PLOS Computational Biology: New Articles

Multiscale modeling of T cell exhaustion: A mathematical framework integrating continuous dynamics with spatial heterogeneity

2026-08-26T14:00:00Z

by Chenghang Li, Yuhong Zhang, Xue Liu, Yipu Qu, Xiulan Lai, Jinzhi Lei

Continuous antigen exposure drives T cells into a progressive state of dysfunction known as exhaustion, enabling tumors to evade immune surveillance and promoting disease progression. Despite its importance, predictive modeling of T cell exhaustion remains a major challenge due to the complexity of its regulatory dynamics. To address this challenge, we developed a mathematical framework that characterizes the dynamic regulation of T cell exhaustion and its impact on tumor-immune interactions. Here, we integrate multi-source data, population dynamics modeling, and agent-based modeling to track the progressive stages of CD8+ T cell exhaustion. Our model demonstrates that immune checkpoint blockade significantly delays exhaustion and promotes the expansion of tumor-reactive T cells compared to untreated conditions. From a pseudo-potential energy perspective, we show that the core mechanism of immunotherapy lies in expanding the tumor-reactive T cell pool, which consequently reduces the overall state of exhaustion within the system. We find that T cell activation and exhaustion signals jointly govern tumor-immune dynamics. Enhancing activation alone without restricting exhaustion can inadvertently accelerate the loss of T cell function. In contrast, combining enhanced activation (via anti-CTLA-4) with suppressed exhaustion (via anti-PD-1) is essential for achieving a sustained antitumor response. Furthermore, spatial simulations confirm that a high-activation and low-exhaustion state effectively restricts tumor spread, maintaining substantially lower tumor densities compared to low-activation, high-exhaustion scenarios. Our framework provides quantitative insights into T cell exhaustion and a theoretical foundation for optimizing combination immunotherapies.

Simulation and inference methods for non-Markovian stochastic reaction networks

2026-08-26T14:00:00Z

by Thomas P. Steele, David J. Warne

Stochastic models of reaction networks are widely used to capture intrinsic noise in complex systems in the life sciences. Typical formulations of these models are based on Markov processes for which there is extensive research on efficient simulation and inference. However, there are complex processes in biology, such as gene transcription and translation, that introduce history dependent dynamics requiring non-Markovian processes to accurately capture the stochastic dynamics of the system. This greater realism comes with additional computational challenges for simulation and parameter inference. We develop efficient stochastic simulation algorithms for well-mixed non-Markovian stochastic reaction networks with stochastic delays that depend on system state and time. Our methods generalize the next reaction method and τ-leaping method to support arbitrary inter-event time distributions while preserving computational scalability. We also introduce a coupling scheme to generate exact non-Markovian sample paths that are positively correlated to an approximate non-Markovian τ-leaping sample path. This enables substantial computational gains for simulation and Bayesian inference through multilevel Monte Carlo and multifidelity schemes. We demonstrate the effectiveness of our approach using several non-Markovian examples, showing substantial gains in both simulation accuracy and inference efficiency. These results extend the practical applicability of non-Markovian models in systems biology and beyond.

The complex swarming dynamics of malaria mosquitoes emerge from simple minimally-interactive behavioral rules

2026-08-26T14:00:00Z

by Antoine Cribellier, Bèwadéyir Serge Poda, Roch K. Dabiré, Abdoulaye Diabaté, Olivier Roux, Florian T. Muijres

Swarming is a widespread collective behavior in animals, often thought to emerge from complex interactions among individuals. Here, we show that mating swarms of malaria mosquitoes can emerge from simple behavioral rules with minimal interaction between individuals. We analyzed two published experimental datasets with three-dimensional flight tracks of Anopheles coluzzii mosquitoes swarming in the lab under simulated sunset conditions and above a visual marker. We found that individuals alternate between straight flight and rapid turning maneuvers known as saccades. These saccades are triggered at the edge of the swarm, and their directions are biased along the sunset direction. This behavior was found consistent between both laboratory datasets, including swarming of solitary individuals; this indicates that inter‑individual interactions may play a limited role within the small-to-medium sized swarms that we studied. We developed a simple agent-based model incorporating three behavioral rules: attraction to the swarm center, alignment normal to the sunset horizon, both driven by environmental cues, and repulsion through short-range collision avoidance. This parsimonious model reproduced key features of the observed laboratory mosquito swarms, including looping flight paths, central density peaks, and directional alignment. Even in the absence of direct inter‑individual interactions, the model generated realistic emergent swarm dynamics, suggesting that social coordination is not required for swarm emergence. Although our lab-based findings remain to be confirmed across more naturalistic conditions, they suggest that the most parsimonious hypothesis behind male mosquito swarming is that it is primarily guided by environmental perception rather than social interaction. This minimal framework offers a new perspective on insect swarming and may apply broadly to the many other insect species that form mating swarms. Understanding the behavioral rules of mosquito swarming could inform strategies for vector control and improve the effectiveness of interventions targeting mosquito reproduction.

Modeling the influences of non-local connectomic projections on geometrically constrained cortical dynamics

2026-08-26T14:00:00Z

by Rishikesan Maran, Eli J. Müller, Ben D. Fulcher

The function and dynamics of the cortex are fundamentally shaped by the specific wiring configurations of its constituent axonal fibers, also known as the connectome. However, many dynamical properties of macroscale cortical activity are well captured by instead describing the activity as propagating waves across the cortical surface, constrained only by the surface’s two-dimensional geometry. It thus remains an open question why the local geometry of the cortex can successfully capture macroscale cortical dynamics, despite neglecting the specificity of Fast-conducting, Non-local Projections (FNPs) which are known to mediate the rapid and non-local propagation of activity between remote neural populations. Here we address this question by conducting a range of investigations using a mathematical model of macroscale cortical activity, in which cortical populations interact both by a continuous sheet and by an additional set of FNPs wired independently of the sheet’s geometry. By simulating the model across a range of external inputs, timescales, and idealized connectome topologies, we demonstrate that the addition of FNPs strongly shape the model dynamics of rapid, stimulus-evoked responses on fine millisecond timescales, but contribute relatively little to slower, spontaneous fluctuations over longer order-of-seconds timescales, which increasingly resemble geometrically constrained dynamics without FNPs. Our results suggest that the discrepant views regarding the relative contributions of local (geometric) and non-local (connectomic) cortico-cortical interactions are context-dependent: While FNPs specified by the connectome are needed to capture rapid communication between specific distant populations (as per the rapid processing of sensory inputs), they play a relatively minor role in shaping slower spontaneous fluctuations (as per resting-state functional magnetic resonance imaging).

Manifold-constrained plasticity enables stable learning in recurrent neural circuits

2026-08-25T14:00:00Z

by Camille Godin, Jean-Philippe Thivierge

The activity of large neuronal populations is often confined to low-dimensional manifolds that can drift over time, posing a challenge for learning rules that assume stable, full-rank representations. Here, we introduce SPLiT (Synaptic Projection Learning with intrinsic Tracking), a synaptic plasticity rule for recurrent neural networks that combines unsupervised manifold tracking with supervised learning. SPLiT uses an online Oja rule to continuously estimate the intrinsic low-dimensional activity subspace and performs normalized least-mean-squares learning in manifold coordinates. In this way, SPLiT keeps synaptic weights aligned with evolving population dynamics. Using a rate-based recurrent network, we show that SPLiT reliably learns time-varying target signals under both constrained dynamics, where activity is restricted to a fixed low-dimensional manifold, and unconstrained dynamics exhibiting changes in the dominant activity subspace. We show that manifold-constrained learning with SPLiT yields faster convergence, less sensitivity to recurrent gain, smaller weight updates, and is more robust to noisy teaching signals. Analytical results show that SPLiT learns the optimal decoder projected onto the instantaneous principal subspace and maintains bounded error under drift. Together, these findings provide a mechanistic account of how synaptic plasticity can leverage the structure of neural manifolds to enable stable and efficient supervised learning despite representational drift.

The multi-omics fallacy in microbiome science

2026-08-25T14:00:00Z

by Rebecca Lewandowski

Artificial intelligence and machine-learning-assisted multi-omics have expanded the scale and ambition of microbiome research, but they have also sharpened an older interpretive problem. Biologically plausible structure is too easily mistaken for biological explanation. This Perspective defines the multi-omics fallacy as claim inflation that occurs when integrating microbial, host, environmental, spatial, and clinical data is assumed to move interpretation from association toward verified mechanism without a corresponding gain in measurement, localization, temporal resolution, functional linkage, or perturbation. The risk is not that computational integration lacks value. It is that predictions, imputations, inferred pathways, feature attributions, and cross-layer networks can acquire mechanistic authority before their biological status has been established. In microbiome science, where stool readouts, taxonomic abundance, inferred function, and predicted metabolites often serve as proxies for host-microbial interaction, this slippage can make uncertain claims appear more complete than the evidence allows. Computational confidence can amplify structured artifact when systematic error becomes learnable. This Perspective proposes a model-to-mechanism burden of proof that distinguishes prediction from explanation, imputation from observation, attribution from causality, cross-layer coherence from mechanism, and diagnostic performance from biological validity. This framework is intended to strengthen, not constrain, computational microbiome science by clarifying which outputs support classification or hypothesis generation and which require direct measurement, localization, temporal analysis, functional validation, or perturbation. Used this way, computational models can help expose uncertainty, prioritize experiments, identify fragile claims, and sharpen biological questions. The result would be a more powerful form of computational microbiome science, one in which models do not stand in for mechanisms but guide the work needed to earn them.

Theory and evidence of amplitude control by frequency detuning in a coupled neuronal oscillator system

2026-08-25T14:00:00Z

by Adam C. Lu, Seyed AmirHossein Ourang, Jeffrey D. Moore

Neuronal oscillator circuits that generate rhythmic movements must operate flexibly and reliably to produce the varied motor patterns that animals exhibit naturally. Rodents rhythmically “whisk” their vibrissae for haptic perception, and they dynamically adjust the whisking range to serve different perceptual goals. Whisking is controlled by a brainstem oscillator circuit that is coupled to breathing, yet how whisking amplitude is modulated remains unknown. Here we propose and evaluate an amplitude control mechanism based on principles of synchronization in coupled oscillators. Specifically, a re-analysis of rat behavioral data demonstrates that whisking exhibits kinematic signatures and phase dynamics of amplification via entrainment with “sniffing”, a mode of high-frequency breathing. A neuronal network model of the whisking oscillator circuit suggests that whisking amplitude can be modulated by shifting the oscillator’s intrinsic frequency relative to the sniffing frequency, analogous to the engineering technique of “detuning”. Based on these results, we propose that detuning between coupled neuronal oscillators may represent a general computational strategy for gain control in nervous systems.

High reelin expression may explain why a subgroup of entorhinal cortex neurons functions as an initial nucleation site of Alzheimer’s disease

2026-08-25T14:00:00Z

by Asgeir Kobro-Flatmoen, Jagir R. Hussan, Peter J. Hunter, Stig W. Omholt

The entorhinal cortex (EC) plays a crucial role in memory functions. Long before the clinical symptoms of Alzheimer’s disease (AD) emerge, it has already undergone significant degeneration, making it a primary site for the onset of the disease. The reasons for this remain elusive. Layer II (LII) neurons of the anterolateral EC are especially prone to display a very early increase in intracellular amounts of amyloid-β peptide (Aβ) and hyperphosphorylated tau protein (p-tau). The expression of the large glycoprotein reelin is extraordinarily high in ECLII neurons compared to most other cortical neurons and proximity ligation assay data and other immunohistochemical data strongly support the notion that reelin binds to Aβ in these neurons. Here, based on the premise that reelin may function as a sink for intracellular Aβ, we show by computational modeling that, in a senescent physiology predisposing to frequent inflammation-driven Aβ42 production bursts over a decades-long period, the intracellular amount of Aβ42-reelin complexes can accumulate to extraordinarily high levels in anterolaterally positioned LII neurons compared to the vast majority of cortical neurons. This is consistent with experimental data showing that intracellular accumulations of Aβ42 positive material ranged from 20 to 80% of total soma volume in EC neurons from patients with idiopathic AD. Based on known tau protein biology, we also show that this extreme intracellular aggregation that overloads the lysosomal degradation machinery, manifesting chronic homeostatic dysregulation, can lead to the production of p-tau fragments prone to aggregation. Together, our findings may contribute to the resolution of why the EC is so strongly associated with the very early etiology of AD.

Double shrinkage transfer causal learning: An application to alzheimer’s disease

2026-08-24T14:00:00Z

by Yunxin Shi, Lulu Pan, Yu Hu, Yongfu Yu, Guoyou Qin

Amyloid-Beta 42 (ABeta42) is a key biomarker of cerebral amyloidosis in Alzheimer’s disease, and estimating its causal effect on subsequent cognitive outcomes is important for understanding disease progression in racial minority populations, such as African Americans. However, causal effect analyses in such populations are often challenged by limited sample sizes, leading to biased or unstable estimates when only target-population data are used. Existing transfer causal learning methods borrow information from a larger related source domain, but may be unstable when source-domain estimation is high-dimensional or when irrelevant covariates introduce noise into transferred nuisance models. We propose a double shrinkage transfer causal learning (DSTCL) estimator that regularizes both the initial source domain estimation and the source-target parameter differences, and then estimates causal effects using a doubly robust framework. We evaluated DSTCL in simulations under varying source sample sizes, inter-domain similarities, covariate dimensions, and other settings, and then applied it to the Alzheimer’s Disease Neuroimaging Initiative dataset using African American participants as the target domain and non-Hispanic White participants as the source domain. In simulations, DSTCL achieved small bias, low mean squared error, and stable confidence interval coverage, generally outperforming competing methods. In the ADNI application, lower baseline ABeta42, reflecting greater amyloid burden, was associated with worse 12-month cognitive performance among African American participants, with DSTCL producing the narrowest confidence interval among the methods compared. These findings suggest that DSTCL provides a practical framework for transfer causal inference in biomedical studies with limited target population.

Mechanochemical modeling of exercise-induced skeletal muscle hypertrophy

2026-08-24T14:00:00Z

by Ingvild S. Devold, Marie E. Rognes, Padmini Rangamani

Skeletal muscle displays remarkable plasticity, adapting its size and strength in response to mechanical loading, especially, from exercise. This process, known as hypertrophy, is fundamental to athletic training and rehabilitation, but is challenging to quantitatively predict due to its multifactorial, multiscale nature. Specifically, skeletal muscle hypertrophy results from an integration of macroscopic mechanical stimuli with the intracellular signaling pathways that govern muscle growth. In this work, we present a multiscale computational model that mechanistically integrates these mechanical and biochemical stimuli and offers a framework for predicting the outcomes of different types of exercise on skeletal muscle growth. The framework couples a transversely isotropic hyperelastic model for tissue-level mechanics with a system of ordinary differential equations representing the IGF1-AKT-mTOR-FOXO signaling pathway, a key regulator of protein synthesis and degradation. We link these scales using a volumetric growth model, where the signaling dynamics inform a growth tensor that drives changes in muscle cross-sectional area. This approach enables the simulation of long-term muscle adaptation, providing a mechanistic tool to investigate how different exercise protocols lead to macroscopic hypertrophy. Simulations from our model capture the temporal dynamics of hypertrophy under varying load protocols and highlight how feedback between protein synthesis and muscle growth regulates the dose-response relationship to prevent unbounded growth. Using muscle geometries derived from the Visible Human dataset, we study how human variations in muscle geometry affect hypertrophy. Finally, we demonstrate that the mechanochemical coupling between muscle geometry and signaling not only predicts macroscopic shape changes but also provides buffering from local signaling heterogeneity. Ultimately, this framework offers a predictive computational tool for optimizing training regimens and understanding the multiscale determinants of muscle adaptations.

How host mobility formulations shape estimates of pathogen dispersal and epidemic risk in non endemic regions

2026-08-24T14:00:00Z

by Charley Presigny, Piero Poletti, Stefano Merler, Manlio De Domenico

The global warming effects of climate change favor the stable residence of invasive vectors, such as the mosquito Aedes albopictus, in temperate areas, like the Mediterranean basin or north America. Thus, favorable weather conditions, together with human mobility, increase the likelihood of large-scale vector-borne disease epidemics in areas where they were historically not endemic. While mathematical modeling acknowledges the relevance of human mobility in the persistence or prevalence of vector-borne diseases in endemic areas, the combined role of host-mediated vector transport and long-range airline travel in the emergence of epidemics in non-endemic areas is poorly understood. In this context, the impact of specific modeling formulations of human mobility on the epidemic predictions remains elusive. To bridge this gap, we compare two alternative models - one based on the force of infection from different source locations and one based on the explicit physical diffusion of individuals - that incorporate intra- and inter-country human mobility (considering air traffic flows), vector dispersal, and human-mediated vector mobility (e.g., vectors in cars/trains), with applications to diseases such as chikungunya, dengue and Zika. Through extensive computational and analytical analysis, we assess how implicit assumptions and the inclusion of different mobility components influence model estimates of epidemic emergence risk, the timing of potential epidemics and the risk of epidemic spread. The comparison between the two frameworks is illustrated using synthetic networks and Italy as a realistic case study. Our findings show that different model formulations of human mobility yield divergent risk assessment, affecting predictions of large scale outbreaks and importation times driven by explained mechanistic difference. Despite these differences, both models show consistent qualitative patterns in the spatial risk and the likelihood of concurrent local outbreaks. Integrating multiple mobility components and carefully identifying assumptions that reflect observed mechanisms are essential for understanding epidemics in currently non-endemic areas.

The perils of omitting omissions when modeling evidence accumulation

2026-08-21T14:00:00Z

by Xiamin Leng, Alexander Fengler, Amitai Shenhav, Michael J. Frank

Response deadlines are commonly imposed in decision-making research to incentivize speedy decisions and sustained attention. This procedure frequently leads to a proportion of trials in which no response is made, and these omissions are often simply removed from the data during analysis. Here we show that this seemingly trivial assumption is in fact quite consequential for parameter estimation. We propose that omissions should instead be treated as observations to inform inference of the underlying generative process that led to their occurrence. Using new tools from likelihood-free inference applicable to a broad class of sequential sampling models (SSMs), we enable fast computation of omission probability without explicit integration, and clarify the degree to which omitting omissions – even in seemingly benign settings – can lead researchers astray. We explore this phenomenon in the setting of SSMs with constant and time-varying boundaries, and show that parameter recovery is improved by incorporating a model of omission probability. We show that the benefits of modeling omission probability are distinct from benefits gained from past approaches of modeling attentional lapses to account for these omissions. Our findings shine a light on the consequences of omitting omission in modeling choice behavior with SSMs, and demonstrate how joint modeling of observed and omitted data can improve parameter inference and therefore the reliability of downstream scientific conclusions.

An in silico framework for dissecting the mechanistic origins of in vivo recorded neuronal activity

2026-08-21T14:00:00Z

by Bjorge Meulemeester, Arco Bast, María Royo, Rieke Fruengel, Su Saka, Foivos Kastrinakis, Marcel Oberlaender

How can we identify the mechanistic origins of the electrophysiological activity that is recorded from neurons in the living brain? A promising strategy for addressing this question is to generate biologically realistic models of in vivo recorded neurons, and simulate how they transform synaptic inputs from the network into their observed neuronal activity. For this purpose, we here provide our approaches for the generation, simulation, and analysis of network-embedded neuron models as an open source, fully documented and freely available software environment: In Silico Framework (ISF). ISF is centered around the concept of achieving “model consensus” about the mechanistic origins of in vivo recorded activity across biologically diverse sets of models. To achieve such model consensus, ISF offers three key workflows. First, ISF enables users to generate models that are equally well constrained by empirical data at subcellular, cellular and network scales, while the set of models as a whole is constructed to exhibit maximally diverse parameters, spanning the full ranges permitted by the empirically observed biological variability at each scale. Second, ISF enables users to identify those subsets of model configurations that predict the in vivo observations without being tuned to do so. Third, for each of those model configurations, ISF enables users to identify which mechanisms at subcellular, cellular and network scales are necessary to predict the in vivo observations, and which mechanisms are dispensable. Thereby, ISF can reveal which mechanisms are common across model configurations, and whether the diversity of model configurations could account for the variability of the in vivo observed activity across animals, cells and trials. In essence, by achieving such model consensus, ISF predicts mechanisms that are robust across biological variability, and which may hence indeed be used in vivo. Finally, ISF enables users to derive model consensus for in silico manipulations, to identify which experimental strategies would be best suited to test the predicted mechanisms in vivo. We exemplify how we have used this iterative in silico - in vivo approach of ISF to dissect the mechanistic origins of sensory responses in the barrel cortex. By making ISF available as a standalone online resource, we believe it will facilitate the generation, simulation and analysis of models that reveal mechanistic origins of in vivo recorded activity beyond the barrel cortex for which it was originally designed.

Non-Markovian dynamics and effective reproduction number in COVID-19: Evidence from Cyprus contact tracing data

2026-08-21T14:00:00Z

by Pavlos Alexandros Dimitriou, Matteo D’Alessandro, Brian L. Chang, Valentinos Silvestros, Elisavet Constantinou, Costas Pitris, Panayiotis Kolios, Piet Van Mieghem

Using contact tracing data provided by the Cyprus Ministry of Health, infection trees for the first four waves of the COVID-19 epidemic are constructed. In these trees, nodes represent infected individuals, while links indicate the direction of transmission between them. For each infection tree of N nodes, the hopcount distribution from the root node to all other nodes is calculated. The empirical distribution is then compared to the hopcount distribution of infection trees generated by a non-Markovian SI process on a complete graph, with Weibull infection times characterized by a shape parameter α. We compute the values of the shape parameter α that best fit the empirical distribution and find that only values of α>1 are obtained, while the Markovian case is characterized by α=1. A Weibull distributed infection time with shape parameter α>1 is characterized by a unimodal density function with a peak at finite time, consistent with previous findings in the literature. Our analysis therefore suggests that the spreading process is most likely governed by non-Markovian dynamics, and that non-Markovianity can be detected solely from the topology of the infection trees. Finally, we analyze the evolution of the empirical distribution of the number of secondary infections caused by each node in the infection trees across different time windows to estimate the effective reproduction number. In practice, the average number of secondary infections seems to often provide a lower bound of the reproduction number computed by the Cyprus Ministry of Health. When the last level of the trees, composed predominantly of terminal nodes that do not generate further infections, is excluded, the estimate reflects more accurately the dynamics of the epidemic.

Eleven quick tips for Biomedical Federated Learning

2026-08-21T14:00:00Z

by Kyle Ellrott, Venkat S. Malladi, Jean-Christophe Bélisle-Pipon, Emek Demir, Yael Bensoussan, Serghei Mangul, Alex A. T. Bui, Paul C. Boutros

Modern statistical and machine learning techniques are effective at describing, testing hypotheses and making predictions from complex data. This effectiveness is strongly influenced by the volume and heterogeneity of available data. In many fields, including much of biomedicine, large centralized datasets are not available because of cost, privacy, regulatory or other restrictions. In these cases, smaller datasets are distributed across a large number of independent sites. Medical record data is a classic example of this challenge: the total number of patients may be large, but their records are distributed across many health systems and cannot easily be centralized. Federated learning (FL) is a machine learning paradigm that enables training and validation of a shared model in settings of decentralized data. FL can improve model accuracy and generalizability by increasing sample size, but has trade-offs ranging from operational complexity to data-privacy risks to the potential to introduce unexpected imbalances in model accuracy. We outline ten tips for successfully and sustainably implementing FL for Biomedical applications, ensuring both ethical data governance and improved model performance in sensitive domains.

Ten quick tips for causal analysis of biomedical omics data

2026-08-20T14:00:00Z

by Gleb Svinin, Rebecca Ting Jiin Loo, Nikhilesh Vasantha Kumar, Varsha Venkatesha Murthy, Dilara Uzuner Odongo, Ramón Díaz-Uriarte, Ana Conesa, Gianluca Bontempi, Susana Vinga, Ilaria Granata, Daniel Domingo-Fernández, Paola Lecca, Marieke L. Kuijjer, Jesse H. Krijthe, Ina Koch, Laurence Calzone, Simona Ester Rombo, Fátima Sánchez-Cabo, Enrico Glaab

A portable recalibration workflow for reference-based variant calling in non-human genomes

2026-08-20T14:00:00Z

by Hyeonjung Lee, Sunhee Kim, Michelle Audrelia Sunartha, Chang-Yong Lee, Young-suk Lee

A key computational step in reference-based variant calling is distinguishing true genetic variants from sequencing errors. Advanced tools and workflows have been developed to handle this by computational modelling of technical errors from the sequencing machines. However, these recalibration workflows have largely been evaluated for human data only and its exact applicability for non-human data remains unknown. Here, we conducted a systematic evaluation of variant calling on human, rice, sheep, and chickpea data, and found that existing workflows introduce unexpected statistical bias, thus leading to suboptimal variant calls for non-human data. To address this problem, we present simple guidelines for constructing a “pseudo-”database (pseudoDB) of genetic variants as a scalable and portable solution for recalibration and variant calling. With human data, our pseudoDB-based workflow performs comparably to existing dbSNP-based GATK3 workflows and those using DeepVariant, Strelka2, and FreeBayes. We extend this to other non-human genomes, namely cattle, brown bear, swan goose, African oil palm, Komodo dragon, and stevia, altogether resulting in the identification of up to 242.0% unique genetic variants. The majority of newly identified variants are within the non-coding regions, hinting at the rich diversity of genome regulation in the non-human population. Our pseudoDB-based workflow is agnostic to reference genomes and modular for easy integration with other computational workflows for human and non-human resequencing data.

CLDN18.2 antibody design with protein language models: A deep learning optimization framework

2026-08-20T14:00:00Z

by Tao Qu, Lingyan Yuan, Weiran Cui, Jiatian Tang, Zhitong Bing, Xianghong Xu, Jizheng Duan, Qiong Yang, Hui Cai

CLDN18.2 is a promising tumor-specific antigen; however, the development of therapeutic antibodies against it is challenged by the need for simultaneous optimization of affinity and developability. To address this, we present cdrGPT, a deep learning framework based on GPT-2 for de novo generation of complementarity-determining region H3 (CDRH3) sequences. Our approach integrates pre-training on the Observed Antibody Space (OAS) database with structural templating derived from the known antibody zolbetuximab. Generated sequences were iteratively refined through rejection sampling and fine-tuned against a multi-parameter objective function encompassing predicted affinity and MHC class II binding risk. From an initial set of 50,000 sequences, this screening pipeline yielded 313 high-confidence candidates. Subsequent analysis using evolutionary scale modeling 2 (ESM2) embeddings, principal component analysis (PCA), and clustering revealed three structurally distinct clusters, with intra-cluster cosine similarities exceeding 0.99. Validation of seven representative sequences from the dominant cluster using AlphaFold3 confirmed high structural fidelity to the zolbetuximab template, demonstrating a root mean square deviation (RMSD) of 1.331 Å for the CDRH3 loop and positional deviations of less than 0.4 Å for key paratope residues. These results indicate that the designed variants preserve the core binding mode of the parent antibody. This study establishes a feasible pipeline for integrating AI-generated CDRH3 loops into functional antibody scaffolds, providing a foundation for the accelerated development of therapeutics targeting CLDN18.2 and other clinically relevant antigens.

Enzyme kinetics shapes the growth response of metabolic networks

2026-08-19T14:00:00Z

by Leon Seeger, Fernanda Pinheiro, Michael Lässig

Microbes adjust their metabolism to environmental challenges by changing protein expression levels, metabolite concentrations, and reaction rates. Average expression levels in large proteome sectors change coherently, while individual proteins show divergent shifts even within the same pathway. Here, we establish a metabolic model that integrates local enzyme kinetics and global network architecture to predict the joint growth response of proteins and metabolites. Under nutrient limitation, we predict a remarkably simple pattern of proteome reallocation with growth rate: protein expression levels change linearly but heterogeneously. For a given enzyme, the direction of change is determined by its local kinetic constants – catalytic rate and substrate affinity – and by the degree of nutrient restriction affecting its embedding pathway. This double-graded growth response of the proteome is mediated by restriction-dependent metabolite levels, which are predicted to decrease with growth rate in a nonlinear way. The model establishes three specific growth laws: protein expression changes of individual enzymes are negatively correlated with their expression and with their substrate saturation at high growth; average changes of pathways and larger functional sectors are correlated with their internal variance. These predictions are in quantitative agreement with measured system-wide proteomics and metabolomics data of E. coli. Enzyme-specific response patterns are a starting point for model-guided interventions into bacterial metabolism.

The SATvac model of CD8⁺ T cell expansion and contraction phases considering memory and effector cell differentiation

2026-08-18T14:00:00Z

by Seyedeh Fatemeh Seyyedizadeh, David A. Christian, Thomas A. Adams II

Primary immune responses induce CD8⁺ T cell responses characterized by activation via antigen presenting cells, expansion, differentiation into effector and memory phenotypes, and contraction resulting in long-term memory populations. In this study a mathematical stochastic agent-based model is developed to simulate all phases of the CD8⁺ T cell response following vaccination. Importantly, the model successfully captures the stochastic nature of T cell dynamics throughout the response. It predicts T cell population with high accuracy while addressing mouse-to-mouse variability, highlighting its robust predictive power. This predictive model aims to improve T cell vaccination strategies by both informing the biology of the T cell response and streamlining vaccine development.

Reassessing adult surfactant replacement therapy with mechanics-informed reinforcement learning

2026-08-18T14:00:00Z

by Philippe Meliga, Gregor Roncin, Alejandro Yepes Peñaranda, Elie Hachem

Surfactant replacement therapy (SRT) remains clinically limited to neonatal applications, in part because the mechanical feasibility of achieving efficient delivery in adult lungs is poorly understood. Previous computational studies have largely been descriptive or based on parameter sweeps, providing limited guidance on how to design efficient adult protocols under anatomical constraints. Here, we introduce a computational framework that integrates mechanistic modeling of surfactant propagation in anatomically motivated airway trees with deep reinforcement learning (DRL) to identify efficient delivery strategies across scales and airway geometries. The approach leverages a reduced-order model of plug transport and redistribution that captures the key mechanics of surfactant coating in complex airway networks while remaining lightweight enough for large-scale optimization. A custom DRL agent autonomously explores delivery parameters—including aliquot volume, flow rate, patient posture, and surfactant properties—to optimize protocol-level performance across diverse anatomical and physiological conditions. Under the branch-level coverage-based reward adopted here, systematic optimization of delivery parameters improves distal delivery at both pediatric and adult scales, while clarifying how these gains depend on prescribed volume, airway asymmetry, and control complexity. Increasing the number of aliquots improves access to distal regions and, in favorable geometries, shifts the onset of high-coverage regimes toward lower prescribed volumes, whereas posture becomes especially informative in asymmetric trees. Extending the control space to include surfactant rheology provides an additional lever, particularly in constrained adult settings, but does not overcome the structural limitations imposed by strong geometric asymmetry. Overall, these results establish a physics-based framework for AI-assisted optimization of intrapulmonary liquid delivery and clarify the respective roles of dose partitioning, posture, and formulation tuning. They also show that the interpretation of delivery success depends strongly on the evaluation metric: branch-level coverage provides a functionally oriented measure across heterogeneous airway trees, whereas stricter homogeneity metrics yield substantially more pessimistic assessments, especially in adult asymmetric geometries. These findings do not predict clinical efficacy; rather, they provide a controlled mechanistic feasibility benchmark and testable design hypotheses within physiologically realistic bounds.

Cerebellum-inspired neural network of supervised learning with tensor-based sparse coding for multi-class classification

2026-08-18T14:00:00Z

by Runguang Zhou, Douglas Zhou, Songting Li, Xiaoyu Chen

Under the Marr-Ito-Albus framework, the cerebellum performs supervised learning in Purkinje cells upon the unsupervised sparse representations generated within granule cells, contributing fundamentally to associative learning in motor control. However, the specific mechanisms through which cerebellar circuitry and plasticity rules enable supervised learning, and properties of the sparse coding induced by cerebellar architectural constraints, remain poorly characterized. To address this, we first established a sparse coding mechanism inspired by anatomical and physiological properties of the granular layer, including input-sharing connectivity from mossy fibers, Golgi-cell-mediated localized feedback inhibition for winner-take-all sparsification, and activity-dependent bias adjustments. This sparse coding, implemented via efficient tensor-based computations, preserves local neighborhood structures as revealed in a geometric interpretation. Furthermore, we demonstrated that error signals transmitted via climbing fibers to Purkinje cells, integrated with intrinsic plasticity, drive efficient learning of the sparse representations for multi-class classification, leading to accurate population coding for judgments in the cerebellar nuclei. The resultant cerebellum-inspired neural network model achieved a test accuracy of 90.50±0.10% on Fashion-MNIST, performance comparable to a backpropagation-trained single-hidden-layer feedforward neural network, while providing more than 5× computational acceleration. The model further exhibited proof-of-principle closed-loop control capability in simplified motor tasks, including balancing a cart-attached pole and controlling a robotic arm to reach targets. Our study delineates cerebellum-inspired sparse coding and supervised learning mechanisms, and demonstrates robust performance of the cerebellum-inspired neural network across multiple task domains. These results suggest that cerebellum-inspired architectures would provide useful design principles for lightweight neural networks in resource-limited and latency-critical applications.

ERFMTDA: Predicting tsRNA–disease associations using an enhanced rotative factorization machine

2026-08-18T14:00:00Z

by Wei Lan, Dong Wang, Wenyi Chen, Xuhua Yan, Qingfeng Chen, Shirui Pan, Yi Pan

tRNA-derived small RNAs (tsRNAs) have emerged as a novel class of regulatory molecules implicated in the pathogenesis of numerous human diseases, positioning them as promising biomarkers and therapeutic targets. Existing computational methods provide a cost-effective alternative to experimental method, but they tend to ignore biological attributes and complex feature interactions. To overcome these limitations, we propose ERFMTDA, an enhanced rotative factorization machine framework for predicting potential tsRNA-disease associations. ERFMTDA explicitly models complex interactions among heterogeneous biological features while integrating latent structural representations derived from the global association matrix. In addition, a biologically informed negative sampling strategy based on motif-level sequence similarity is introduced to improve the reliability of negative samples. Extensive experiments demonstrate that ERFMTDA consistently surpasses the other eleven state-of-the-art methods. Two case studies on diabetic retinopathy and hepatocellular carcinoma further corroborate the model’s ability to prioritize biologically meaningful tsRNA–disease associations.

Can intrinsic loop energetics predict G-Quadruplex topology?

2026-08-18T14:00:00Z

by Michał Jurkowski, Mateusz Kogut, Mikołaj Ławicki, Jacek Czub

Naturally occurring DNA G-quadruplexes (G4s) regulate many cellular processes, such as gene expression and replication, whereas designed G4s serve as building blocks for controllable nanodevices. Composed of four G-tracts and three loops, G4s exhibit high structural polymorphism, mainly associated with different geometries adopted by the loops. Understanding the biological role of G4s and facilitating their targeted design require detailed knowledge of their folding preferences, which likely originate from the stability of individual loops. However, sequence-dependent preferences for loop geometries and their effects on the overall G4 conformational landscape remain unclear. Here, we used molecular dynamics simulations to systematically evaluate folding free energies of all five standard G4 loop geometries across four loop lengths. Our results reveal that loop-geometry preferences strongly depend on the loop spatial span and, with increasing length, shift from low-span toward large-span geometries. Crucially, we discovered that the overall G4 fold is primarily dictated by internal loop stabilities in three-tetrad G4s but not in two-tetrad ones, where inter-loop interactions emerge as vital stability determinants. Moreover, internal loop stabilities explain the exceptional structural diversity of G4s with three-nucleotide loops reported by experimental studies. Finally, we show that loop-geometry preferences arise from an interplay between the electrostatic repulsion of phosphate groups and loop overstretching.

Contrastive learning to fine-tune feature extraction models for the visual cortex

2026-08-17T14:00:00Z

by Alex Mulrooney, Zhi Li, Austin J. Brockmeier

Predicting the neural response to natural images in the visual cortex requires extracting relevant features from the images and relating those feature to the observed responses. In this work, we optimize the feature extraction in order to maximize the information shared between the image features and the neural response across voxels in a given region of interest (ROI) extracted from the BOLD signal measured by functional magnetic resonance imaging (fMRI). We adapt contrastive learning (CL) to fine-tune a convolutional neural network, which was pretrained for image classification, such that a mapping of a given image’s features are more similar to the corresponding fMRI response than to the responses to other images. We exploit the Natural Scenes Dataset as organized for the Algonauts Project, which contains the high-resolution fMRI responses of eight subjects to tens of thousands of naturalistic images. We show that CL fine-tuning creates feature extraction models that enable higher encoding accuracy in both early and higher visual ROIs as compared to the features from the pretrained network. Quantitatively, the performance is similar to a baseline approach that directly uses a regression loss at the output of the network to tune it for fMRI response encoding. We investigate inter-subject transfer of the CL fine-tuned models, including subjects from the Natural Object Dataset, another lower-resolution dataset with 9 subjects. We also pool subjects for fine-tuning, which further improves encoding performance in early ROIs. Finally, we examine the performance of the fine-tuned models on common image classification tasks, explore the landscape of ROI-specific models by applying dimensionality reduction on the Bhattacharya dissimilarity matrix created using the predictions on those tasks, and show that these landscapes match those based on representational similarity analysis. Finally, we generate images via Stable Diffusion based on vector-space prompts created by aligning the CL-tuned models embeddings for different ROIs, showing that generated images have similar embeddings to the original but that estimates of the intrinsic dimension are lower for generated versus original representations.

IBAS: Interaction-bridged association studies discovering novel genes underlying complex traits

2026-08-17T14:00:00Z

by Dinghao Wang, Pathum Kossinna, Karen Ardila, Senitha Kumarapeli, M. Ethan MacDonald, Jingjing Wu, Qingrun Zhang

Genetic contributions to complex traits are often mediated through coordinated gene–gene interaction networks, yet most existing association frameworks focus on marginal single-gene effects and overlook higher-order dependency structures. Direct modeling of interactions remains challenging due to combinatorial complexity and statistical instability. We introduce Interaction-Bridged Association Study (IBAS), a general framework that incorporates pathway-level interaction patterns into genotype–phenotype association analysis without explicitly enumerating interactions. IBAS leverages transcriptomic reference data to construct low-dimensional representations of pathway activity, which guide SNP-weighting and gene-level association testing within a kernel-based framework. In perturbation-based simulations, IBAS demonstrates improved stability and reproducibility compared to conventional TWAS and gene-based methods, while maintaining well-calibrated Type I error under phenotype permutation. Application to the WTCCC datasets identifies both known and novel genes across multiple complex diseases, including candidates with modest marginal effects missed by standard approaches. These findings are supported by replication in an independent cohort, and analyses across multiple reference tissues revealing both shared and tissue-specific signals. Overall, IBAS provides a statistically robust and computationally tractable framework for incorporating interaction effects into association mapping, extending beyond the single-gene paradigm and enabling more comprehensive characterization of complex trait. IBAS is available on GitHub at: https://github.com/QingrunZhangLab/IBAS

Ten simple rules for effective use of generative AI for code development in environmental science

2026-08-17T14:00:00Z

by Rachel A. King, Laurel Abowd, Carlo W. Broderick, LM Bradley, Max F. Czapanskiy, Mona M. Farnisa, Erica M. Ferrer, Carmen Galaz García, Darian Gill, Nicole M. Greco, Juliette Jacquemont, Justin A. Kadi, Li Kui, Gretchen LeBuhn, Liying Li, Abigail Meyer, Marisa Morse, Evan Patrick, Samantha Shanny-Csik, Nicholas A. C. Tucker, Zhe Wang, Caitlin R. Fong

Reconciling contradictory models of subthalamic nucleus contributions to basal ganglia beta oscillations

2026-08-17T14:00:00Z

by Ka Nap Tse, G. Bard Ermentrout, Jonathan E. Rubin

Recent computational studies of Parkinson’s disease have yielded contradictory findings regarding the role of the subthalamic nucleus (STN) in pathological beta oscillations, with some models implicating STN as essential for beta generation and others suggesting that STN suppresses oscillations. This work addresses these discrepancies by systematically investigating how the specific features of the integrate-and-fire neurons used in these models influence simulated basal ganglia network dynamics. Using both rate models and spiking network simulations incorporating coupled subthalamopallidal and pallidostriatal circuits, we demonstrate that the choice between leaky integrate-and-fire (LIF) and quadratic integrate-and-fire (QIF) models to represent STN neurons fundamentally impacts the phase relationship between STN and external globus pallidus prototypical (Proto) neuron populations. QIF STN neurons establish in-phase coupling with Proto neurons, which enhances beta oscillation amplitude, while LIF STN neurons develop anti-phase relationships, which suppresses beta power. Through intervention experiments and parameter sweeps across physiologically relevant firing rates, we show that these phase-related effects persist robustly across network conditions, and we mathematically establish conditions under which these results are guaranteed to hold. Our findings reveal that the fundamental mathematical structure underlying spike generation, rather than other biophysical details, determines whether the subthalamopallidal loop acts as a beta amplifier or suppressor. This mechanistic insight reconciles contradictory findings in the literature, demonstrates that seemingly minor modeling choices can have profound consequences for understanding disease mechanisms and therapeutic targets, and offers predictions for determining which model framework reflects the biological reality.

The limitations of non-mechanistic methods for characterizing pathogen-pathogen interactions: A simulation study

2026-08-17T14:00:00Z

by Sarah C. Kramer, Sarah Pirikahu, Cana Kussmaul, Lulla Opatowski, Matthieu Domenech de Cellès

Pathogen-pathogen interactions occur when infection with one pathogen influences one’s chance of infection or disease due to another. Increasingly, evidence suggests that interactions are a common feature of infectious disease epidemiology. However, due to both the nonlinearities and stochasticity inherent to infectious disease transmission, and the frequency of confounding (e.g., by shared seasonal forcing), simple, correlative methods for characterizing interactions may be prone to failure. Here, we perform a simulation study to evaluate several more complex non-mechanistic approaches for inferring causality from time series data: generalized additive models (GAMs), Granger causality, transfer entropy, and convergent cross-mapping (CCM). Specifically, we use a two-pathogen mechanistic transmission model, calibrated to produce dynamics resembling outbreaks of influenza and respiratory syncytial virus (RSV), to generate synthetic datasets with a range of values for interaction strength and duration. We then apply each method to all synthetic datasets. We find that Granger causality, transfer entropy, and CCM all fail to consistently infer whether data contain signal of an interaction; in particular, methods tend to incorrectly identify interactions where none are modeled (average sensitivity = 80.6%, 92.1%, 72.1%, respectively; average specificity = 31.0%, 33.3%, 33.1%). Furthermore, we find little to no association between point estimates from each method and true interaction strength. In contrast, GAMs infer the existence of interactions more accurately than the other methods (sensitivity = 85.2%, specificity = 72.5%), and consistently yield larger point estimates for stronger interactions. However, their practical utility is limited by an inability to evaluate interaction asymmetry (i.e., whether the effect of pathogen A on pathogen B is identical to that of B on A). Overall performance patterns were similar when methods were applied to two real-world datasets from Hong Kong and Canada. We conclude that accurately and comprehensively characterizing pathogen-pathogen interactions based on outbreak data remains a significant challenge. For this reason, it is critical that any proposed methods be rigorously evaluated before being used to draw conclusions about interactions.

Structure-aware deep learning enhances m6A prediction and reveals cell type-associated RNA structural signatures

2026-08-14T14:00:00Z

by Mingze Sun, Di Zhang, Zhiyuan Li, Yihan Lin

N6-methyladenosine (m6A), the most abundant mRNA modification in eukaryotes, plays essential roles in gene regulation and disease pathogenesis. Computational prediction of m6A sites offers a scalable alternative to costly experimental approaches, yet current methods rely predominantly on linear sequence features. This overlooks potentially informative RNA structural context, which is associated with local methylation patterns and may provide complementary predictive information beyond linear sequence motifs. To incorporate this complementary information, we propose SMART-m6A (Sequence–structure Multifeature Attention RNA Transformer for m6A), a deep learning framework that integrates sequence and structural information through parallel convolutional feature extraction and structure-guided attention for multifeature fusion. SMART-m6A achieves superior predictive performance compared to existing methods, with particularly clear advantages in sequence-ambiguous candidates. Beyond prediction accuracy, learned attention patterns reveal strong concordance with experimentally validated m6A-binding protein recognition sites and identify potentially novel regulatory motifs. Through systematic ablation studies and targeted structural-input perturbation analyses, we show that sequence and structure provide complementary predictive information, and that sites with greater prediction sensitivity to structural perturbation exhibit distinct local structural profiles between cell lines. Collectively, this work demonstrates the predictive value of sequence-derived structural features in m6A modeling and provides a multifeature deep learning framework for accurate and interpretable structure-aware epitranscriptomic prediction.

PLOS Computational Biology: New Articles

Multiscale modeling of T cell exhaustion: A mathematical framework integrating continuous dynamics with spatial heterogeneity

Simulation and inference methods for non-Markovian stochastic reaction networks

The complex swarming dynamics of malaria mosquitoes emerge from simple minimally-interactive behavioral rules

Modeling the influences of non-local connectomic projections on geometrically constrained cortical dynamics

Manifold-constrained plasticity enables stable learning in recurrent neural circuits

The multi-omics fallacy in microbiome science

Theory and evidence of amplitude control by frequency detuning in a coupled neuronal oscillator system

High reelin expression may explain why a subgroup of entorhinal cortex neurons functions as an initial nucleation site of Alzheimer’s disease

Double shrinkage transfer causal learning: An application to alzheimer’s disease

Mechanochemical modeling of exercise-induced skeletal muscle hypertrophy

How host mobility formulations shape estimates of pathogen dispersal and epidemic risk in non endemic regions

The perils of omitting omissions when modeling evidence accumulation

An in silico framework for dissecting the mechanistic origins of in vivo recorded neuronal activity

Non-Markovian dynamics and effective reproduction number in COVID-19: Evidence from Cyprus contact tracing data

Eleven quick tips for Biomedical Federated Learning

Ten quick tips for causal analysis of biomedical omics data

A portable recalibration workflow for reference-based variant calling in non-human genomes

CLDN18.2 antibody design with protein language models: A deep learning optimization framework

Enzyme kinetics shapes the growth response of metabolic networks

The SATvac model of CD8+ T cell expansion and contraction phases considering memory and effector cell differentiation

Reassessing adult surfactant replacement therapy with mechanics-informed reinforcement learning

Cerebellum-inspired neural network of supervised learning with tensor-based sparse coding for multi-class classification

ERFMTDA: Predicting tsRNA–disease associations using an enhanced rotative factorization machine

Can intrinsic loop energetics predict G-Quadruplex topology?

Contrastive learning to fine-tune feature extraction models for the visual cortex

IBAS: Interaction-bridged association studies discovering novel genes underlying complex traits

Ten simple rules for effective use of generative AI for code development in environmental science

Reconciling contradictory models of subthalamic nucleus contributions to basal ganglia beta oscillations

The limitations of non-mechanistic methods for characterizing pathogen-pathogen interactions: A simulation study

Structure-aware deep learning enhances m6A prediction and reveals cell type-associated RNA structural signatures

The SATvac model of CD8⁺ T cell expansion and contraction phases considering memory and effector cell differentiation