Biology subjects

Wu, Y.

Publications and source records attributed to Wu, Y..

At least 37 records · Page 2Linked to original sources

bioRxiv2026

Dietary magnesium supplement enhances memory through Hippo signaling

Dietary magnesium (Mg2+) supplement enhances memory in mammals and Drosophila, but whether and how Mg2+ changes molecular states of neurons, remains unknown. Here we used single-cell RNA sequencing to map cell-type specific transcriptional responses to memory-enhancing dietary Mg2+. These analyses revealed changes in Hippo-pathway gene expression in memory-relevant {beta} Kenyon Cells (KCs) of the mushroom bodies. Targeted knock-down of Merlin, hippo (hpo) and warts demonstrated requirement in {beta} KCs for Mg2+-enhanced, but not baseline memory. Further study of Hippo signaling identified Yorkie (yki) transcriptional coactivator and DNA-binding partner Scalloped (sd) to be required for baseline and Mg2+-enhanced memory. Moreover, releasing Yki::Sd of inhibition from Tondu-domain containing growth inhibitor (Tgi) specifically in {beta} KCs enhanced memory, independent of diet. Analyses of Yki::Sd target genes identified several required in {beta} KCs for baseline memory, including crb and Glut1 which are also Mg2+-induced in a Hippo-dependent manner in {beta} KCs. These genes suggest that aiding the metabolic demands of long-term memory and facilitating structural neuronal plasticity are key mediators of Mg2+-induced memory enhancement. We propose pharmacological enhancement of Sd-dependent transcription, and activity of particular Sd-target genes may have therapeutic potential for memory disorders.

neuroscience↗

bioRxiv2026

APOE4 reduction in vascular-associated fibroblasts decreases cerebral amyloid angiopathy and associated vascular dysfunction

Cerebral amyloid angiopathy (CAA), characterized by cerebrovascular amyloid deposition, is strongly associated with Alzheimers disease, ischemic injury, and lobar hemorrhages. APOE4 is the strongest known genetic risk factor for CAA, yet the cellular sources of APOE4 that drive pathogenic vascular amyloid deposition remain poorly defined. Here, we identify vascular-associated fibroblasts (VAFs) as a key source of APOE4 that mediates CAA and cerebrovascular dysfunction. In post-mortem brain tissue from APOE4/E4 patients with confirmed CAA, APOE colocalizes with COL6 VAFs around A{beta}-laden vessels. Therefore, to study the contribution of APOE4 derived from VAFs in CAA, independently of parenchymal sources, we used 5xFAD mice in which the endogenous mouse Apoe coding exons are replaced with floxed human APOE4 sequences. This enables selective suppression of APOE4 expression in VAFs upon tamoxifen induction of Col1a2-CreERT+. APOE4 reduction in VAFs markedly reduced CAA, with the strongest effects observed in leptomeningeal vessels. Additionally, A{beta}40, which preferentially associates with vascular amyloid, was significantly reduced in the vascular fraction. VAF APOE4 depletion resulted in fewer small microbleeds and an approximately 50% reduction in parenchymal border macrophages within the leptomeningeal compartment, whereas astrocytic reactivity remained unchanged. Functional assessment using in vivo multi-parametric photoacoustic microscopy revealed improved arterial dilation and blood flow in amyloid-laden leptomeningeal arteries. Together, these findings establish VAF-derived APOE4 as a compartment-specific driver of CAA and cerebrovascular dysfunction. One Sentence SummaryVascular-associated fibroblasts are a major source of pathogenic APOE4 that drives vascular dysfunction in cerebral amyloid angiopathy.

neuroscience↗

bioRxiv2026

A two-step organoid screen identifies transient p53 activation as a senolytic strategy for radiation-induced salivary gland dysfunction

Radiotherapy significantly improves the survival of head and neck cancer patients. However, it frequently damages the salivary glands, causing persistent hyposalivation and xerostomia. Radiation-induced cellular senescence within the salivary gland stem/progenitor cell niche contributes to hyposalivation, suggesting that selective elimination of senescent cells could promote tissue recovery. However, the effectiveness of senolytic compounds is strongly context dependent, varying with tissue of origin and potentially impairing regenerative cells required for functional restoration. Here, we developed a two-step organoid-based screening methodology that integrates the preferential induction of apoptosis in irradiated, senescence enriched, salivary gland organoids with the preservation of stem/progenitor cell regenerative capacity. As a first step, the senolytic activity of candidate drugs was tested to identify drugs that induced apoptosis in senescent cells of irradiated organoids but were not or less toxic to unirradiated organoids. As a second step, the confirmed senolytic compounds were tested for effects on stem cell self-renewal potential. Using this approach, we identified Nutlin3a, an MDM2 inhibitor and p53 activator, as the compound with the most favorable toxicity profile in salivary glands. Last, we confirmed the anti-senescence and pro-regenerative potential of Nutlin3a in an in vivo mouse model of radiation-induced hyposalivation, where localized Nutlin3a delivery resulted in a robust increase in saliva production, as early as two weeks after administration. These findings establish an organoid-based strategy for identifying tissue-adapted senolytic interventions and propose transient local p53 activation with Nutlin3a as a potential approach to mitigate radiation-induced salivary gland dysfunction.

cell biology↗

bioRxiv2026

Evidence Accumulation in Synaptic Efficacy: Activity-Silent Decision-Making with Short-Term Plasticity

Decision-making requires accumulation of evidence over time, yet exactly how neural circuits implement this computation remains debated. Classical models posit that evidence accumulation is realized by ramping neural activity, whereas recent experimental findings suggest an alternative mechanism in which neural activity remains largely silent before abruptly transitioning into an active decision state. This mechanism leaves a fundamental question unanswered: how is evidence accumulated during the neural activity-silent period? To address this question, we trained excitatory-inhibitory recurrent neural networks (RNNs) with short-term synaptic plasticity (STP) on a reaction-time decision task. We found that after training, the networks learned to leverage STP to make decisions in an activity-silent manner. Choice-tuning analysis revealed a clear functional organization of the network, and connection-pruning analysis uncovered the coupling between choice-selective excitatory and inhibitory neural groups. Guided by the architecture of trained RNNs, we constructed a minimal neural circuit model, in which evidence accumulates in facilitated synaptic efficacy that progressively reshapes the circuit's phase portrait, eventually driving the network state across the stability boundary and triggering a rapid transition to a choice state. We further compared activity before a choice between the theoretical STP and Static models to examine the potential energy benefit of activity-silent decision-making. Together, our results suggest STP as a mechanism for evidence accumulation in activity-silent decision-making and offer insights into the design of energy-efficient artificial neural networks.

neuroscience↗

bioRxiv2026

Single-Particle STORM Imaging Quantifies Coating Heterogeneity Among Cell Membrane-Coated Nanoparticles

Cell membrane-coated nanoparticles (CNPs) are a potentially transformative biomimetic targeted-delivery platform, which can engage therapeutic targets through source-cell membrane protein functions or homotypic interactions. However, their therapeutic development is limited by the current inability to accurately resolve membrane coating completeness, and CNP particle-to-particle variation. Here, we introduce a STochastic Optical Reconstruction Microscopy (STORM)-based approach that resolves population composition and quantifies membrane coating coverage at the single-particle level. Applying it to dendritic cell membrane- and HeLa cell membrane-coated nanoparticles (DCmPs and HeLamPs), we revealed substantial heterogeneity within both particle types and a cell-type dependence of coating outcomes, corroborated by confocal imaging and flow cytometry. We also quantified membrane protein composition, including peptide major histocompatibility complex class I (pMHC-I) among DCmPs, and confirmed that DCmPs preferentially engaged antigen-specific T cells in coculture. This single-particle framework establishes a quantitative standard for CNP characterization, providing a foundation for quality control and rational design.

bioengineering↗

bioRxiv2026

Unified modeling of mutation-induced protein stability responses with OmegaTherm

Accurate prediction of mutation-induced protein stability responses, including changes in folding free energy and melting temperature, is critical for protein engineering. Existing models represent these related but non-equivalent measurements separately and often focus on single mutations. Here we present OmegaTherm, a sequence-based framework that jointly learns both stability responses for single and multiple mutations. OmegaTherm introduces a shared protein language model to capture transferable mutation context, measurement-specific encoders to preserve measurement identity, and paired-measurement-informed distillation to learn complementary knowledge from heterogeneous data. An antisymmetric prediction architecture further enforces physical consistency between forward and reverse mutations. Across independent benchmarks, OmegaTherm achieves state-of-the-art performance in quantitative prediction, mutation classification and protein-level ranking. Beyond benchmark prediction, OmegaTherm-derived stability landscapes correlate strongly with fitness landscapes from deep mutational scanning, enrich favorable variants and reveal function-associated mutation-sensitive regions. These results establish unified representation learning as an effective strategy for integrating related but non-equivalent biochemical measurements.

bioinformatics↗

bioRxiv2026

Lactate Receptor Activation Alleviates Senescence and Preserves Homeostasis of Aged Arteries

Arteries are among the first tissues to exhibit age-related dysfunction, yet the metabolic mechanisms driving vascular senescence remain poorly understood. Here, analysis of human aortic transcriptomic data identified HCAR1, encoding the lactate receptor GPR81, as one of the genes most significantly downregulated with age. We therefore investigated whether age-associated loss of GPR81 contributes to cellular senescence within the vessel wall. Senescent human endothelial cells and vascular smooth muscle cells accumulated neutral and oxidized lipids and exhibited increased labile iron and ferroptosis. Silencing GPR81 in early-passage cells recapitulated this metabolic phenotype together with multiple hallmarks of cellular senescence. Moreover, endothelial-specific deletion of GPR81 in young mice was sufficient to induce senescent cell accumulation, impaired lipid homeostasis, endothelial dysfunction, and elastin disorganization. Conversely, pharmacological activation of GPR81 with the agonist CHBA restored fatty acid metabolism, promoted glycolytic reprogramming, and attenuated ferroptotic stress and senescence-associated phenotypes. In lamin A knock-in (LAKI) progeroid mice, CHBA reduced arterial lipid accumulation and cellular senescence, shifted vascular cell composition toward a youthful state, improved endothelial integrity, and restored extracellular matrix homeostasis. Together, these findings identify age-associated loss of GPR81 as a driver of vascular metabolic dysfunction and cellular senescence and establish pharmacological GPR81 activation as a promising therapeutic strategy for preserving vascular homeostasis and mitigating age-associated cardiovascular disease.

bioengineering↗

bioRxiv2026

Decoding subtype development and function in human pluripotent stem cell-derived midbrain dopaminergic neurons

Midbrain dopaminergic (mDA) neurons comprise molecularly and functionally distinct subtypes with differential vulnerability in neurodegenerative and psychiatric disorders. However, the mechanisms specifying subtype identity remain poorly understood, and protocols for the selective derivation of human mDA subtypes are lacking. Here we establish a strategy to derive substantia nigra (A9) and ventral tegmental area (A10) mDA neurons from human pluripotent stem cells (hPSCs). A9 identity is specified by dual-SMAD activation through Activin A and BMP7 at the midbrain floor-plate progenitor stage, whereas A10 identity is promoted by BMP inhibition. A9 mDA neurons are purified based on ALDH1A1 expression, and subtype identity is maintained by continued TGF-{beta} modulation and ESRRB activation in vitro and upon transplantation in vivo. Single-cell RNA sequencing and biochemical analyses demonstrate that hPSC-derived A9 neurons exhibit increased oxidative phosphorylation, neuromelanin-like pigmentation, elevated dopamine synthesis and release, and electrophysiological properties characteristic of A9 mDA neurons in vivo. Integration with human fetal midbrain datasets confirms strong transcriptional concordance between in vitro-derived and in vivo mDA subtypes. Neuromelanin-like structures produced by hPSC-derived A9 neurons trigger pro-inflammatory cytokine secretion from hPSC-derived microglia, and A9 neurons are primed to upregulate MHC-I genes in response to interferon-{gamma}, features which may contribute to the selective vulnerability of A9 neurons. Together, these results establish a robust in vitro platform to interrogate human mDA subtype development, function, and selective vulnerability, enabling mechanistic studies relevant to Parkinson's disease and the development of cell-based therapies.

developmental biology↗

bioRxiv2026

High-order enhancer hubs buffer allelic regulatory variation through kinetic compensation

Diploid genomes carry millions of heterozygous variants in cis-regulatory DNA, yet most genes produce similar RNA output from two parental alleles. How this balance is maintained is unclear. We developed Nanopore-HiChIP, a long-read method that maps high-order enhancer hubs on each haplotype. Over half of these enhancer hubs differ in chromatin architecture and transcription-factor occupancy between homologous chromosomes, but their target genes show substantially lower rates of allele-specific expression than genes lacking hub regulation. Single-cell kinetic modeling shows that burst frequency and burst size change in opposite directions, thereby preserving balanced transcriptional output. This hub-mediated kinetic buffering is enriched at haploinsufficient genes and coincides with smaller effects of expression quantitative trait loci. Enhancer hubs therefore absorb allelic regulatory variation through kinetic compensation, protecting dosage-sensitive transcription.

genomics↗

bioRxiv2026

Multiple introductions shaped genomic diversity and demographic history of invasive box tree moth (Cydalima perspectalis) populations in North America

The box tree moth (Cydalima perspectalis) is an invasive pest of boxwood plants (Buxus spp.) that has rapidly expanded across Europe and, since its first detection in Canada in 2018, across North America. Understanding its introduction pathways and early population structure is critical for effective surveillance and management. We integrated genome-wide single nucleotide polymorphism (SNP), mitochondrial cytochrome c oxidase I (COI), and morphological data to investigate the early invasion history of C. perspectalis across North America. Genome-wide analyses revealed pronounced genetic differentiation between North American and European populations, together with substantial genetic structure within North America. Delaware was genetically distinct from other North American populations and showed greater affinity with European populations, consistent with a separate introduction history. In contrast, New York, Massachusetts, Michigan, Ohio, and Ontario showed greater genetic similarity, suggesting shared ancestry and/or regional connectivity, whereas Virginia and West Virginia exhibited additional differentiation. Mitochondrial data supported shared ancestry among most North American populations, while Delaware shared a haplotype group with the Croatian European population. D-statistic analyses identified asymmetric allele sharing among several North American and European populations, consistent with multiple introductions and subsequent admixture. Morphometric analyses revealed significant geographic variation in several body and wing traits. Together, these complementary datasets indicate that multiple introductions, regional genetic structuring, and admixture have shaped the North American invasion. These findings highlight the value of integrating genomic approaches with nursery and border surveillance to identify introduction pathways and inform targeted containment and long-term management of this emerging horticultural pest.

genomics↗

bioRxiv2026

Behaviorally prioritized entity-relation structure captures human visual cortical representations of natural scenes

Understanding natural scenes requires identifying visible entities and representing how those entities are related. Recent studies have shown that artificial neural networks (ANNs), large language models (LLMs), and vision language models (VLMs) can predict visual cortical responses to natural images. However, the neural organization of relational scene meaning remains poorly understood, in part because these models typically encode scene content in global feature spaces that are difficult to decompose into separable entity and relation components. Here, we combined scene-graph annotations, behavioral measurements, and large-scale neural datasets to characterize structured relational representations during natural vision. We used RotatE, a knowledge-graph embedding model, to represent head-relation-tail triplets annotated for images from the 7T Natural Scenes Dataset. Triplet embeddings reliably captured cortical representational structure across the visual hierarchy. Behavioral judgments further revealed systematic differences in triplet accessibility associated with visual, relational, and graph properties. Prioritizing more behaviorally accessible triplets improved neural correspondence and explained unique variance beyond object co-occurrence, ANN image features, and LLM caption embeddings. Decomposing triplet representations into entity and relation components revealed partially dissociable cortical contributions, with lateral parietal cortex showing sensitivity to both. Triplet-based semantic information also remained spatially grounded: visual-field-specific triplet models preferentially predicted voxels with matching retinotopic preferences. Finally, cross-species comparison indicated that triplet-based semantic features were relatively more aligned with human high-level visual cortex than with macaque inferotemporal cortex. Together, these findings provide new insights into the representation of semantic relational information in the human visual cortex during natural scene perception.

neuroscience↗

bioRxiv2026

Cardiomyocytes upregulated PD-L1 expression to alleviate cardiac injury induced by irradiation combined with anti-PD-1 antibody: an in vitro and in vivo study

Objective: Preclinical studies suggest that irradiation combined with anti-PD-1 antibody (iRT) exacerbates cardiac injury in mice, partly through CD8 T-cell-mediated mechanisms. We investigated radiation-induced PD-L1 expression in cardiomyocytes and explored its potential role in iRT-associated cardiac injury. Methods: AC16 cells were exposed to different doses of irradiation. PD-L1 expression was assessed at different time points by flow cytometry, qRT-PCR, western blotting, and immunofluorescence. Cell viability, apoptosis, and cytokine levels were evaluated. PD-L1 was knocked down using siRNA to assess its function. PBMCs or CD8 T cells were co-cultured with irradiated or non-irradiated AC16 cells at different E:T ratios, with or without PD-1 blockade. A C57BL/6J mouse model of iRT-induced myocardial injury was established, and myocardial PD-L1 expression and T-cell infiltration were evaluated. Results: Irradiation upregulated PD-L1 expression in AC16 cells, with increased membrane and cytoplasmic signals at 48 h and prominent nuclear-associated fluorescence at 72 h. Irradiation increased IL-6, MCP-1, CCL5, and CXCL10 release, decreased IL-10, reduced cell viability, and increased apoptosis. PD-L1 knockdown further increased inflammatory cytokine and chemokine release but did not significantly alter apoptosis or further reduce viability in irradiated AC16 cells. In the presence of a high density of activated CD8 T cells, disruption of PD-1/PD-L1 signaling was associated with a further reduction in AC16 cell viability. In vivo, myocardial PD-L1 expression was increased, whereas CD8 T-cell infiltration remained limited and focal. Conclusion: PD-L1 upregulation in cardiomyocytes may exert an adaptive protective and immunoregulatory role during iRT-associated injury, potentially by limiting inflammatory mediator release and modulating susceptibility to activated CD8 T-cell mediated injury. CD8 T-cell density and activation are the key issues.

cancer biology↗

bioRxiv2026

High-throughput physics-based enzyme engineering

Enzyme engineering aims to tailor natural enzymes for industrial and therapeutic applications, yet physically grounded rational design has been limited by a trade-off between accuracy and cost, leaving the field heavily dependent on expert intuition. Here we present a scalable physics-based framework that combines field-aware machine learning with molecular mechanics to capture enzyme electrostatics at quantum-mechanical accuracy while enabling efficient, atomistic exploration of reaction free-energy landscapes. Coupled with microkinetic modelling, the framework translates molecular free-energy landscapes into catalytic rates and selectivity across competing, multistep reaction pathways. Applied to a newly engineered oxidative amidase (OxiAm), the framework predicts catalytic rate constants with near-experimental accuracy, quantitatively resolves the selectivity between hydrolysis and aminolysis, and generalizes across substrates, mutations and enzyme homologues. Transition-state ensemble analysis further reveals the reaction mechanism and guides the design of enzyme variants for pharmaceutical synthesis. By bringing chemical accuracy and high-throughput sampling to enzyme catalysis, this approach shifts rational design from static, empirical practice toward dynamic, free-energy-driven design, and should accelerate the engineering of biocatalysts.

biophysics↗

bioRxiv2026

Incomplete references leave bulk deconvolution targets non-identifiable, but identification has a computable precision price

Background. Reference-based deconvolution estimates cell-type proportions from bulk profiles. Incomplete references compromise these estimates, yet many remedies return point estimates. We separate the operator provenance that reproduces an estimate from the information and precision that identify its target. Results. Two operator histories at one reduced reference give different estimators: 28.2% of 196,420 sample-deletion pairs differed by >0.1 total variation; locking every learned component made paths identical. Across nine cohorts, operator history reversed 160 of 952 association signs, 31 with a significant path, and changed significance for 126. Observationally equivalent completions can fill the open simplex and reverse retained-type rankings. Under a joint zero-exposure condition, shared structure leaves inherited bounds unchanged; one to six uncalibrated views also gave identical bounds. A profile library contracted estimator-output envelopes by 98.93% yet covered the full-reference effect for 42.77%, with 19 wrong-sign certificates; conditional sharp bounds stayed at [-1, 1]. Treating RNA yields as exact collapsed intervals to points covering none of seven flow-measured targets: contraction without coverage is false certainty. Calibrated cross-modal anchors contract width to 0.51 at six types but certify no valid sign. Exact RNA yields identify cell fractions from RNA contributions; a decisive sign in blood requires {+/-}2.6% proxy accuracy with near-total contraction of donor-heterogeneity and dynamic-range envelopes. Conclusions. Incomplete-reference deconvolution is an identification problem, not only an estimation problem. Remedies must be scored on shrinkage, coverage, certification and false certification against held-out targets. fitdrop implements this scoring and a precision frontier for planning decisive measurements.

bioinformatics↗

bioRxiv2026

Test-Retest Reproducibility of Single- and Cross-Population White Matter Atlases in Diffusion MRI Tractography

Diffusion MRI tractography enables noninvasive mapping of white matter fiber tracts. Atlas-based white matter parcellation supports automated tract identification by assigning individual streamlines to atlas-defined clusters and anatomical tract labels. Because clustering-based white matter atlases are constructed from cohort-specific tractography data, they capture common white matter organization represented in the atlas-construction population. Although major white matter anatomy is shared across populations, subtle population-related anatomical variability may influence atlas representation and test-retest correspondence. Therefore, the reproducibility and cross-population generalizability of tractography-based white matter atlases are important considerations for quantitative neuroimaging studies. In this study, we evaluated whether incorporating cross-population anatomical variability during atlas construction improves test-retest reproducibility. To do so, we compared a single-population ORG atlas constructed from a Western cohort with the cross-population East-West White Matter Atlas constructed from both Eastern and Western cohorts. Test-retest diffusion MRI scans from the Human Connectome Project Young Adult (HCP-YA) dataset and the Connectivity-based Brain Imaging Research Database (C-BIRD) were analyzed as independent Western and Eastern test-retest cohorts, respectively. Whole-brain tractography was reconstructed for each dMRI scan and parcellated using both atlases. Reproducibility was assessed using tract detection rate at both cluster level and anatomical tract level, weighted Dice coefficient, and the relative difference of mean fractional anisotropy (FA). Both atlases showed stable tract detection across test-retest scans in both cohorts. Compared with the ORG atlas, the East-West White Matter Atlas achieved higher overall spatial overlap and lower test-retest variability in mean FA, although atlas performance varied across individual tracts. These findings suggest that integrating cross-population information during atlas construction improves the reproducibility and generalizability of white matter atlas mapping across independent populations and imaging protocols.

neuroscience↗

bioRxiv2026

Cell type-specific histone acetylation landscape in Alzheimer's disease reveals a putative role of MITF in microglia

Alzheimer's disease (AD) is characterised by aberrant amyloid beta and tau aggregation, neuroinflammation, demyelination and neurodegeneration, which have been linked to changes in cell-specific gene expression signatures. Among the mechanisms driving cell-type-specific transcriptional changes, histone acetylation plays a central role in regulating gene activity. While global alterations in histone acetylation have been implicated in AD, the contribution of individual cell types to these epigenetic changes remains poorly understood. To decode cell-type-specific changes in gene regulation in AD pathogenesis, we profiled histone H3 lysine 27 acetylation (H3K27ac) in microglia, oligodendrocytes and neurons from the prefrontal cortex of individuals with late-stage AD and non-dementia controls. Oligodendrocytes had the highest number of differential H3K27ac regions in AD, followed by microglia. Genes nearest to differential H3K27ac in purified microglia were enriched for phagocytosis, lipid processing, inflammatory and disease-associated cell state signature genes. Gene network analysis revealed downregulation of homeostatic genes in AD microglia and upregulation of immune activation, including signatures of lipid-handling and monocyte-derived macrophages. Oligodendrocyte co-regulated regions were indicative of increased MHC class I antigen presentation and altered neuron-oligodendrocyte interactions in AD. We identified H3K27ac allele-specific variants (ASVs) enriched near endolysosomal and ubiquitin-proteasome-associated genes in microglia and neurons. ASVs coincided with AD genome-wide association study (GWAS) risk loci, including CLU in oligodendrocytes and HLA-DRB1 in microglia. DNA motif analysis identified putative transcription factor drivers of AD glial dysregulation, including the lysosomal-associated MITF, Cap'n'collar (CNC) family (BACH1 and NFE2) and AP-1 activation in microglia. DNA binding of the MITF protein in human microglia was localised to lysosomal-associated genes and enriched in H3K27ac regions upregulated in AD and near disease-associated microglia (DAM) genes. Collectively, these findings implicate lysosomal dysfunction and upstream transcriptional regulation via MITF as key processes in AD microglia.

neuroscience↗

bioRxiv2026

Concurrent model evidence computation and posterior sampling in continuous attractor network subspaces

Extensive studies suggest the brain performs Bayesian inference to infer the latent world states. It is a fundamental neuroscience question that how canonical recurrent neural circuits in the brain implement Bayesian inference. Many existing theoretical studies focused on how the recurrent circuits compute the posterior, while largely overlooking how the circuits compute the model evidence (normalization constant in Bayes' theorem), a key quantity that measures how well a model explains the observed data.Thus, it remains largely unknown about how the recurrent circuits compute the model evidence. The present study performs rigorous theoretical analyses of the continuous attractor networks, a canonical recurrent circuit model, and reveals that the nonlinear circuit dynamics can simultaneously compute the posterior and model evidence in first two dominant subspaces within the circuit dynamics. Specifically, the circuit dynamics in the stimulus feature subspace implements the Langevin posterior sampling, and the circuit dynamics in the subspace of total neuronal activity computes the model evidence in a way analogous to the evidence lower bound in stochastic variational inference. We further extend the circuit model to compute the model evidence of multiple inputs, and simulations validate the computation in the network. Our work for the first time reveals the concurrent model evidence and posterior sampling in subspaces in continuous attractor networks, significantly deepen our understanding of the computational algorithms adopted by the neural circuits.

neuroscience↗

bioRxiv2026

De novo design of ligand binding proteins using large language models alone

Protein design has rapidly advanced with the advent of sequence- and structure-based machine learning models. However, reasoned design, which applies physicochemical principles and rules derived from sequence-structure-function relationships, has not seen the same benefits from generative machine learning models. Here, we test the ability of common large language models (LLMs; e.g. Claude, ChatGPT, and Gemini) to consider design principles to generate de novo proteins that bind metals and lipophilic small molecules without copying existing sequences. Common LLMs alone are able to 1) generate protein sequences to adopt a desired fold and bind the target ligand and 2) explain the principles that motivate the design choices. Following structure prediction and filtering, we selected a small set of designs (6 to 12 designs per query) for experimental validation, affording metal binders in one round of LLM-based design (25% hit rate) and perfluorooctanoic acid binders in two rounds (25% hit rate in the second round of design). Importantly, the LLMs produce detailed justification to accompany the de novo designed sequences, providing a conceptual framework on which designs can be evaluated. While the successful designs have some deviations from the prompted parameters and LLM-articulated design rationale, these campaigns provide a case study that highlights the utility of LLMs in making protein design more comprehensible and accessible to users without sophisticated design expertise.

biochemistry↗