Summary of Nature publications, focusing on contents relevant to AI and life sciences
Humans, Deep Learning, Diploidy, Genome, Human, Genomics, Haplotypes, Nanopore Sequencing, Nanopores, Sequence Analysis, DNA, Telomere
Producing complete, telomere-to-telomere genome assemblies currently requires combining multiple expensive sequencing technologies. Researchers developed HERRO, a deep learning framework that corrects errors in Oxford Nanopore long reads (Simplex) while preserving haplotype-specific differences—achieving up to 100-fold accuracy improvement for diploid human genomes. The key challenge HERRO addresses is distinguishing genuine sequence variants between haplotypes or repetitive genomic regions from sequencing errors, avoiding the homogenization that naive correction would cause. When paired with the Verkko 2 assembler, HERRO-corrected reads yielded assemblies with up to 32 chromosomes resolved end-to-end, including both sex chromosomes, with NGA50 values consistently above 100 Mb across multiple human genomes. The approach works with two generations of Nanopore chemistry and transfers to non-human species. By enabling high-quality assemblies from a single sequencing platform, HERRO could substantially reduce the cost and DNA input requirements for reference-grade genome projects.
Humans, Deep Learning, History, 20th Century, History, 21st Century, Human Migration, Recurrent Neural Networks, Time Factors
Tracking international migration consistently across time and countries has long been hampered by sparse, incompatible data. This study tackles that problem by building a new dataset of annual migration flows between 230 countries and regions from 1990 to the present. The researchers integrated official statistics, census-derived stock estimates, net migration figures, and prior flow reconstructions into a unified framework powered by an ensemble of deep recurrent neural networks. The models incorporate geographic, economic, cultural, and political variables to capture both long-run migration trends and rapid shifts triggered by conflicts or crises, while also quantifying uncertainty. The resulting estimates outperform existing five-year flow datasets on held-out data and offer finer temporal resolution, uncovering migration dynamics that coarser data had obscured. The authors release all data, code, and trained models, providing a reproducible resource intended to support demographic research and policy in a world where population movements are increasingly volatile.
Animals, Humans, Algorithms, Nanopore Sequencing, Sequence Analysis, DNA, Telomere
Telomere-to-telomere (T2T) genome assembly—where every chromosome is fully resolved from end to end—has previously required Oxford Nanopore ultra-long reads, which are expensive and hard to generate from samples lacking established cell lines. Researchers developed hifiasm (ONT), a new assembly algorithm that achieves near-T2T results using standard ONT simplex reads instead. Benchmarked against existing tools on the same datasets, it cuts computational requirements by roughly an order of magnitude while resolving more complete chromosomes. By removing the dependency on ultra-long sequencing, the approach makes high-quality T2T assembly practical for a much wider range of biological samples.
Animals, Ventral Tegmental Area, Prefrontal Cortex, Reward, Optogenetics, Cognitive Flexibility, Mice, Cues, Male, Neurons, Models, Neurological, Learning, Neural Pathways
Cognitive flexibility—the ability to update behavior when cue-reward relationships change—depends on the medial prefrontal cortex (mPFC), but the underlying circuit mechanisms were poorly understood. Researchers built a meta-learning extension of a standard reward prediction error model to better capture how mice adjust licking behavior when cue-reward contingencies are degraded or enhanced. Using longitudinal two-photon calcium imaging and single-cell holographic optogenetics in mice, they identified a specific subset of mPFC neurons that causally encode contingency degradation. They then showed that these neurons project to the ventral tegmental area (VTA), a core reward-processing hub, with most mPFC→VTA projection neurons carrying this degradation signal. Optogenetic stimulation of these ensembles was sufficient to accelerate behavioral updating. The work maps a prefrontal-to-midbrain circuit through which the brain selectively suppresses outdated learned behaviors when reward contingencies change.
Models, Molecular, Capsid, Animals, Protein Engineering, Protein Multimerization, Molecular Weight, Humans
Viral capsids achieve large sizes by using a single protein subunit in slightly different conformations to tile both pentagonal and hexagonal arrangements — a property called quasisymmetry. Replicating this architecture synthetically has been difficult. Researchers developed a computational design approach based on geometric frustration to build two-component protein cages that mimic this strategy. Trimeric and dimeric protein building blocks self-assemble into hexagonal lattices that cannot tile a flat surface, so they spontaneously close into spheres by incorporating pentagonal defects — just as in natural viruses. By tuning the curvature encoded in the dimer component, cage diameters ranging from 40 to over 200 nm and masses from 2 to over 50 MDa were achieved. These cages were further engineered to carry RNA-protein cargo and enter cells. When expressed in mammalian cells, fluorescently labeled versions served as probes for studying how molecular size affects movement through the cytoplasm and protein localization, with potential uses in drug delivery.
Capsid, Capsid Proteins, Cryoelectron Microscopy, Models, Molecular, Nanostructures, Protein Subunits
Viruses build enormous shells from thousands of identical protein subunits by placing the same piece in geometrically distinct positions — a principle called quasisymmetry. Replicating this feat through rational protein design is difficult because the same subunit must fold similarly yet interact differently depending on its location. Researchers hypothesized that quasisymmetry could emerge spontaneously from symmetry breaking when building blocks with precisely programmed curvatures interact strongly with each other. Combining a parametric description of cage geometry with RoseTTAFold diffusion-based generative modeling, they designed a wide variety of such assemblies. Electron microscopy confirmed cages ranging from T=3 to T=36, containing 180–2,160 subunits and spanning 68–220 nm in diameter, plus clathrin-like non-icosahedral structures. Cryo-EM revealed how interface-level symmetry breaking drives the formation of pentagonal and hexagonal units in T=3 architectures. These results offer a practical framework for engineering large, single-component protein vessels for drug and biologic delivery.
Humans, Bacteria, Evolution, Molecular, Genome, Bacterial, Phylogeny, Recombination, Genetic, Selection, Genetic, Gastrointestinal Microbiome
Researchers investigated how bacterial populations in the human gut become ecologically distinct, examining thousands of microbial genomes from geographically and ethnically diverse people. By filtering out recombination signals that obscure evolutionary history, they identified genome-wide selective sweeps — events where a single well-adapted clone outcompetes all others in its niche before diversifying again — as a widespread mechanism shaping gut microbiome structure. This pattern was detected in at least 66 taxa spanning 25 bacterial families and produces clusters of closely related genomes sitting on long branches in phylogenetic trees, resembling global epidemics. These sweep clusters appear to spread worldwide within decades and have recurred throughout human history. Crucially, different sweep clusters associate with distinct host states including aging, colorectal cancer, inflammatory bowel disease, and type 2 diabetes, confirming their ecological differentiation and offering a mechanistic explanation for why certain heritable microbial strains show consistent disease associations.
Artificial Intelligence, Artificial Life, Biosynthetic Pathways, Computer Simulation, Generative Artificial Intelligence, Large Language Models, Models, Biological, Saccharomyces cerevisiae, Synthetic Biology
Researchers propose a framework called 'virtual yeast' — an AI-driven computational model designed to simulate eukaryotic cell behavior using baker's yeast (Saccharomyces cerevisiae) as the testbed. The architecture breaks cellular complexity into eight functional modules covering genetic, metabolic, and structural systems, each handled by a specialized AI tool coordinated through a large language model. The system draws on three data pillars: mechanistic knowledge, subcellular architecture, and dynamic cell states, combining representation learning and generative modeling in a closed-loop pipeline capable of autonomously designing and running experiments. The goal is to optimize biosynthetic pathways, generate and rank biological hypotheses, and speed up target discovery. By pairing biological accuracy with autonomous AI reasoning, the framework is intended as a generalizable blueprint for building virtual eukaryotic cells more broadly, with implications for synthetic biology and systems-level understanding of cell biology.
Animals, Mice, Hippocampus, Male, Neural Pathways, CA3 Region, Hippocampal, CA1 Region, Hippocampal, Dentate Gyrus, Action Potentials, Mice, Inbred C57BL, Gyrus Cinguli, Cerebral Cortex, Neurons
How the hippocampus transforms incoming signals into outputs relevant for navigation and memory remains poorly understood at the circuit level. Researchers implanted high-density electrode arrays (up to 1,024 channels) in mice to simultaneously record spiking activity across the dentate gyrus, CA3, CA2, CA1, and retrosplenial cortex (RSC) during behavior. Using partial canonical correlation analysis, they identified low-dimensional 'communication subspaces' linking pairs of regions while controlling for a third area's influence. These subspaces captured distinct input-output transformations in CA1, with membership constrained by firing properties and anatomical location—predominantly deep sublayers of the CA3–CA1–RSC axis. Overlapping neuronal pools could be recombined to support different interareal interactions as tasks or brain states changed. Notably, CA1–CA3 subspace reactivation during post-experience sleep correlated with replay, whereas CA1–RSC subspaces did not, suggesting that hippocampal circuits balance plasticity and stability by reconfiguring predetermined connectivity motifs to flexibly encode new experiences.
Death, Sudden, Cardiac, Humans, Electrocardiography, Deep Learning, Biomarkers, Sweden, Taiwan, Arrhythmias, Cardiac, Female, Male, Death Certificates, United States, Stroke Volume, Reproducibility of Results
Researchers used deep learning to analyze a large Swedish dataset linking electrocardiograms (ECGs) to death certificates, aiming to improve prediction of sudden cardiac death (SCD) beyond the standard biomarker—left ventricular ejection fraction (LVEF). The resulting model identified a high-risk group (2.2% of patients) with a 7.0% annual SCD rate, surpassing the 4.6% rate seen with reduced LVEF, and critically, 86% of the model's high-risk cases were missed by LVEF screening. Patients in this high-risk group who received defibrillators showed a 54% lower-than-expected mortality. The model was validated in US and Taiwanese healthcare datasets. By pairing the predictive model with a generative ECG model, the researchers visualized a previously undescribed waveform feature driving the predictions. Linking this novel biomarker to electrophysiological principles, they also propose a new hypothesis about the underlying mechanism of sudden cardiac death.
Animals, Female, Humans, Male, China, Dental Enamel Proteins, Fossils, History, Ancient, Hominidae, Molecular Sequence Data, Neanderthals, Phylogeny
Researchers extracted and analyzed ancient enamel proteins from six Homo erectus fossils (~400,000 years old) recovered from three Chinese sites: Zhoukoudian, Hexian, and Sunjiadong. All six specimens — five male, one female — shared two amino acid variants in the enamel protein AMBN. One variant (A253G) is entirely novel and absent from every other known human lineage, while the other (M273V) was previously found only in Denisovans. This finding suggests that the M273V variant may have entered the Denisovan lineage through interbreeding with populations related to these Chinese H. erectus. It also implies that genomic regions in Denisovans previously labeled 'super-archaic introgression' — some of which later reached modern humans — likely trace back to H. erectus. The results raise the possibility that late Middle Pleistocene H. erectus and Denisovans overlapped geographically in East Asia and interbred.
Comments (0)
No comments yet. Log in to leave a comment.
You must be logged in to leave a comment.
Login to Comment