Nature

Medical large language models (LLMs) have shown promise in medical knowledge retrieval and clinical reasoning. However, their capacity for quantitative reasoning over dense, longitudinal health data remains limited, largely due to the lack of effective training strategies that leverage large-scale, natural-language-based prospective cohorts. To address this, we present HealthFound, a medical LLM trained on 12.4 million self-constructed examples. These include 15-year longitudinal biomedical records from 501,936 participants, as well as datasets for medical knowledge and general reasoning. We propose a novel three-stage training framework comprising progressive curriculum supervised fine-tuning, rejection-sampled retention fine-tuning, and task-verifiable policy optimization. This enables the model to effectively parse heterogeneous quantitative follow-up data and perform robust quantitative reasoning. HealthFound achieves state-of-the-art performance across multiple newly constructed benchmarks. It attains the highest accuracy on eight public medical benchmarks for general medical competence. On 1,162 UK Biobank tasks predicting future disease onset from biomedical measurements, it outperforms the second-best model by over 10 percentage points on average. In two independent external validations (MIMIC-IV and NHANES), it demonstrates superior zero-shot generalization, achieving improvements of over 6% across 484 disease endpoints. The model also achieves top results on novel tasks for literature-derived question answering and quantitative indicator imputation. Furthermore, HealthFound generates interpretable reasoning rationales that are consistent with established biomedical knowledge. Collectively, our results demonstrate that HealthFound enables medical LLMs to evolve from static knowledge repositories into dynamic quantitative reasoning engines over longitudinal health data.

Functional neural networks emerge as developing neurons form synaptic connections. Revealing how these connections are organized and perturbed in disease requires linking single-neuron connectivity to molecular state across thousands of networks, which is challenging with current methods. We developed COSCO (Connectomics by Sequencing in Cerebral Organoids), combining barcoded monosynaptic rabies tracing with single-cell RNA sequencing to reconstruct over 10,000 networks across organoids. We demonstrate that organoid circuitry self-organizes into stereotyped network motifs governed by cell-intrinsic wiring rules aligning with features of mouse cortical connectivity. Projecting network topology onto transcriptomic space reveals that wiring correlates with cell-adhesion and synaptic gene expression programs. In a Tuberous Sclerosis Complex model, pathological hyperexcitability coincides with topological rewiring: mutant networks become more heterogeneous, driven by the transcriptional distribution of upper layer excitatory cell states. This work establishes cerebral organoids as a model to study self-organization of human connectivity, and its disruption in neurodevelopmental disease. O_FIG O_LINKSMALLFIG WIDTH=199 HEIGHT=200 SRC="FIGDIR/small/755792v1_ufig1.gif" ALT="Figure 1"> View larger version (66K): org.highwire.dtl.DTLVardef@50ade8org.highwire.dtl.DTLVardef@c97320org.highwire.dtl.DTLVardef@1a53a47org.highwire.dtl.DTLVardef@9520ad_HPS_FORMAT_FIGEXP M_FIG C_FIG

Pooling cBioPortal studies into one patient-by-gene matrix is the standard preliminary to pan-cancer analysis, and the natural join key is the patient. Three Memorial Sloan Kettering clonal-hematopoiesis studies deposit blood normals carrying the same patient identifiers as tumors deposited elsewhere, so a patient-keyed merge records a hematopoietic clone as a mutation in that patients solid tumor. We linked 1,282,483 coding calls across 140 genes from 483 studies to patients, classified every call by specimen type, and rebuilt the matrix with and without blood-derived calls. Blood normals supply 1.7% of calls but concentrate sharply: 52.8% of patients flagged DNMT3A-mutant, 21.3% of TET2 and 19.7% of CHEK2 have no tumor-derived call in that gene, against 0.4% for TP53. The affected calls carry a clonal-hematopoiesis signature: DNMT3A supplies 42.0% of blood-normal and 0.61% of tumor calls, at a median allele fraction of 0.4% against 33.4%. The contamination reaches 65 other study cohorts, almost all of them solid-tumor studies, affecting more than one patient in ten in fifteen of them. It is a property of cohort composition rather than disease: across 45 cancer types the distortion tracks the share of patients with a paired blood normal at r = 0.994, and where one disease carries two unharmonized labels the rates differ up to fivefold before correction and agree after it. Correction lowers the pan-cancer DNMT3A rate by 45% and moves survival estimates further than chance removal of the same patients would. Two filter lines, applied before aggregation, remove it.

DNA lesions impede RNA polymerase II, yet the transcriptional consequences of that stress are less understood than the lesions themselves. We reanalysed total RNA and matched whole genomes from 438 liver lesions across four divergent mouse backgrounds, one of them the separate species M. caroli, 428 of them induced by a single dose of diethylnitrosamine (DEN) and 10 arising spontaneously, together with untreated liver and, critically, histologically normal liver from treated, tumour-bearing animals (paired normal liver). Three post-transcriptional changes accompany transformation in every background. Canonical splicing efficiency falls by 3.6 to 4.8 percentage points, and in genes that acquire them, cancer-specific junctions account for a median of 20-25% of junctions and arise at weaker splice signals. Output shifts toward circular RNA: a five-method consensus pipeline identifies 56,418 recurrent backsplice junctions, 9,718 of them detected in all four backgrounds, with ~2-fold enrichment in circRNA abundance and an expanded repertoire of distal configurations. Adenosine-to-inosine editing sites fall 2.2- to 3.8-fold, most steeply at intronic sites. Exposure changes not only these outputs but their coordination. In untreated C57BL/6J liver, splicing efficiency, mean relative backsplice span and editing co-vary positively (pairwise {rho} +0.78 to +0.87). That coordination is lost in exposed tissue, and paired normal liver shows the same loss as tumour. Pooling conditions obscures the effect because opposing estimates cancel. In C57BL/6J, paired normal liver already shows tumour-like splicing, circRNA abundance and editing roughly eight months after one exposure, while backsplice architecture has begun to reorganise. These samples carry a median of 337 callable somatic single-nucleotide variants against 78,843 in tumours and show no elevation of Mki67 or cell-cycle programmes. The post-transcriptional phenotype is therefore a durable property of damaged tissue established before morphological transformation.

Extracellular recordings carry information about neuronal location, and pooling per-unit predictions across neighboring electrodes improves its readout, but whether this local pooling should be learned is untested. We present MYCEL (Message-passing Yields Cellular Embeddings of Location), a graph neural network whose nodes are spike-sorted units carrying waveform and spike-timing features selected on cell-type ground truth, and whose edges encode recording geometry. On the Allen Brain Observatory Visual Coding Neuropixels dataset, learned message passing outperforms fixed aggregation, with the largest gains where single units are least informative. On the International Brain Laboratory Brain Wide Map, MYCEL matches the strongest population model at one-tenth the pairwise interactions. On held-out animals, the embeddings recover laminar depth within visual cortex. Retrained with only the edge geometry changed, the same framework decodes anatomy in macaque visual cortex and human medial temporal lobe. Thus, how neighboring neurons are combined matters as much as the features that describe them.

The adult mammalian brain has limited regenerative capacity yet retains substantial potential for functional reorganization after experience, injury or disease1-3. While local plasticity at injury sites has been described4,5, the brain-wide consequences of a focal injury and the biological processes that drive them remain unknown. To address this, we performed an unbiased, whole brain screen of neuronal activity at single cell resolution following a focal injury to the primary visual cortex (V1) in mice. Olfactory processing areas emerged as loci of remote activation. We further show that V1 injury promotes the recruitment of adult born neurons into the olfactory bulb (OB) and potentiates cortical feedback onto bulbar circuits. Combining high-density multielectrode array recordings with two-photon calcium imaging revealed OB circuit refinement characterized by increased synchrony and sharpened tuning of principal output neurons. These circuit-level functional enhancements coincide with improved behavioral performance in odor-guided tasks. Motor cortex lesions do not elicit similar cellular and behavioral responses, suggesting that such distal adaptation only develops when injury is coupled with increased olfactory demand. Together, these data demonstrate that the adult brain can recruit alternative sensory circuits distant from a focal lesion to undergo adaptive, functionally relevant reorganization, and implicate adult neurogenesis as a contributing mechanism. Our findings expose a previously underappreciated degree of remote plasticity and reveal a novel role of adult neurogenesis in sensory compensation.

Across human malignancies, the ability of cancer cells to switch states (plasticity) drives progression, metastasis, and therapy resistance 1-3. Here we investigated the transcriptional and epigenetic basis of plasticity in small-cell lung cancer (SCLC), an aggressive, treatment-refractory cancer 4, 5. SCLC tumours display plasticity along a neuroendocrine (NE) to nonNE axis; however, the trajectories of lineage transition and the gene regulatory network underlying it, remain poorly defined 6. Using single-cell multi-omics, we resolved distinct transcriptional SCLC cell states, including a previously unrecognised Intermediate state comprised of low-identity immunogenic SCLC cells that serve as an obligatory route in NE to nonNE lineage transition. Using single-cell gene expression and chromatin accessibility data, we next developed a Boolean Bayes model for transcription factor networks (BoBa-T) to predict regulators of SCLC cell states. We validated RORB as a novel gatekeeper of the NE state that represses the nonNE identity in SCLC. RORB inactivation drives NE cells into the Intermediate state, upregulates immunogenic programs, and inhibits tumour growth in immunocompetent hosts. Our work identifies a plastic, immunogenic waypoint for SCLC lineage switching and reveals strategies to target it therapeutically.

Transposable elements are a major source of gene-regulatory functions. Yet this contribution is understood largely from ancient insertions already fixed in the genome: the successes, seen long after the fact. The polymorphic phase before fixation or loss remains poorly resolved because such alleles could be scored only as present or absent. We resolved segregating transposable-element alleles as sequences, epigenetic states, and transcript components using 231 haplotype-resolved human pangenomes with matched long-read multi-omics. Their origins extend beyond recent transposition; their regulatory consequences depend on methylation and transcript context and are often missed by gene-level analyses; they provide mechanisms at trait-associated loci, while selection favors element presence and absence alike. Regulatory innovation before fixation is therefore not one-way accumulation, but a mixture of gain, remodeling, and loss.

Structural variants (SVs), including variable number tandem repeats (VNTRs), are a major source of human genetic variation, but their impact on the proteome remains poorly characterized. Using a long-read assembly-based reference panel, we imputed 54,578 general SVs and 15,826 VNTRs in 54,306 UK Biobank participants to interrogate their effects on the abundances of 2,923 plasma proteins. SVs and VNTRs together explained ~8.0% of variant-based heritability for protein abundance, with 123 proteins driven predominantly (>80%) by SVs. We identified 8,065 independent SV-protein and 4,101 VNTR-protein associations (with 601 highly unlikely to be driven by small genetic variants) that frequently perturb active regulatory elements, topologically associating domain boundaries, and post-transcriptional mechanisms. Integrative analyses with gene expression and complex trait SV association data identified 1,353 protein-trait association pairs, such as a highly pleiotropic insertion in MAN1A2 associated with 86 distinct phenotypes and a 3' UTR deletion in SULT2A1 linked to gallstone disease, illustrating plausible mechanisms whereby SV-induced regulatory perturbations mediate disease susceptibility. This SV-protein atlas bridges a critical gap between large-scale genomic alterations and proteomic variation, providing novel mechanistic insights into human disease biology and therapeutic discovery.

Viral-induced DNA damage triggers both eukaryotic and prokaryotic innate immunity. Here, we report that self-DNA ends generated routinely during normal termination of bacterial replication suffice to trigger the P2-encoded OLD antiphage system. Normally, these physiological triggers are cleared by the dual helicase-nuclease activities of the conserved RecBCD enzyme, which acts as an immune safeguard analogous to mammalian TREX1 exonuclease and plant DBR1 debranching enzyme. However, when this clearance fails, the transient self-DNA ends persist and are recognized by ATP-bound OLD dimers via a positively charged channel across their ATPase subunits. ATP hydrolysis then activates OLD to cleave tRNAs, causing cytotoxicity. Single-cell microscopy reveals that dozens of OLD molecules associate with kilobases of adjacent DNA, amplifying the sparse physiological DNA end signal. Our findings reveal how an endogenous byproduct poses a constant autoimmune threat in bacteria, highlighting a universal molecular conflict across kingdoms and the convergent solutions evolved to resolve it.

As the most recently evolved neuroanatomical trait of our species, our globular head shape ('globularity') constitutes a promising point of entry to probe possible associations between brain morphology and cognitive traits that may be distinctive of Homo sapiens. To shed additional light on the underlying biology of globularity and its cognitive consequences, we capitalize on data pairing genetic variation to brain shape differences from magnetic resonance imaging (MRI) scans of 34,595 adults of European ancestry, and embed genome-wide associations of globularity within genetic variation underpinning up to 75 brain-related traits. Through extensive hypergeometric tests, we observe significant associations between globularity and white matter traits at the gene level, but stronger associations with cerebellar volume, ventricular size, and the fornix at the genomic variant level. Combining conjunctional false discovery rate (FDR) with genetic correlations, we notice homogeneously concordant relations between globularity and the volume of the brain ventricles, and homogeneously discordant relations with glymphatic function. We also unravel more selective, spatially organized relations between globularity and cerebellar volume, white matter connectivity and sulcal width. Spatial permutation tests reveal that sulcal width associations are systematically distributed along a sensorimotor-association gradient of cortical organization. Additionally, correlations of brain- and lifespan-wise gene expression patterns linked to globularity and the brain ventricles highlight recurrent associations with the pentose phosphate metabolic pathway. Finally, deploying an open repository of genetic variant age estimations and an extensive catalog of single nucleotide changes distinguishing hominins, we uncover a coincidence between the establishment of globularity and the time of emergence of modifications in the cerebrospinal fluid proteome. Overall, we provide a finer-grained genetic architecture for globularity, with novel biological associations to traits across evolutionary and developmental timescales, leading us to draw implications for the early development of cognitive networks but also for liabilities to neurodegeneration late in life.

Human centromeres harbor extensive structural variation that has remained largely inaccessible to human genetics because their repetitive -satellite arrays are absent from standard reference and imputation panels. Here we catalog 188,589 centromeric structural variants (SVs) from haplotype-resolved long-read assemblies of 196 Human Pangenome Reference Consortium individuals aligned to the telomere-to-telomere CHM13 reference. Resolving recurrent SVs into monophyletic components along centromeric haplotype phylogenies makes them imputable from pericentromeric SNPs (mean R2CV = 0.407 vs. 0.065 unresolved). In 124 individuals with haplotype-resolved nanopore methylation, we identify 418 same-chromosome SV-CpG associations that form local hypomethylation-enriched domains in the centromeric region, most strongly for deletions. We also show that a truncated 113-bp monomer on chromosome 8 is associated with the position of the CENP-A-associated centromere dip region (CDR) at n = 151, independent of array length. Imputed into 630 individuals with expression and splicing data, centromeric SVs are associated with 141 genes, 14 of them new relative to the SNP-based cis-QTL analysis, including the H3K9 methyltransferase SUV39H1. In 315,204 All of Us participants, we identify 170 associations between 61 SV predictors and 27 clinical phenotypes, including substance-related disorders and breast cancer. Our study brings a previously inaccessible class of variation into population-scale genetics and links centromeric SVs to centromere organization, gene expression, splicing and complex traits.

Arterial pulsations are thought to drive perivascular transport by displacing vessel walls radially. In 11 awake mice, all pial arteries also bend with each heartbeat, displacing the wall by three times the pulse-related dilatation per wall, while veins barely move. The bend has a spatial period of about 20 diameters, arrives with the dilatation, at above 1.6 m/s, and dissipates 450 times less energy in the wall.

Diverse human cell types establish specialized functions through lineage- and context-specific regulatory programs. Interpreting non-coding genetic risk requires integrated multi-omic reference maps that directly connect regulatory DNA to cellular expression across human tissues. Here we present a single-nucleus multi-omic atlas comprising 459,856 transcriptomic and chromatin accessibility profiles from 21 adult human tissues and four donors, including paired measurements from 160,688 nuclei. The atlas resolves nine cell lineages, 61 broad cell types and 313 subclusters, and identifies 1,085,062 candidate cis-regulatory elements (cCREs), including 161,270 novel elements absent from ENCODE. Regulatory activity was dominated by cell identity but refined by tissue context. Joint profiling enabled 871,177 cCRE-gene associations and revealed lineage-specific regulatory architectures. Cross-tissue accessibility further identified lineage-restricted and constitutively inaccessible chromatin domains, the latter showing preferential hypomethylation across human cancers. Furthermore, we leverage this dataset to train sequence-to-function models to predict chromatin-accessibility effects for 548,656 fine-mapped variants, identifying 18,133 high-effect variants, including 1,120 broadly active variants. Models trained for eight endothelial subtypes further resolve predicted variant effects across vascular beds. Together, this atlas provides a comprehensive cellular and computational framework for interpreting regulatory sequence, context-dependent gene control, and complex trait genetics across the human body.

Diet exerts selection on bacterial species within the mammalian gut microbiota1,2,3,4,5, but the extent, modes, genomic bases, and repeatability of gut bacterial adaptive evolution in response to dietary variation within hosts are poorly understood. Here, we show that within-host dietary transitions drive ubiquitous and highly reproducible adaptive evolution of gut bacterial populations. We deeply profiled allele-frequency dynamics in gut microbiota of an experimental cohort of litter-matched wild-derived isogenic mice either transitioned from low-fat/high-carbohydrate diets to high-fat/low-carbohydrate diets or maintained on their ancestral low-fat/high-carbohydrate diet. Dietary transitions induced significantly parallel allele-frequency changes across replicates in 61/62 core gut bacterial species within five months, whereas 0/62 species displayed significant across-replicate parallelism in the control group, indicating deterministic adaptive evolutionary responses across the gut microbiota. Adaptation proceeded through multiple modes, including strain sorting and gene-wise responses. The latter were enriched in a subset of gene families and functions across bacterial species, including invertases. Culturing revealed an invertible DNA switch in Phocaeicola sartorii that, in all replicates transitioned to high-fat/low-carbohydrate diets, localized a promoter sequence upstream of cell-surface glycan foraging genes. These findings elucidate widespread gut bacterial adaptive evolution mediated by within-host dietary variation.

Mitochondrial DNA (mtDNA) mutations cause inherited disease, accumulate with age, and recur in tumors, yet their effects remain unresolved because existing models interrogate one engineered allele at a time. We developed mtDIVE, a random C[->]T mutagenesis platform for pooled mitochondrial forward genetics in human epithelial cells and zebrafish embryos. mtDIVE generates genotypes ranging from single substitutions to >100 mutations per molecule, impairs respiration, and causes lineage-specific developmental delay. Coupling single-cell transcriptomics with mtDNA genotyping revealed that selection on mutant genomes and their transcriptional consequences depend on metabolic state and developmental lineage. Obligate OXPHOS eliminates highly mutated cells by apoptosis, whereas glycolytic conditions promote their competitive displacement. Individual loci exert modest but non-redundant effects that predict cell state beyond aggregate mutation burden, and cancer-recurrent variants are enriched among mutations linked to reduced mitochondrial transcription. Thus, mtDIVE links mtDNA genotypes to cellular phenotypes, separating the consequences of mutation identity from mutational load.

The mesoderm coordinates organogenesis and generates the cardiovascular, visceral and musculoskeletal systems, yet in mammals this germ layer is difficult to access, impeding its genetic manipulation and study. Here, we establish embryonic day (E) 7.5 exocoelomic cavity nano-injection as a scalable route for selective genetic targeting of mouse mesoderm. Combining in utero lentiviral barcoding with single-cell and spatial transcriptomics, we reconstructed mesodermal and neural crest clonal relationships across >590,000 cells and 31,000 multicellular clones from E9.5-E10.5 embryos and E16.5 livers and hearts. In liver, capsular mesothelium was closely related to hepatic stellate cells, whereas fibroblast and vascular smooth muscle lineages diverged earlier. In heart, epicardium was clonally coupled to three spatially restricted mesenchymal sub-lineages but not to cardiomyocytes, and lineage tracing separated epicardial- and neural crest-derived populations within transcriptionally convergent valve mesenchyme. This approach enables Cre-independent manipulation of mammalian mesoderm and resolves organ mesenchyme development at clonal resolution.

Hematophagy has arisen independently many times across Metazoa, and recurrent anticoagulant protein families in blood-feeders are often read as convergent recruitment - the Kunitz/BPTI domain a paradigm case, with the leech an oft-cited low-Kunitz exception. We re-examine this at the gene level and ask whether a confirmatory cross-phylum test of this blood-feeding/anticoagulant association is feasible with public genomes. Applying an auditable gene-level protocol (one longest-isoform representative per gene; conservation-checked protein to gene mapping) to eight metazoan lineages, we find no consistent, universal elevation of whole-genome gene-level Kunitz-repertoire size in these blood-feeders (blood-feeder median 18 genes vs non-blood-feeder median 39; a descriptive comparison of non-independent taxa, not a formal test). Protein-entry counts inflate gene-level Kunitz counts by up to ~4.6x (mosquito 23 to 5), and neither this inflation nor proteome-wide isoform density (1.0-2.7x) tracks diet, so protein-entry comparisons are an unreliable basis for repertoire claims. Separately, deterministic bookkeeping under a fixed topology and a no-reversal rule counts 12 independent blood-feeding origins (11 if the ancestral lamprey is treated as parasitic with two losses); a non-exhaustive screen of annotated public genomes yielded only one candidate blood/non-blood pair (bedbug), and, under the pre-registered simulation scenario, only a cross-origin heterogeneity endpoint is attainable within a realistic origin ceiling, and only under strong heterogeneity (among-origin SD >= 3-4). An exploratory, feasibility-grade secretome-composition estimate did not meet the pre-registered criterion. We offer a gene-level, annotation-aware re-analysis, a caution about isoform/annotation bias in cross-phylum comparisons, and an account of what current data can and cannot support.

Single-cell sequencing has charted the clonal landscape of tumor-infiltrating T cells, but the antigens that drive these responses remain largely undefined, limiting rational immunotherapy design. Here, we introduce TWISTAR, a platform that systematically deorphanizes single-cell-derived TCRs against tumor transcriptomes. By directly validating TCR engagement, TWISTAR exclusively recovers functionally immunogenic antigens. Applying TWISTAR to melanoma, breast, and hepatocellular carcinoma patients responding to immune checkpoint blockade, we uncover that tumor-reactive TCRs recognize a broad repertoire of immunogenic targets beyond neoantigens, including peptides derived from unannotated coding sequences and novel transcripts. Many of these antigens are shared across patients and cancer types, revealing communal targets. Furthermore, we observe immune dominance, consisting of multiple clones converging on few immune dominant epitopes. Together, these findings reveal an unexpectedly broad landscape of tumor immunogenicity, expanding opportunities for the development of pan-cancer vaccines.

Sequence-to-function models predict regulatory activity and variant effects from DNA sequence alone, and leading models such as Borzoi compute long-range interactions with self-attention. The reference genome caps unique training sequences, and new assays add labels, not sequences. This constraint favors architectures that learn more from limited sequence diversity. We compared long-range blocks with the rest of the architecture and the training data held fixed, replacing Borzoi's transformer blocks with Hydra, a bidirectional state space block built on Mamba-2. Hydra blocks ran 5.7-fold faster on 8,192-token inputs, and models built on them predicted binned coverage on held-out sequences and classified fine-mapped eQTLs more accurately. Hydra's linear cost in sequence length let us run the blocks at 32 bp resolution and remove the U-net decoder, leaving a convolution-SSM model smaller and more accurate than the one it replaced. Scaling this architecture up on an augmented track collection produced Cerberus, an ensemble of eight models over 786 kb of input. We evaluated it on GTEx v11 fine-mapped eQTL, sQTL, and paQTL benchmarks with negatives matched on allele frequency, gene expression, and phenotype-specific positional annotations to reduce confounding by these properties. Cerberus is more accurate than Borzoi on all three phenotypes, 0.692 versus 0.668 mean eQTL AUPRC across 48 tissues, and estimates eQTL effect sizes better, 0.379 versus 0.321 Spearman $\rho$. Against AlphaGenome, it trades leads in eQTL classification, ahead on variants 3 to 100 kb from the transcription start site and behind in coding sequence and within 3 kb. Interpretation of the trained blocks shows interaction range growing with block depth, from a 0.9 kb median half-life in the first block to 79 kb in the seventh, and the deepest heads anchoring on promoters, enhancers, and CTCF peaks. We release the models, the training data, the training and evaluation code, and the benchmark sets, supporting variant scoring and transfer learning.