cochlea. Read the full breakdown
A clinically-oriented foundation model for intraoperative pathology
AI | Frontier Score 68 | Prospective clinical
CRISP is a foundation model for intraoperative (frozen-section) pathology, built on over 100,000 frozen sections from ten medical centers and then tested on more than 15,000 slides across nearly 100 diagnostic tasks. It generalized across institutions, tumor types, and anatomical sites, including sites and rare cancers it had not seen. The number that matters most: in a prospective cohort of over 3,000 patients, it directly informed surgical decisions in 92.6% of cases, while human-AI collaboration cut diagnostic workload by 35%, avoided 105 ancillary tests, and detected micrometastases with 87.5% accuracy. Computational pathology has advanced for years without the large-scale prospective validation needed to enter routine surgical workflows, and this moves past retrospective benchmarks toward the point of care. The gains come from a collaboration setting, so the model assists pathologists rather than replacing them, and external validation outside the development network is the next test. It extends the clinical picture in our latest medical breakthroughs of 2026 roundup. Read the full breakdown
Multicellular rosette formation guides epithelial tissue assembly in pancreatic organoids
Biology | Frontier Score 68 | Organoid model
In branched pancreatic ductal adenocarcinoma organoids, this work shows that the shift from a disordered mesenchymal state to an ordered columnar epithelium runs through the formation of multicellular rosettes. Fluctuating acto-myosin contractions on the emerging apical side set up a tug-of-war that spaces the rosettes regularly, with the distance between neighbors depending on branch diameter, a relationship the authors reproduce with a minimal model of apical constriction. A lumen then opens as an inner cell mass undergoes apoptosis, leaving an epithelial layer lining the cavity. Mesenchymal-to-epithelial transitions are essential to building functional tissue, yet how a disordered cell mass self-organizes into a lumen-bearing epithelium has been unclear. The argument here is that mechanics, not biochemical signaling alone, guides that assembly. The system is an organoid built from cancer cells and the spacing conclusion rests on a minimal model, so the open question is whether the same mechanical rules operate during morphogenesis in a living pancreas. Read the full breakdown
Cancer treatment alters mutant selection in normal esophagus
Biology | Frontier Score 68 | Observational, 70 patients
Aging epithelia are colonized by competing somatic mutant clones. This study asked whether cancer treatment itself changes which clones dominate normal tissue, by sequencing healthy esophageal epithelium taken from 70 patients after treatment for esophageal cancer. The selection proved treatment-specific: patients who had chemoradiation (CROSS) showed expanded TP53 and PPM1D clones, while a chemotherapy group (FLOT) showed increased selection for RAC1, NFE2L2, and MTOR mutations, consistent with those mutants conferring 5-fluorouracil resilience in normal epithelium. In other words, different regimens favor different mutant genes in the surrounding healthy tissue, not just in the tumor. Because it compares groups by the treatment they received, the work reports associations rather than proof that a regimen caused specific clones to expand, and the clinical consequence of carrying more of these clones is not established here. Larger, longitudinal cohorts are the way to confirm the patterns and trace what they mean for patients. Read the full breakdown
Evolution of flowering phenology over 44 years of climate change
Climate | Frontier Score 68 | Resurrection experiment
Direct evidence that wild populations are evolving to track climate change is scarce, and most inferences lean on space-for-time substitutions whose assumptions are hard to check. Combining those comparisons with environmental manipulation and a resurrection experiment (reviving stored seed to compare past and present genotypes), this study followed the annual plant Lactuca serriola across 44 years of climate change in Israel and the Golan Heights. Contemporary plants flowered over 7 days earlier, a putative adaptation to hotter, drier conditions, and flowered earlier under experimental drought too. The twist: plants from wetter populations flowered earlier than those from drier ones, so the spatial gradient predicted the opposite direction to the change actually seen over time. That mismatch is the durable lesson, because it warns that space-for-time studies can mispredict future responses. This is one annual species in one region, so it shows adaptive evolution can happen under climate change, not how general or how sufficient it is for a population to persist. Read the full breakdown
2. Also new this period
Below the top of the table, the period's remaining 23 entries fan out across neuroscience, materials chemistry, planetary science, and clinical AI. Two of them are connectomes: the complete Drosophila male central nervous system (166,700 neurons, Frontier Score 65), now aligned against the earlier female brain for the first synapse-resolution comparison of the sexes, and the foundational FlyWire adult-brain connectome (score 38), the reference map every later fly connectome is measured against. The male CNS work appears in Cell - the paper. The clinical-AI thread continues too, with an explainable decision-support tool for non-small-cell lung cancer and a method that makes a self-driving car's reasoning legible to its driver. The rest range from Denisovan remains identified by ancient proteins to a benchmark finding that only four of 60 language models reason like humans on policy questions.
- Compound and heterogeneous relationships between climate extremes and global net migration - Climate, Frontier Score 68
- Facial palsy reveals the sensorimotor contribution to facial-emotion recognition - Neuroscience, Frontier Score 68
- Phenomic profiles and disease patterns of 1.9 million participants from Our Future Health - Biology, Frontier Score 67
- Seasonally asymmetric warming reshapes grassland stability across timescales - Climate, Frontier Score 67
- Deconstruction of lysergic acid diethylamide - Biology, Frontier Score 67
- Ancient proteins identify various Denisovan remains from Southwest China - Other, Frontier Score 67
- Planetary shielding and surface precipitation of solar energetic particles during BepiColombo's Mercury flyby - Space, Frontier Score 66
- Divergence between physical and emotional heat resilience in Chinese cities - Climate, Frontier Score 66
- Plausible nonsense and deliberative reasoning: benchmarking LLMs against human judgment - AI, Frontier Score 66
- Identifiability of Bayesian models of perception - Neuroscience, Frontier Score 66
- Explainable deep learning improves human mental models of self-driving cars - AI, Frontier Score 65
- Imaging the Meissner effect in lanthanum hydride using diamond quantum sensors - Materials, Frontier Score 65
- Clinical usability of an explainable AI decision support tool in NSCLC - AI, Frontier Score 65
- Sexual dimorphism in the complete Drosophila male central nervous system connectome - Neuroscience, Frontier Score 65
The lower half of the period's additions leans toward materials chemistry and single-system method demonstrations, the kind of work whose evidence is real but narrow at first: a catalyst measured in one electrolyte, a polymer tested in mice and minipigs, a chiroptical material characterized before any device exists. A lower Frontier Score does not mean a result is weak or unimportant. It usually means the evidence is early-stage, the system is singular, or the claim is a method awaiting independent replication, exactly the distinctions the score is built to make legible.
- Increasing risk of vapor pressure deficit exceeding critical thresholds for tree growth - Climate, Frontier Score 64
- Disentangling the effect of space and electric field on photoelectrochemical water splitting - Materials, Frontier Score 61
- Catalytic asymmetric spiroannulation to access polycyclic spiro enones - Materials, Frontier Score 59
- Operando electrochemiluminescence imaging of intermediate evolution in electrocatalytic nitrate reduction - Materials, Frontier Score 54
- Localized glutamine leakage drives the spatial structure of root microbial colonization - Biology, Frontier Score 48
- A skin-permeable polymer for non-invasive transdermal insulin delivery - Biology, Frontier Score 46
- Hierarchical chiral supramolecular assemblies with strong and invertible chiroptical properties - Materials, Frontier Score 43
- Maximized atom utilization in a high-entropy metallene for nitrate electroreduction to ammonia - Materials, Frontier Score 41
- Neuronal wiring diagram of an adult brain - Neuroscience, Frontier Score 38
3. How these scores are built
Every entry on Frontier carries a single number from 0 to 100, and that number is not an editor's verdict. It is recomputed from a fixed methodology across three axes: Evidence (how strong and direct the support is, from a prospective clinical cohort down to a single-system demonstration), Impact (how much the result changes what its field can do), and Novelty (how far it departs from what was already known). The scores in this dispatch are the live values at the time of writing, and they can move as a result is replicated, extended, or superseded. The underlying records are traced to their primary sources and built on open scientific metadata (OpenAlex, released under CC0, alongside other open indices), so anyone can follow a score back to the paper it describes. The full, sortable index of every scored breakthrough lives on the Frontier homepage.
Correction, 27 September 2026: this dispatch originally counted 37 new entries. Three were not primary research and have been removed from the index: a news piece ("Electric charge makes drops of water more destructive"), a research briefing ("Phone-call tutorials improve learning cost-effectively during education emergencies") and a perspective essay ("Why did some viruses evolve to be giants while others did not?"), which also appeared in the top-10 chart. Six field labels have also been corrected: two climate studies had been filed under space, three entries filed under climate or the catch-all category belong in materials or biology, and the hominin walking study moved from climate to the catch-all category. The counts above reflect the corrected index. The original version also said biology and climate made up more than half of the new entries; at 15 of 37 they did not. On 28 September three more labels were corrected when Frontier recalibrated how it assigns fields: the non-Hermitian phase transition moved from quantum to physics, the Our Future Health cohort from neuroscience to biology, and the language-model benchmark from the catch-all category to AI.