Biology subjects

Zhang, Y.

Publications and source records attributed to Zhang, Y..

At least 73 records · Page 4Linked to original sources

bioRxiv2026

A unified framework for quantitative splicing and transcript prediction

RNA splicing is frequently disrupted by genetic variation, contributing to a wide range of human diseases. Identifying the variants responsible for altered splicing is therefore crucial but has long been challenging. Deep learning models that predict splicing from genomic sequence have substantially advanced their detection, but accurately resolving their quantitative effects on splicing and the resulting transcripts remains an unmet need. Here we introduce SpliceAI2, a deep learning model trained end-to-end on splice site, splice junction, and transcript usage across multiple mammalian species. SpliceAI2 reconstructs complete transcripts directly from genomic sequence and considerably improves identification of splice-altering variants and quantification of their effects in diverse benchmarks spanning cryptic splicing, splicing quantitative trait loci, and massively parallel reporter assays. Human population datasets provide independent validation of these predicted functional effects through evidence of natural selection and associated changes in plasma protein abundance. Application to a rare disease cohort reveals that predicted splice-altering variants are highly enriched in disease-associated genes and account for 15% of the excess genetic burden, with approximately half residing in deep intronic regions likely missed by exome sequencing. Together, these findings establish SpliceAI2 as a major step forward in interpreting splice-altering variants across the human genome.

genomics↗

bioRxiv2026

Learning Interpretable Switching Dynamics in Shared Neural-Behavioral Latent Space

Modern recording technologies enable simultaneous measurement of high-dimensional neural activity and rich behavioral variables, creating both an opportunity and a modeling challenge: identifying the shared latent structure that mediates the brain-behavior relationship and the underlying dynamics through which it evolves across behavioral states. Existing approaches focus on different parts of this challenge: shared representation learning methods identify neural-behavioral subspaces but typically do not model dynamical structure, while dynamical models capture temporal variation and discrete state transitions in neural activity, but do not explicitly disentangle shared neural-behavioral structure from modality-specific variability. We introduce Switching Shared Latent Dynamics (SSLD), a unified framework that learns a latent representation shared between neural activity and behavior and endows this shared space with nonlinear switching recurrent dynamics to reflect the hypothesis that behaviorally relevant neural representations exhibit structured dynamical changes across behavioral states. Complementary private latent variables capture modality-specific variability, providing a window into neural activity that does not directly interact with behavior. We evaluate SSLD on a simulated dataset and four diverse experimental datasets spanning species, brain regions, and recording modalities: monkey motor and premotor cortex during reaching, somatosensory cortex during a bump task, widefield calcium imaging across mouse dorsal cortex during (a) self-initiated decision-making and (b) spontaneous movements. Across all four datasets, SSLD accurately reconstructs neural and behavioral signals, recovers discrete states in the dynamics that align with experimentally defined behavioral epochs, and isolates behaviorally-relevant neural information in the shared latent while preserving private neural variability. Ablations confirm that shared representation learning, behavioral supervision, and switching dynamics each contribute to performance. SSLD offers an interpretable approach to modeling shared neural-behavioral dynamics that complements ongoing efforts toward foundation models for neuroscience.

neuroscience↗

bioRxiv2026

Defense-associated reverse transcriptases synthesize DNA in a telomerase-like manner

Defense-associated reverse transcriptases (DRTs) constitute an emerging class of bacterial defense systems that harness reverse-transcriptase activity to generate cDNA products and counter phage infection. A distinct subset of DRT10 systems evolutionarily related to telomerase generates tandem-repeat DNA to confer antiphage immunity. Here, we biochemically characterize the reverse transcription activity of DRT10 and determine cryo-EM structures of DRT10-ncRNA complex in three functional states, revealing an unanticipated 2:1 RT:ncRNA assembly mode. Two DRT10 protomers dimerize through a central pair of interlocked {beta}-hairpins and are further embraced by a single ncRNA. This architecture distinguishes DRT10 from previously characterized DRT complexes, which either lack an ncRNA component or exhibit an equivalent stoichiometry of RT:ncRNA. Strikingly, despite the nearly identical conformations of the two DRT10 protomers, cDNA product density was resolved in only one catalytic pocket, revealing different product occupancy within the dimer. Moreover, along with flanking linkers, the stem loops of the ncRNA as fixed points to supply steric barriers help define the template boundary. These findings reveal a distinctive ncRNA-mediated DRT10 dimeric assembly and establish a mechanistic framework for protein-primed reverse transcription constrained by the length of the RNA flanking linker.

biochemistry↗

bioRxiv2026

Neutrophils orchestrate pre-metastatic immunosuppressive bone marrow remodeling during prostate cancer metastasis

The dynamic reprogramming of the bone marrow (BM) microenvironment during metastatic progression is poorly defined. Through single-cell profiling of BM samples spanning benign conditions, high-risk bone metastatic potential localized prostate cancer (HMP) and established bone metastatic prostate cancer (PCa), we uncovered a primary tumor-to-BM signaling axis that orchestrates niche formation through neutrophil-mediated remodeling. Tumor-associated neutrophils (TANs) emerge during the HMP stage, persist throughout metastatic progression, correlate with poor patient prognosis, and exhibit direct immunosuppressive activity that contributes to a BM environment permissive to metastatic colonization. Functional and genetic studies showed that tumor-associated systemic IGF1 reprogrammed BM neutrophils through IGF1R-MAPK/ERK-KLF4 signaling, with IL-1{beta}, reactive oxygen species and nitric oxide contributing to TAN-mediated T cell suppression. Tumor-cell Igf1 perturbation, IGF1R blockade and neutrophil-specific Igf1r deletion attenuated TAN-associated BM remodeling, and reduced subsequent bone metastatic progression. In bone-metastasis models, combined IGF1R and PD-1 blockade suppressed bone-metastatic progression. Translationally, our investigator-initiated clinical trial (NCT06866548) of teprotumumab plus tislelizumab in patients with established metastatic castration-resistant prostate cancer (mCRPC) provided initial safety data and exploratory clinical observations. Together, these findings identify IGF1R-dependent neutrophil reprogramming as a functionally important component of the evolving PCa bone metastatic microenvironment and provide a rationale for therapeutic targeting of IGF1R signaling together with immune checkpoint blockade.

cancer biology↗

bioRxiv2026

VINTER: a generative vision-language model for annotating and interrogating spatial tissues

Spatial transcriptomics links gene expression to tissue structure, yet current models learn embeddings whose meaning is assigned afterwards or read histology without measured expression. Tissue decisions are therefore rarely made from morphology, genes and neighbourhood together, or examined through them. Here we introduce VINTER, a generative vision-language model aligned on 1.45 million histology-caption pairs that reads histology, measured gene expression and spatial neighbourhood as one sequence. It returns the probability of every tissue class, and because genes enter as text, its molecular evidence can be edited with the image held fixed. Reading all three sources together, VINTER annotated tissue more accurately than spatial-domain methods and a spatial foundation model, and transferred across cohorts without target labels. Editing only the neighbourhood resolved the follicular B-cell and plasma-cell programmes behind its reading of tertiary lymphoid structure (TLS) maturation in lung and kidney cancer. Its class probabilities exposed what single labels conceal in triple-negative breast cancer: immune heterogeneity within stroma and cancer hidden beneath benign labels. Without retraining, the same model transferred to Visium HD and single-cell Xenium data and resolved follicular architecture within colorectal cancer TLS. VINTER thus turns spatial annotation into an interrogable reading of tissue across platforms and resolutions.

bioinformatics↗

bioRxiv2026

PPARγ brakes cytosolic DNA-induced macrophage inflammation by super-enhancer metabolic reprogramming

Cytoplasmic DNA sensing triggers intense innate immune responses that require tight regulation. PPAR{gamma} is a master metabolic regulator in macrophages, but its role in cytoplasmic DNA immunity remains unclear. Using poly(dA:dT) to mimic cytoplasmic DNA stimulation, we combined transcriptomics, epigenomics (PPAR{gamma}/H3K27ac ChIP-seq), and functional assays. Poly(dA:dT) induced a cytokine storm with MAPK activation and apoptosis/necrosis. PPAR{gamma} activation by rosiglitazone (RGZ) antagonized this response, reducing cell death and transcriptionally inhibiting MAPK genes (e.g., DUSP4/10). Mechanistically, PPAR{gamma} bound promoters/enhancers via PPRE motifs and distal intergenic regions, remodeling H3K27ac (reduced at inflammatory loci, increased at metabolic genes). RGZ induced de novo super-enhancers coupling PPAR metabolic programs (e.g., CPT2) with immune suppression by targeting TLR4, CEBPB, and PIM3. Thus, PPAR{gamma} acts as a brake in cytoplasmic DNA immunity, integrating metabolic reprogramming and inflammation suppression through a multi-level genomic/epigenetic program. This study highlights PPAR{gamma} as a key immune-metabolic integrator and therapeutic target for diseases driven by excessive cytoplasmic DNA sensing.

immunology↗

bioRxiv2026

ddkg.skill: A Compositional Agent Skill for Translating Biomedical and Bioinformatics Questions into Cypher for the Data Distillery Knowledge Graph

Biomedical knowledge graphs can connect information across genes, phenotypes, tissues, pathways, experiments, and clinical resources, but they are difficult to query correctly without detailed knowledge of the graph. Large language models can help write Cypher code for knowledge graphs, yet a query focused on a bioinformatics task that looks reasonable may still use the wrong identifier, relationship direction, source, intermediate node, or output unit if the user is not completely trained on the schema. We developed ddkg.skill, an Agent Skill that can be loaded into a compatible LLM session to support querying of the NIH Common Fund Data Ecosystem Data Distillery Knowledge Graph (DDKG), a biomedical property graph integrating more than 180 ontologies, over 40 genomics datasets, and dozens of additional cross-domain biomedical datasets. Rather than utilizing a simple markdown prompt, ddkg.skill contains a controller, a compositional library of DDKG-specific references and structured tables, primary documentation, validated query patterns, a routing table, and a Python script that checks the internal links among these materials. The skill build for the December 2025 DDKG release contains 38 bundled files and 239 routing relationships and is identified by checksum so users can state the exact build used to compose a query. The evaluation reported here was performed separately on an earlier build with 38 files and 227 routing entries. In that orthogonal nine-test evaluation against a live December 2025 DDKG instance, five of seven tests targeting sources absent from the worked examples produced correct executed results. The evaluation also exposed a failed query, a cross-source comparison that was not biologically well posed, and errors in the skills own reference material that informed later revisions. The skill guides an AI through entity resolution, graph inspection, query construction, and validation for biomedical and bioinformatics queries, while keeping the DDKG itself as the source of returned results. This design provides a portable and versioned method for giving general-purpose AI systems practical knowledge of a complex biomedical knowledge graph to empower complex bioinformatics data integration. Three executed biomedical use cases further demonstrate cross-species phenotype-to-expression querying, genomic overlap between 4DN chromatin-loop anchors and GTEx eQTLs, and integration of congenital heart phenotype associations with ClinVar and GTEx heart evidence.

bioinformatics↗

bioRxiv2026

Achieving near single-molecule sensitivity in separation-free protein detection using engineered kinetic proofreading

Unlike nucleic acids, proteins cannot be directly amplified. Ultrasensitive detection must therefore couple efficient analyte-to-signal conversion with strong suppression of target-independent background. Consistently detecting ~100 molecules (below which Poisson sampling noise alone causes CV to exceed 10%) across protein targets in a format as scalable and accessible as PCR remains unsolved. Separation-based ultrasensitive immunoassays require target binding stable enough to survive stringent washing and are therefore constrained by wash-induced signal loss and, more fundamentally, the relative scarcity of exceptionally tight binders. Performance consequently varies across targets; detection limits below 1,000 molecules remain rare. Conventional separation-free proximity assays avoid washing but incur background from random probe collisions, with reported detection limits frequently exceeding 100,000 molecules. Here we introduce Successive Proximity Extension Amplification Reaction (SPEAR), a separation-free architecture in which autonomous engineered DNA nanodevices emulate biological kinetic proofreading. Two successive irreversible proximity-extension steps compound discrimination between analyte-dependent and background reactions, suppressing spurious signals while preserving efficient analyte-to-signal conversion. This design reduces reliance on high-affinity binders and enables sub-fM detection from 1 uL samples through simple 3-step mixing followed by standard qPCR readout. Across 11 biomarkers tested, SPEAR consistently achieved sub-100-molecule detection limits. Where compared, it outperformed ultrasensitive separation-based platforms, analytically and clinically, including for low-abundance Alzheimer's disease biomarkers. SPEAR also precisely quantified SARS-CoV-2 neutralizing antibody titers from finger-prick dried blood spots, concordant with the cell-based reference assay on venous blood. It detected a distinct mixed phosphorylation state of Tau, and targeting this state resolved previously observed capillary-venous discordance and improved amyloid-status discrimination. These results establish engineered autonomous successive molecular discrimination as a practical strategy for near-single-molecule protein detection, extending the sensitivity, scalability and accessibility of PCR-based testing to non-amplifiable analytes.

bioengineering↗

bioRxiv2026

Autonomic and neurotrophic signaling in alcohol-associated liver disease

Alcohol-associated liver disease (ALD) is a leading cause of liver-related morbidity and mortality, yet its pathogenesis remains incompletely understood. While neural regulation is critical for hepatic homeostasis, how autonomic and neurotrophic signaling is regulated in the liver during ALD progression remains largely unexplored. Here, we performed an integrative analysis of bulk and single-cell transcriptomic datasets of human livers across the spectrum of ALD. We identified stage-dependent remodeling and coordinated expression patterns of autonomic and neurotrophic genes. Autonomic signaling exhibited a consistent decline, with reduced expression of sympathetic (DBH, TH) and the parasympathetic (ChAT) markers. In contrast, pan-neuronal and neuroplasticity-associated genes (UCHL1, TUBB3, STMN2) were upregulated, indicating active neural remodeling. In parallel, factors such as MANF, GDNF, and NRG2, with the receptor NTRK1, were upregulated and positively correlated with clinical indicators of disease severity, inflammation and fibrosis. In contrast, trophic factors including NTF3 and CDNF, and receptors such as GFRA1 and GFRA3, were downregulated and inversely associated with these parameters. Single-cell analyses revealed that neurotrophic factors are predominantly expressed by hepatic stellate cells, supporting their role as resident neuroglia-like cells in the liver. In another independent cohort, we observed a coordinated downregulation of NTF3 and sympathetic nerve markers in alcohol-associated hepatitis (AH), suggesting impaired NTF3-mediated neurotrophic support may contribute to sympathetic dysfunction in AH. Collectively, these findings uncover previously unrecognized neuro-regulatory dimensions in ALD and suggest that dysregulated autonomic and neurotrophic signaling may contribute to disease progression. Conflict of interestThe authors have declared that no conflict of interest exists.

bioinformatics↗

bioRxiv2026

A direct RNA sequencing workflow for simultaneous profiling of tRNA abundance and modifications in mouse liver

Transfer RNA (tRNA) is highly modified and difficult to profile using reverse-transcription-based sequencing. Here, we established a direct RNA sequencing (tRNA-DRS) workflow to simultaneously characterize tRNA abundance and modification signals in mouse liver. The workflow included small-RNA enrichment, deacylation, splint-adapter ligation, barcoding, reverse transcription, and nanopore sequencing on PromethION with modification-aware basecalling. From 1,678,666 raw reads, 1,421,279 (84.67%) reads were retained after 20-200 nt length filtering, and 17,094 uniquely mapped reads (MAPQ [≥] 10) were obtained. These reads covered 244 tRNA genes and represented all 23 tRNA isotypes. tRNA abundance was strongly skewed, with Asp, Gly, Ser, and Lys accounting for 56.6% of uniquely mapped reads. Read-body coverage indicated that most uniquely mapped reads spanned substantial portions of tRNA molecules (mean 71.1%, median 75.9%; 83.3% reads covering [≥]50% of the reference). Modification analysis identified 41,013 candidate sites, of which 711 passed stringent filtering (modification fraction [≥]50%, coverage [≥]5), spanning eight modification types and dominated by pseudouridine and m5C. Filtered sites showed position-specific clustering and isotype-dependent patterns. Together, these results support tRNA-DRS as a practical single-assay framework for simultaneous tRNA expression and epitranscriptomic profiling in mammalian tissue.

molecular biology↗

bioRxiv2026

The RNA-binding protein TRIM71 Promotes Supporting Cell Reprogramming and Hair Cell Regeneration by Activating Sall4

Hair cell (HC) loss in the auditory organ (cochlea) is irreversible and represents the leading cause of deafness in humans. The absence of HC regeneration is primarily due to the limited plasticity of cochlear supporting cells (SCs). Using a novel doxycycline-inducible TRIM71 transgenic mouse model, this study demonstrates that overexpression of the RNA-binding protein TRIM71 promotes spontaneous cochlear HC regeneration in neonatal mice. Experiments in cochlear organoids indicate that TRIM71 reprograms SCs into progenitor-like cells and that this reprogramming is further enhanced by coactivation of follistatin (FST). FST augments TRIM71-induced increases in chromatin accessibility at progenitor and HC-specific gene loci, facilitating their transcriptional activation in SCs. Notably, FST also enhances TRIM71-induced reactivation of Sall4, which, like Trim71, is highly expressed in early otic progenitors. Using cochlear organoids and explants, we show that the early otic transcription factor SALL4 is required for TRIM71-induced HC formation. By contrast, SALL4 overexpression inhibits HC formation by maintaining reprogrammed SCs in an undifferentiated, proliferative state. These findings establish TRIM71 as an activator of HC regeneration in vivo, provide mechanistic insights into TRIM71's role in SC reprogramming, and identify SALL4 as its essential effector.

cell biology↗

bioRxiv2026

Baseline growth rate shapes pollutant impacts in aquatic environments: Evidence from antibiotic responses in the harmful cyanobacterium Microcystis

Hormesis is an important consideration in chemical risk assessment across most aquatic environments. Evidence suggests that hormetic responses are shaped by organismal physiological processes. Accordingly, predictions of hormesis under environmentally realistic conditions should account for physiological traits. Yet the role of these traits remains largely unexplored. This study investigated how baseline growth rate modulates antibiotic responses in the harmful cyanobacterium Microcystis and the underlying mechanisms. Results showed that at an antibiotic concentration of 200 ng L-1 (representing environmental concentration levels), slow-growing Microcystis populations exhibited an obvious growth enhancement more than 4-fold greater than that of fast-growing populations, and this enhancement was positively correlated with cellular reserves. A baseline growth rate below 0.19 day-1 emerged as the threshold below which hormetic responses were consistently observed. This study provides pioneering insights into how organisms' physiological states influence pollutant-induced hormetic responses in aquatic environments. Our findings point to a central principle for extrapolation that laboratory-derived toxicological responses may not translate directly to the field unless key physiological traits of the organisms are explicitly considered. In this regard, baseline growth rate and energy reserves may be key physiological traits for reliably predicting toxic effects at the population level. This study also makes an initial attempt to conceptualize a physiology-anchored framework for strengthening risk assessment of low-level emerging contaminants.

ecology↗

bioRxiv2026

Communication as a Two-Person Neural Process: fNIRS Hyperscanning in Adults with and without Post-Stroke Aphasia

Communication is inherently interactive, yet most neuroimaging studies examine individuals in isolation. Here, we used functional near-infrared spectroscopy (fNIRS) hyperscanning to investigate interbrain synchrony (IBS) during naturalistic communication. In Study 1, 13 healthy familiar dyads completed Joint Singing, Joint Reading, and Picture Guessing tasks. IBS was quantified using wavelet transform coherence across bilateral frontal and temporoparietal regions. Joint Singing produced higher IBS than the other tasks in several regions, including the dorsolateral prefrontal cortex and temporoparietal junction, and selected effects were greater in true than pseudo-dyads. In Study 2, we applied the same paradigm to five individuals with chronic post-stroke aphasia interacting with trained speech-language pathologists. IBS could be estimated in all dyads after lesion-informed channel exclusion, although regional coverage varied with lesion location. These findings demonstrate that fNIRS hyperscanning can capture task-dependent interpersonal neural coupling during spoken interaction and can be extended to communication involving individuals with aphasia. This approach provides a foundation for studying how neural coordination between communication partners relates to functional communication and partner-supported interaction.

neuroscience↗

bioRxiv2026

Shared Fine-Scale Functional Organization in the Human Subcortex Revealed by Connectivity Hyperalignment

Subcortical structures are central to cognition, emotion, and social processing, and their dysfunction underlies a wide range of neurological and psychiatric conditions. Yet subcortical fine-scale functional organization has remained largely inaccessible to group-level fMRI analysis: nuclei are small and low-SNR, and their functional topography varies idiosyncratically across individuals in ways that anatomical alignment cannot resolve. This idiosyncrasy carries real stakes. For example, deep brain stimulation targets with more variable functional connectivity across individuals show correspondingly less consistent clinical outcomes. Here we show that this seemingly idiosyncratic subcortical organization in fact reflects shared, fine-grained structure that can be recovered by modeling each individual's connectivity in a common functional space (connectivity hyperalignment, CHA), rather than relying on anatomical correspondence. Applying CHA to fMRI data collected during naturalistic movie viewing, we found that aligning subcortical voxels by their cortical connectivity, rather than their anatomical location, sharply increased cross-subject correspondence in both stimulus-evoked response and connectivity profiles. Critically, this alignment sharpened rather than blurred local structure: neighboring voxels became more functionally distinct from one another, and cortical connectivity targets became substantially more decodable across subjects. These results indicate that human subcortical nuclei possess a common fine-scale functional architecture that is masked by anatomical variability but recoverable at the individual level. This reframes a central methodological constraint in subcortical neuroimaging: rather than trading spatial resolution for signal quality via smoothing, group-level analyses can achieve voxel-level precision by aligning function rather than anatomy, opening a path to studying cortical-subcortical circuit organization in greater detail.

neuroscience↗

bioRxiv2026

Designing enzymes for new-to-nature chemistry and non-natural substrates with AlphaProtein Novo

Creating highly active enzymes for arbitrary reactions is a transformative goal for the molecular sciences (1). Traditional protein engineering (2,3) is constrained by a reliance on existing natural starting points, which are often challenging to discover. De novo design offers the potential to overcome this by creating enzymes from first principles (4), but has yet to achieve practically relevant catalytic properties. Here we present AlphaProtein Novo (AP Novo), a machine-learning pipeline that demonstrates, for the first time, that de novo enzyme design can outperform natural sequence mining in addressing challenging chemistry. We used AP Novo to design new-to-nature nitrene transferases to synthesise the pharmacophore piperidine with unprecedented product selectivity, and to create enzymes that degrade the environmental toxin DEHP under conditions that denature natural enzymes. We also obtained state-of-the-art catalytic efficiencies on two well-studied model reactions, and through iterative design and analysis found that key enablers of success were mechanism-inspired metrics based on AlphaFold 3 predictions and sequence-ensemble-based scoring of designed backbones. Our results show that de novo design has become a powerful complement to natural enzyme diversity for discovering catalysts for a range of applications.

synthetic biology↗

bioRxiv2026

High-throughput functional genomics and transcriptomics reveal molecular mechanisms regulating the synergistic fungicidal activity of copper and zinc

Fungal diseases cause devastating crop infestations and millions of deaths annually. To improve disease treatment development, we must understand the biological processes these treatments target and the fungal cellular responses they elicit. Current methods to identify genetic responses of fungi to antifungal drugs and fungicides typically involve laborious molecular genetics experiments or transcriptomics. We utilized high-throughput screens in the basidiomycete yeast Rhodotorula (Rhodosporidium) toruloides to rapidly identify genes associated with the fungal response to a copper-zinc fungicide developed by VM Agritech and compared identified genes to those transcriptionally activated in response to the fungicide in R. toruloides and the filamentous ascomycete Neurospora crassa. These 'omics technologies combined with molecular genetics and biochemical analyses showed copper and zinc synergize to inhibit fungal growth due, at least in part, to an increase in cellular zinc concentrations when copper is present. Intriguingly, although we identified genes involved in the copper-zinc fungicide response using both high-throughput screens and transcriptomics, there was little overlap in genes identified through the two methods. Our findings demonstrate the power of combining functional genomics and transcriptomics approaches in identifying a more comprehensive map of the genetic mechanisms necessary for responding to fungicides.

microbiology↗

bioRxiv2026

MOSAIC: Learning Graph Node Embeddings from Spectrally Isolated Dominant Subspaces of Accumulated Diffusion Operators

Graph embedding transforms complex networks into low-dimensional representations for analysis and downstream machine learning. Existing methods often rely on eigenvalue ordering, random-walk sampling, or neural optimization. We introduce MOSAIC, Multistep Operator Spectral Accumulation with Isolated Components, an unsupervised framework that constructs an accumulated diffusion operator from successive powers of a graph transition matrix. MOSAIC uses Krylov subspace analysis to identify dominant spectral components according to their convergence behavior. This approach emphasizes persistent and spectrally separated diffusion modes rather than selecting embedding coordinates solely by eigenvalue magnitude. We evaluated MOSAIC using a biomedical graph projection derived from the Data Distillery Knowledge Graph that contained 5,171 human genes and 10,147 phenotypes. Across embedding dimensions from 16 to 128, MOSAIC achieved node-type classification AUROCs from 0.9863 to 0.9926 and exceeded FastRP, Node2Vec, HashGNN, and GraphSAGE. For gene-phenotype link prediction, validation and test associations were withheld before embedding generation. Under progressive removal of all edges or specific edge classes, MOSAIC retained more node-label information than FastRP, including after extreme graph sparsification. In a focused DICER1 analysis, 106 of 137 documented phenotype associations received scores above 0.90, demonstrating concordance with known pleiotropic biology and prioritizing additional relationships for independent evaluation. These results support MOSAIC as a fully unsupervised framework for learning informative graph representations. Although demonstrated using a biomedical graph projection, MOSAIC is domain independent and can be applied to other complex networks

bioinformatics↗

bioRxiv2026

Integrating cell-state and enhancer constraints discriminates models of PU.1 regulation during hematopoietic differentiation

Distinct PU.1 levels are associated with progenitor, B-cell, and macrophage states during hematopoietic differentiation. A previous dynamical-systems model attributed these levels to multistability generated by PU.1 autoregulation coupled to feedback between PU.1 abundance and cell-cycle progression, but its predictions were not quantitatively evaluated against the motivating measurements. Here, we test this mechanism using reported measurements of PU.1 abundance, its synthesis rate, and cell-cycle duration, and develop progressively more mechanistic alternatives. The original growth-feedback model neither represents the two reported progenitor groups nor reproduces the nonmonotonic relationship between PU.1 abundance and cell-cycle length. Even with experimentally measured cell-type-specific growth rates, PU.1 autoregulation alone does not reproduce the observed PU.1 abundance and synthesis rates across cell states. Reproducing the cell-state measurements requires an additional cell-state-dependent input to PU.1 production. Guided by enhancer regulation, we model positive feedback between PU.1 and C/EBP and constrain it with reporter measurements, but this architecture cannot jointly reproduce the reporter and cell-state datasets. Adding RUNX1 permits data-consistent stable states when high effective cooperativity in PU.1-dependent C/EBP activation is allowed. Thus, combined dynamical and cis-regulatory constraints support additional regulation of PU.1 production beyond autoregulation, cell-type-specific dilution, and C/EBP feedback. The RUNX1-extended architecture provides one biologically grounded realization and generates experimentally testable hypotheses for differentiation networks.

biophysics↗