Biology subjects

Wen, Y.

Publications and source records attributed to Wen, Y..

At least 19 recordsLinked to original sources

bioRxiv2026

The human circular Tau 12->7 RNA is translated in mouse neurons and promotes tau aggregation

In addition to numerous mRNA isoforms, the human microtubule associated protein tau (MAPT) gene generates circular RNAs that are not known to be expressed in rodents. Using adeno associated viruses, we expressed human circular Tau RNA generated from backsplicing of exon 12 to exon 7 (circTau 12->7) in mouse brain. We observed rolling circle translation of circTau 12->7, similar to cultured human HEK293 cells. Despite using a ubiquitously active promoter, circTau 12->7 protein and a control circEGFP protein were only observed in mouse neurons, indicating neuron-specific translation of these circular RNAs. CircTau 12->7 expression resulted in tau aggregates resembling neurofibrillary tangles in mouse brain that included endogenous mouse tau protein. Our data support a model where neuron-specific translation of circTau RNAs induces 'linear', mRNA-translated tau aggregation, which could contribute to tauopathies, including Alzheimer's disease. Thus, targeting circTau RNAs could be a therapeutic approach for Alzheimer's disease.

neuroscience↗

bioRxiv2026

Pathological musculoskeletal characterisation of a new model of spontaneous OA in CD1 mice.

Osteoarthritis (OA) is a whole joint disease involving pathological changes in cartilage, synovium, ligaments, menisci and periarticular muscle. This study provides a comprehensive musculoskeletal characterisation of spontaneous OA in male CD1 mice to evaluate their suitability as a model of naturally occurring disease. Knee joints from 3-4 and 6-8 month old mice were analysed histologically, OARSI scoring and immunohistochemistry for markers of chondrogenesis and endochondral ossification. Quadriceps and gastrocnemius muscles were examined for fibre cross sectional area. CD1 mice displayed mild OA at 3-4 months, progressing to moderate to severe disease by 6-8 months. Cartilage loss, synovial hyperplasia and fibrosis increased across compartments. Anterior cruciate and collateral ligaments, and medial meniscus exhibited marked pathological remodelling, including proteoglycan rich rounded cell regions and evidence of endochondral ossification with increased COL2, SOX9, RUNX2 and ASPN expression. Gastrocnemius muscle showed significant atrophy at 6-8 months, correlating with OA severity, whereas quadriceps atrophy was not significant. These findings establish CD1 mice as a robust spontaneous OA model that recapitulates the multi-tissue complexity of human disease. Because CD1 mice are globally available, cost-effective, genetically diverse and frequently, this model offers a practical and translationally relevant platform for investigating whole-joint organ mechanisms of OA progression and for evaluating therapies aimed at modifying disease across the entire joint.

pathology↗

bioRxiv2026

Concurrent anodal tDCS enhances motor learning and increases cortical TrkB phosphorylation

The timing of transcranial direct current stimulation (tDCS) relative to motor practice may influence its behavioral effects, but the accompanying molecular responses remain incompletely characterized. Here, we examined cortical current delivery, rotarod performance, and BDNF-TrkB signaling in mice receiving anodal tDCS concurrently with training or during intertrial intervals. Cortical current measurements indicated that approximately half of the applied current reached the measurement site under the experimental stimulation configuration. During four days of rotarod training, anodal tDCS concurrent with motor tasks improved performance at selected trials and increased footstep frequency, whereas stimulation delivered before or after training trials did not significantly improve performance. Preceding online anodal tDCS with cathodal stimulation provided no detectable additional behavioral benefit. Concurrent anodal tDCS increased TrkB phosphorylation in the stimulated cortex after both one and four days, without a significant increase in BDNF levels. Paused offline anodal tDCS stimulation also increased TrkB phosphorylation after one day despite lacking a detectable behavioral benefit. No significant molecular changes were detected in the sampled non-targeted brain regions. Genetic experiments revealed reduced rotarod performance in BDNF heterozygous mice at several trials and trial-dependent responses to online stimulation in BDNF-KIV mice. Together, these findings show that concurrent anodal tDCS improves rotarod performance and increases TrkB phosphorylation in the stimulated cortex. These results motivate further investigation of how BDNF-TrkB signaling interacts with task-related neural activity during stimulation.

neuroscience↗

bioRxiv2026

Light-cycle time-restricted feeding remodels a hidden layer of the cardiac transcriptome through sex-specific transcript switching

Light-cycle time-restricted feeding disrupts daily cardiovascular and thermoregulatory rhythms, but the molecular effects of light-cycle time-restricted feeding on the heart have been measured only at the level of total gene expression. We used Oxford Nanopore long-read RNA sequencing to resolve the full-length ventricular transcriptome from male and female mice under ad libitum feeding or light-cycle time-restricted feeding across the 24-hour cycle. Greater than 20% of cardiac transcripts represent unannotated variants of known genes absent from the current GENCODE reference annotation. Light-cycle time-restricted feeding reorganizes transcript usage across hundreds of genes, including genes encoding splicing regulators, largely without changing total gene expression. The genes affected are sex-specific, with fewer than 2% of changes shared at the gene, transcript, and transcript-usage levels. We show that transcript-level regulation is a previously underrecognized component of the cardiac response to altered feeding behavior, undetected by conventional short-read approaches.

physiology↗

bioRxiv2026

Exercise preconditioning confers skeletal muscle myometaplasticity

Previously exercise trained skeletal muscle is more growth-responsive to retraining. Using a murine training-detraining-retraining approach and multi-omics, we find that Pecam1 gene expression is lower, but capillarization is appreciably higher in previously trained (preconditioned) relative to naive trained control muscle. Greater capillarity could be permissive for accelerated hypertrophic adaptation. Exercise preconditioned myonuclei feature differential promoter CpG regulation in genes related to Wnt signaling. These epigenetic alterations with retraining align with our prior observations of methylation changes within the same pathway after a longer period of chronic training, suggesting a more rapid response due to preconditioning. Methylome-transcriptome integration and single myonucleus RNA-sequencing expose the polyamine metabolism enzyme Smox as a target that relates to heightened hypertrophic adaptability with retraining. Smox induction is sufficient to cause hypertrophy in aligned myotubes cultured on a stiffness-tuned substrate along with a growth-supportive transcriptional program. Integration of our multi-omics data suggests that Smox regulates repression of Ddit4/Redd1 (an inhibitor of mTORC1 signaling) after retraining. Smox may govern a favorable muscle fiber growth environment in previously trained muscle by sensitizing anabolic potential through polyamine metabolism. A lower adaptive threshold mediated by Smox could contribute to myometaplasticity, or a change to how subsequent muscle adaptations are made.

molecular biology↗

bioRxiv2026

Skeletal Muscle Stem Cell-Derived Myonuclei Adopt Divergent Terminal Transcriptional States in Adult and Aged Muscle In Response to a Hypertrophic Stimulus

Skeletal muscle stem cells (MuSCs) give rise to a fusogenic cell population that provide new myonuclei to muscle fibers. Myonuclear functional heterogeneity has recently become appreciated, but the terminal identity of MuSC-Derived myonuclei remains undefined. We performed single-nucleus RNA-sequencing of myonuclei in Adult and Aged muscle to define MuSC-Derived and resident myonuclear responses to mechanical overload (MOV), which induces a hypertrophic stimulus. We found a MuSC-dependent induction of a youthful transcriptional signature in resident myonuclei after MOV in Aged muscle. Age determined terminal transcriptional states of MuSC-Derived myonuclei toward MTJ in Adult, NMJ in Aged, and muscle spindles in both ages. Microtubule-remodeling genes, Macf1, Map1b, and Nav3, along with the transcription factor Runx1, identified this post-fusion specialization with greater expression of these genes in Adult than in Aged MuSC-Derived myonuclei. In-silico transcription factor KO screen identified Runx1 as a regulator of post-fusion specialization and Esrrg as a driver of spindle (intrafusal) MuSC-Derived myonuclear maturation. By defining the age-associated fate of MuSC fusion to muscle fibers, we provide potential targets for modulating muscle plasticity.

molecular biology↗

bioRxiv2026

Functional evaluation of a natural AAV capsid liver targeting motif in human hepatocytes

Background & Aims: Adeno-associated virus (AAV) vectors are attractive delivery vehicles for therapeutic gene delivery, and a notable feature of most AAVs is their natural tropism for the liver, which leads to significant hepatic uptake following systemic administration. In previous work, we identified 266G as a conserved motif on a variable region on the capsid of many commonly used AAV variants that controls liver uptake in both mice and non-human primates. This single amino acid could be functionally leveraged to engineer AAVs to either de-target from or enhance tropism to the liver. Here, we explored whether these observations extended to the human context. Methods: Two human hepatocyte models were tested: Fah-/-/Rag2-/-/Il2rg-/- (FRG) mice with humanized livers and a bioengineered human microliver platform in vitro. A barcoded AAV capsid library including standard control serotypes were used to assess the role of the 266G motif on gene transfer and transgene expression in both liver systems. Results: In vivo, 266G containing AAVs indeed targeted human hepatocytes superiorly, with some noted dependency on the degree of human-hepatocyte replacement in the chimeric mouse model. Initial studies in the micropatterned primary human hepatocyte co-culture model however demonstrated enrichment of heparin-binding AAVs, and not 266G variants. Notably, incorporation of polyethylene glycol (PEG) into the system modified the AAV transduction potential of those capsids including the liver-targeting motif, recapitulating the hepatocyte transduction pattern observed in vivo. Importantly, when PEG was used, the two human models, both at the DNA and RNA level, did correlate significantly. Conclusions: Our results showed the potential of a combinatorial AAV library for model validation and revealed the human microliver platform-PEG as a reliable system for the development of AAV therapeutics.

molecular biology↗

bioRxiv2026

Consistency Analyses of Open-source Software for Motor Unit Decomposition Using High-density Electromyography Signal

Motor unit (MU) decomposition using high-density surface electromyography (HD-sEMG) has been widely used to characterize MU behavior in neurophysiology and to build neural-machine interfaces for wearable robots. Recently, many open-source software tools for MU decomposition have been made available on GitHub, which could reduce the effort of researchers in the field. However, the consistency among these open-source tools has never been studied, making researchers hesitate to use them. In this study, we collected 7 open-source software tools on GitHub and applied them to decompose MUs from an open-source HD-sEMG dataset (including 11 isometric contraction trials) to investigate the consistency among these tools. To create a comprehensive MU pool for reference, we combined all unique MUs identified by seven tools, visually inspected and removed bad MUs, and manually edited all remaining MU spike trains. Across 7 tools for 11 trials, the number of identified MUs ranges from 167 to 736. The number of valid MUs after expert inspection ranges from 29 to 210, which is 10% to 72% of the reference pool. The rate of agreement between the raw MUSTs and the manually edited MUSTs ranges from 0.86 to 0.94, and the averaged number of edits per MU to correct misalignments ranges from 14 to 39. The results show inconsistency in the implementation and procedures of each tool, which results in an inconsistent number of identified MUs and valid MUs (29 vs 210). In general, a substantial amount of effort is required to process the raw MUSTs from each tool to conduct further research analysis. This study provided a guideline for using open-source software tools for MU decomposition and indicated that it would be beneficial to develop tools to automatically edit the MUSTs.

bioengineering↗

bioRxiv2026

Transient MYC Mimicking the Exercise Response Orchestrates Multifaceted Skeletal Muscle Adaptations

Despite decades of study, MYCs physiological role in adult tissues remains obscured by the models used to investigate it. Most work relies almost exclusively on chronic or constitutive MYC induction that recapitulates the sustained activity found in tumorigenesis and expectedly produces pathological outcomes. What has gone largely untested is how MYC operates when induced in a controlled and physiologically relevant manner, as it is during adaptive processes such as exercise. Using a recombination-independent strategy in adult skeletal muscle, transient MYC bursts drive coordinated hypertrophic-metabolic reprogramming followed by a shift in myosin fiber type, recapitulating adaptations characteristic of concurrent endurance and resistance training. A single pleiotropic transcription factor governing several aspects of muscle health reframes perspectives on a gene understood almost entirely in the context of pathology. We uncover a previously unrecognized muscle-specific role for MYC in controlling muscle cell composition and present a multi-timepoint multi-omic resource for interrogating MYC: data.myoanalytics.com/study/myc_transient/gene.

Cell Biology↗

bioRxiv2026

Three-dimensional imaging reveals preserved intrinsic contractile function in aging human skeletal muscle fibers

Age-related reductions in muscle fiber size and contractile function, particularly in fibers expressing fast myosin heavy chains, contribute to declines in whole-muscle power. However, methodological limitations in estimating fiber size during contractile experiments have likely contributed to conflicting findings regarding whether reduced single-fiber force and power in older adults reflects their smaller size and/or impaired intrinsic contractile function. To address this, we coupled single-fiber contractile experiments with 3D-imaging in 7 young (19-40yrs) and 6 older (69-84yrs) males to assess intrinsic contractile function and compare agreement between 3D-derived cross-sectional area (CSA) and CSA estimates obtained either in air or solution. Fast fiber CSA from older males were [~]28-45% smaller across measurement conditions compared with young, whereas slow fiber CSA did not differ. Accordingly, absolute force and power of fast fibers were 41% and 37% lower. When normalized to CSA from measurements in air or 3D-imaging, size-specific force and power either did not differ or were greater in older adults, indicating preserved intrinsic contractile function in both fiber types. This was supported by no age-related differences in the rate of tension redevelopment (ktr), a size-independent measure of intrinsic contractile function. In contrast, size-specific force and power calculated using solution-based CSA estimates were lower in older compared with young adults, and Bland-Altman analyses demonstrated the poorest agreement between solution-based and 3D CSA measurements. These findings indicate that intrinsic contractile function is preserved with aging and suggest that methodological differences in CSA measurement contributes to the disparate findings in the literature.

physiology↗

bioRxiv2026

Functional Differentiation of Type II and Type I Collagen Articular Models in Synovial Fluid Film Formation and Recombinant Lubricin Retention

Collagen type II (Col-II) and collagen type I (Col-I) are major components of articular cartilage present at different ratios at its surface. Understanding how each of these components mediates the assembly of molecular films derived from synovial fluid (SF), the native lubricant of synovial joints, is critical to explain the loss of mechanical performance in pathological conditions, guide the design of biomaterial implants meant to be in contact with SF, and develop molecular therapies to restore SF properties. This work demonstrates that Col-II articular surface model assists in scaffolding of full SF-derived films, while Col-I model lacks SF film scaffolding capabilities. However, when Col-II and Col-I are exposed to recombinant lubricin (rLub) alone, the major boundary lubricant in SF, both adsorbed and retained similar amounts. These insights, deduced from quartz crystal microbalance with dissipation, diffuse reflectance circular dichroism, and atomic force microscopy, reveal possible mechanisms underlying the loss of mechanical performance of synovial joints in pathology, where Col-I becomes the major collagenous component of the articular cartilage surface, as well as considerations for designing functional biomaterial implants. Furthermore, this work reinforces the idea of rLub as an intra-articular osteoarthritis therapy with the ability to bind to Col-II and Col-I, irrespectively. GRAPHICAL ABSTRACT O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=170 SRC="FIGDIR/small/726594v1_ufig1.gif" ALT="Figure 1"> View larger version (69K): org.highwire.dtl.DTLVardef@138a77borg.highwire.dtl.DTLVardef@7b8512org.highwire.dtl.DTLVardef@15d7060org.highwire.dtl.DTLVardef@17ccf20_HPS_FORMAT_FIGEXP M_FIG C_FIG

biophysics↗

bioRxiv2026

Housing Mice in Thermoneutrality Causes Tissue-specific Changes in Number, Identity, and Phase of Circadian-expressed mRNA Transcripts

Most laboratory mice are housed at room temperature (20-25{degrees}C), which exposes them to chronic mild cold stress because it is below their thermoneutral temperature (30{degrees}C). We hypothesized that mild cold stress suppresses circadian gene expression in peripheral tissues. We performed RNA sequencing on hearts, livers, and diaphragms collected every 4 hours over 48 hours in constant darkness from male mice to identify transcripts with approximately 24-hour rhythms. Thermoneutral housing produced tissue-specific changes in the number, identity, and timing of rhythmic transcripts without altering the expression of core circadian clock genes. In the heart, the number of rhythmic transcripts increased fourfold, whereas the diaphragm showed a 1.5-fold increase. In the liver, the overall number of rhythmic transcripts showed little change, but their identity changed by 30%. Gene Ontology analysis revealed coordinated changes in the temporal organization of metabolic pathways in the heart and liver. Together, these findings demonstrate that ambient housing temperature is a major determinant of tissue-specific circadian gene expression, altering the abundance, identity, and timing of rhythmic transcripts independently of the core circadian clock. SignificanceScientists typically house laboratory mice at room temperature, below their thermoneutrality, forcing them to increase their metabolic rate to maintain core body temperature. Since ambient temperature plays an important role in metabolism, cold stress could disrupt circadian gene expression. Comparing mice housed at room temperature with a warmer, thermoneutral temperature, we found that housing temperature causes tissue-specific differences in rhythmically expressed genes in the heart, liver, and diaphragm, without altering core clock genes. The heart was especially sensitive, with rhythmic genes peaking at the transition between subjective light and dark cycles, increasing fourfold. These results identify ambient housing temperature as an underrecognized variable that biases circadian gene expression in cardio-metabolic tissues, affecting interpretation of preclinical studies of metabolism and disease.

physiology↗

bioRxiv2026

Complete genome-derived metabolic interactions reveal impact of gut ecology on human health

Metabolic interactions govern gut microbiome assembly, yet functional rules remain obscured by genomic incompleteness and fragmentation. Here, we leverage 1,150 complete genomes to construct genome-scale metabolic models, demonstrating that draft genomes introduce systemic artifacts missing critical transport functions. We observe that metabolic interactions are intrinsically driven by genomic traits and niche specialization rather than random association. We stratifies strains into four ecological groups, including active players, resource predators, utilizers, and contributors, each of which exhibits distinct metabolite exchange/competition signatures. This functional architecture is strongly coupled to secondary metabolism. Applying this framework to inflammatory bowel disease revealed distinct temporal dynamics of ecological groups across disease subtypes, where group-specific dysbiosis predicted clinical phenotypes better than whole community profiles. Furthermore, integrating metabolic and co-occurrence networks to identify keystone for classifiers as features significantly improved disease diagnosis. Our study provides novel insights into gut metabolic interaction landscape, bridging the gap between genomic potential and clinical utility for precision microbiome therapeutics.

microbiology↗

bioRxiv2026

Complete Genomes of Cultivated Gut Bacteria Reveal Mobile Genetic Element-Driven Functional Diversity with Therapeutic Implications

Cultivated bacterial isolates are essential for elucidating gut microbiota functions, yet reliance on fragmented draft genomes and limited mobile genetic element (MGE) annotation has constrained high-resolution analyses. Here, we present the Cultivated Complete Genome Reference (CCGR), a comprehensive compendium of 1,150 fully circularized human gut bacterial genomes. Overcoming draft limitations, CCGR achieves chromosome-scale resolution and reveals distinct evolutionary strategies shaping gene topological organization in response to bacterial growth. Furthermore, we decode the dynamic landscape of MGEs, demonstrating that broad-host-range phages follow strict positional patterns by integrating at conserved chromosomal hotspots flanked by nutrient transporter loci. Importantly, we provide mechanistic in vivo evidence that accessory genomic elements drive strain-level functional heterogeneity. Specifically, the plasmid-encoded scrK gene in Levilactobacillus brevis involved in fructose catabolism is required to mitigate fructose-diet-exacerbated colitis. Collectively, CCGR establishes a framework linking chromosomal architecture and MGEs to bacterial evolution and host health, enabling precise, strain-resolved functional investigations.

microbiology↗

bioRxiv2026

Impact of Kernel Dimensionality on the Generalizability and Efficiency of Convolutional Neural Networks to Decode Neural Drive from High-density Electromyography Signal

Convolutional neural networks (CNNs) have been widely used to estimate neural drive from high-density surface electromyography (HD-sEMG) signals in neural machine interfaces owing to their real-time capability. Depending on kernel dimensionality (1D, 2D, or 3D), CNNs can extract temporal, spatial, or spatiotemporal features. Given that motor unit action potentials propagate across both space and time, architectures that exploit spatial features may offer advantages for neural drive estimation. Despite the potential importance of kernel dimensionality, its influence on neural drive estimation remains poorly understood. Existing studies have mainly evaluated CNN generalizability across participants, contraction intensities, or muscles within the same HD-sEMG dataset, while computational efficiency has seldom been considered. As a result, it remains unclear whether different kernel dimensionalities affect cross-dataset generalizability and computational efficiency. In this study, we implemented three CNN architectures--differing only in kernel dimensionality-- to investigate whether exploiting the spatial and spatiotemporal features of motor unit action potentials improves the generalizability and computational efficiency of neural drive estimation from HD-sEMG recorded during lower-limb isometric contractions. We trained the CNNs on one HD-sEMG dataset and evaluated them, without retraining, on two independent, unseen datasets recorded from different participants, sessions, and protocols--one spanning three contraction intensities and the other three muscles. All three architectures are generalized to both unseen datasets. The 2D and 3D CNNs marginally outperformed the 1D CNN with a 0.2% increase in R, while the 3D CNN showed no advantage over the 2D CNN. Computational efficiency depended on kernel dimensionality in a platform-dependent manner. On the CPU, the 3D CNN showed the slowest inference time, which was 2x slower than the 2D and 1D CNN, owing to the higher arithmetic cost of its spatiotemporal convolutions. On the GPU, all three architectures achieved similar inference times of about 1.36 ms/sample. These findings indicate that increased architectural complexity of CNN does not improve generalizability for neural drive estimation, and that a 2D CNN offers the best balance of accuracy and efficiency for a reliable, deployable CNN-based neural drive estimator--particularly on CPU-only or resource-constrained platforms.

neuroscience↗

bioRxiv2026

Decoupling Lineage and Intrinsic Information in Single-Cell Lineage Tracing Data with Deep Disentangled Representation Learning

Single-cell RNA sequencing combined with lineage tracing technologies provides rich opportunities to study development and tumor evolution, yet existing computational methods struggle to disentangle intrinsic transcriptional states from lineage-driven effects. We introduce DeepTracing, a deep generative framework that integrates disentangled representation learning with lineage-aware Gaussian processes to explicitly separate intrinsic cellular variation from lineage constraints. The model constructs a layered latent space and enforces independence via Total Correlation regularization, producing intrinsic, lineage, and unified embeddings. Across extensive benchmarks, DeepTracing consistently outperforms existing approaches. In TedSim simulations, it achieves superior clustering of cell states and effectively recovers phylogenetic structure, surpassing original expression and scVI. Applied to mouse tumor lineage-tracing data, DeepTracing attains higher ARI/NMI for tumor-type classification than scVI and PORCELAN, accurately separating primary and metastatic tumors and recovering known trajectories such as early lymph-node divergence and liver-to-kidney cross-seeding. In larger datasets, it maintains strong performance while preserving both transcriptomic continuity and lineage fidelity. DeepTracing also reconstructs continuous developmental trajectories in mouse ventral midbrain, isolating temporal effects from intrinsic differentiation. These results establish DeepTracing as a scalable and interpretable framework for analyzing multimodal single-cell data in tumor progression. Code availabilityThe source code is publicly available at https://github.com/Yuhong-Wen/DeepTracing.

cell biology↗

bioRxiv2026

Harnessing Tumor-Specific Transcript Diversity Uncovers a Shared Neoantigen Reservoir for Pancreatic Ductal Adenocarcinoma

Pancreatic ductal adenocarcinoma (PDAC) is refractory to immunotherapy due to its immunologically cold microenvironment and the scarcity of mutation-derived neoantigens. Here, we introduce NeoAPP, a computational tool designed to systematically decode neoantigens arising from tumor-specific transcripts (TSTs) generated by transcriptional dysregulation. Multi-cohort transcriptomic profiling of 413 PDAC samples using NeoAPP reveal a median of 351 neoantigens per sample derived from 56 neoantigen-encoding TSTs (neoTSTs), surpassing mutation-derived counterparts in both abundance and patient coverage. Mechanistic analyses show that non-canonical splicing junction and transposable element activation drive neoantigen generation, while FOXA2-regulated promoter usage constitutes a potential major source of neoTSTs. Tumor-derived neoTSTs are also detected in extracellular vesicles and cancer-associated fibroblasts, implicating stromal crosstalk in immune modulation. Vaccination with neoTSTs induce CD8+ T cell responses in HLA-A*02:01/A*11:01 transgenic mice and suppressed tumor growth in syngeneic PDAC models. Collectively, this work establishes TST-derived neoantigens as a dominant and therapeutically actionable antigen reservoir in PDAC, advancing a transcriptome-guided framework for neoantigen discovery with potential to overcome immune resistance in low-mutation cancers. Abstract Figure O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=196 SRC="FIGDIR/small/705024v1_ufig1.gif" ALT="Figure 1"> View larger version (71K): org.highwire.dtl.DTLVardef@82ff31org.highwire.dtl.DTLVardef@401342org.highwire.dtl.DTLVardef@b0771corg.highwire.dtl.DTLVardef@15c03ef_HPS_FORMAT_FIGEXP M_FIG C_FIG

cancer biology↗

bioRxiv2026

Virtual Cells Need Context, Not Just Scale

The intersection of AI and biology has entered a phase of explosive growth, driven by the ambition to build "Virtual Cells" or computational models capable of predicting cellular responses to any perturbation. Following the success of structural biology (e.g., AlphaFold) and large language models, the field has converged on training massive, high-capacity models on large-scale single-cell data. This position paper argues that scaling model capacity is insufficient to solve the Virtual Cell problem because the primary failure mode is a lack of adequate coverage over diverse biological contexts, not insufficient model expressivity. We support this claim by reviewing recent studies showing that simple baselines perform on par with sophisticated architectures within a given biological context, and current models fail to consistently generalize across contexts. We connect this finding to the causal inference literature on transportability and contrast it with domains where scaling has succeeded. We substantiate our argument through analysis of a state-of-the-art model on a 22-million-cell immunology dataset. We conclude that the community faces a causal transport problem that cannot be solved by accumulating more data from the same distributions. Instead, we argue that contextual diversity and causal representation learning deserve increased emphasis, complementing ongoing scaling of model capacity and data volume.

genomics↗