2026 SysMod annual meeting

Integrating systems biology and bioinformatics

July, 2026 | Washington, DC

Methods

Dynamical modeling
Flux balance analysis
Logical modeling
Network modeling
Stochastic simulation
...

Systems

Animals
Bacteria
Humans
Plants
Yeast
…

Applications

Bioengineering
Cancer
Developmental biology
Immunology
Precision medicine
…

Prof. Shayn Peirce-Cottler,
Department of Biomedical Engineering,
University of Virginia

Overview

Advances in genomics are creating new opportunities to understand the biology that require both systems modeling and bioinformatics. The tenth annual SysMod meeting will be a forum for discussion about the combined use of systems biology modeling and bioinformatics to understand biology, and disease. The meeting will take place in July 2026, during the 2026 ISMB/ECCB conference in Washington, DC, USA. The meeting will feature two keynote talks and contributed presentations.

Topics

Methods

Dynamical modeling Flux balance analysis Logical modeling Network modeling Stochastic simulation …

Systems

Animals Bacteria Humans Plants Yeast …

Applications

Bioengineering Cancer Developmental biology Immunology Precision medicine …

Keynote speakers

Shayn Peirce-Cottler
Department of Biomedical Engineering
University of Virginia

Jason Papin
UVA School of Medicine
University of Virginia

Schedule

11:00-13:00	Session I: Computational Disease Dynamics and Therapeutic Modeling
11:00 – 11:15	Welcome and Introduction to SysMod 2026 Abstract The community of special interest (COSI) in systems modeling (SysMod) organizes annual one-day gatherings. In 2026 the meeting comprises three sessions that cover a broad variety of topics, beginning with metabolic modeling, followed by the afternoon session on multiscale modeling and concludes with inference of cellular processes. This year’s meeting features two keynote speakers, Shayn Peirce-Cottler and Jason Papin.
11:20-12:10	Keynote talk: From Single-Cell Decisions to Tissue-Scale Disease: A Multiscale Modeling Framework Shayn Peirce-Cottler University of Virginia, USA Abstract Many of the most prevalent and challenging human diseases—including cancer, cardiovascular disease, and fibrotic disease—emerge from dynamic, multiscale interactions among heterogeneous cell populations and their evolving microenvironments. Yet most therapeutic strategies and computational models focus on isolated molecular targets, limiting their ability to predict or influence emergent tissue-level behavior or patient-specific outcomes. Addressing this gap requires mechanistic models that explicitly link intracellular signaling, cell phenotypic decision-making, and tissue-scale remodeling across space and time. In this talk, I will present our laboratory’s work developing multiscale mechanistic computational models that integrate molecular, cellular, and tissue-level processes to study disease onset and progression. Our approach centers on agent-based modeling, enabling explicit representation of individual cells, stochastic behaviors, and cell–cell interactions while capturing emergent, tissue-scale dynamics. These models are specified and parameterized using data-driven, machine learning approaches, allowing us to incorporate high-dimensional experimental data while preserving mechanistic interpretability. Importantly, our modeling efforts are tightly coupled to experiments in an iterative loop. We design experiments to probe single-cell behaviors and microenvironmental cues, use these data to refine model structure and parameters, and then generate testable predictions regarding tissue-scale adaptations. Using this framework, we have studied microvascular network and tissue remodeling in development, regeneration following injury, and fibrotic disease. Together, these studies demonstrate how multiscale mechanistic models can reveal causal links between cellular decision-making and tissue-level disease phenotypes, providing a predictive foundation for rational therapeutic design.
12:10-12:30	Disentangling cellular dynamics of co-occurring processes via stochastic graph transport Jie Sheng, Noah Cohen Kalafut and Daifeng Wang Abstract Complex systems exhibit dynamics governed by the interplay of intrinsic, co-occurring processes. For instance, emerging single-cell and spatial transcriptomics data reveal that cellular gene expression changes entangle simultaneous biological processes like maturation, proliferation, and spatial organization. Current trajectory inference methods for dynamic analysis typically impose a global smoothness constraint, ordering cells into an aggregated, uni-scale progression that conflates distinct processes and thus overlook their specific dynamic patterns. To address this, we developed VAPOR, a generative framework to infer and disentangle latent dynamics for co-occurring biological processes from single-cell data. To reconstruct continuous transitions from static observations, we employ a Markov process that leverages graph transport on a local stochastic neighborhood graph to estimate feasible state transitions for each cell in the variational autoencoder latent space. Using such transitions, VAPOR models and infers latent dynamics as ordinary differential equations. More importantly, it decomposes the latent dynamics into a set of process-specific components parameterized by transport operators (TOs) and their corresponding weights. Each TO defines a process-specific dynamic component, and its weight for each cell quantifies the process’s contribution to the cell’s dynamics with associated dynamic genes. After simulation studies and benchmarking, we applied VAPOR to diverse real-world datasets, including cross-species brain development, multiplexed cancer drug responses, and spatial transcriptomics of the mouse hippocampus. VAPOR disentangled a variety of temporal and spatial co-occurring processes, such as neuronal differentiation, cell cycle, and extracellular signal-regulated kinase (ERK) suppression after drug treatment. Furthermore, VAPOR also identified corresponding dynamic genes for those processes, such as RBFOX3, showing an earlier progression in macaque than human neurogenesis.
12:30-12:40	Mechanistic ODE modeling of MAPK/AKT signaling reveals scheduling-dependent control of persister dynamics in BRAF-mutant melanoma Ethan Wang Abstract Melanoma relapse is frequently associated with drug-tolerant persister cells, rare tumor subpopulations that survive targeted therapy through adaptive signaling rather than stable genetic resistance. Because these cells remain dependent on MAPK and RTK/AKT signaling, combination strategies that disrupt both pathways may improve control of drug-tolerant populations. Here, we present a mechanistic ordinary differential equation (ODE)-based model to investigate treatment scheduling strategies combining a PROTAC degrader targeting mutant BRAF with an RTK inhibitor. The model integrates MAPK and AKT signaling dynamics, downstream transcriptional programs, and tumor population behavior, with drug action represented through two-compartment pharmacokinetics, bolus dosing, and Hill-type pharmacodynamics. Treatment performance was evaluated using a persister index (PI) together with a composite optimization score (COS) that balances persister burden and drug exposure. Screening across a large set of candidate regimens identified scheduling strategies that reduce simulated persister populations relative to baseline conditions. Regimens incorporating early PROTAC exposure followed by sustained RTK inhibition achieved the largest reductions in PI while maintaining lower modeled toxicity penalties. Additional analysis across simulated tumor phenotypes revealed variability in optimal schedules, reflecting differences in underlying signaling dynamics. Overall, this framework demonstrates that, within a mechanistic modeling context, treatment scheduling can substantially influence persister dynamics and comparative performance metrics. These results provide a basis for exploring combination treatment strategies and generating hypotheses for future experimental validation of scheduling-based approaches to targeting melanoma persister cells.
12:40-13:00	Computational pipeline elucidates how single-cell, lymphocyte motility behaviors drive B-T interactions that mediate the antibody response Nikita Sivakumar, Chanhong Min, Kibaek Choe, Wendy Abstract Lymphocytes rely on cell motility to navigate tissue environments and engage in cell-cell interactions that support the adaptive immune response. Within germinal centers, B-cells and T-cells rely on cell motility to encounter and engage in B-T interactions that enable antibody affinity maturation. Systematically quantifying how B-cell and T-cell motility influences B-T interactions within germinal centers can inform what cellular mechanisms support a healthy antibody response to infection or vaccination. Experimental methods, alone, cannot simultaneously image single-cell motility in vivo and all B-T interactions within high-density tissue environments, like the germinal center. To close this gap, we developed and validated PRISMM (Pipeline for Recapitulating Cell-Cell Interactions using Spatial Motility Modelling). PRISMM first applies unsupervised machine learning to learn single-cell, lymphocyte motility behaviors from intravital, time-lapse imaging data. PRISMM then applies agent-based simulations to infer how these motility behaviors drive B-T interactions. The application of our pipeline identified that germinal center T-cells tend to take on fast and exploratory movements, while B-cells tend to move more slowly. Our simulations show that these distinct motility behaviors allow B-cells to maximize unique interactions with T-cells within confined volumes. This result suggests that baseline lymphocyte motility behaviors optimize B-T interactions during a physiological antibody response. PRISMM leverages the single-cell nature of motility data to identify distinct, in vivo cell motility behaviors. Our pipeline can then predict how these single-cell motility behaviors drive emergent cell-cell interactions. Using PRISMM we can understand how cell motility influences cell-cell interactions in healthy and disease contexts.
13:00- 14.20	Lunch Break
14.20-16.00	Session II: Multi-Scale Systems Biology and Network Modeling
14:20-15:10	Keynote talk: Predicting microbial metabolic function for therapeutic development Jason Papin University of Virginia, USA Abstract With the explosion of data characterizing the genotype-phenotype relationship of microbes under diverse conditions, there remains the challenge to interrogate such data to better understand and predict metabolic functions of microbes. Computer models have become indispensable tools to address these challenges. We will discuss recent methods to construct and test computer models of microbial metabolism and how these models are shaping the way we think about and design therapeutic strategies to treat human disease.
15:10-15:30	Metabolic signatures of immune checkpoint inhibitor response in gynecologic cancers: Insights from flux balance analysis Gideon Idumah, Lin Li, Lamis Yehia, Haider Mahdi and Ying Ni Abstract Modifiers of immune checkpoint inhibitor (ICI) responses in cancer patients are complex and remain poorly characterized, especially in gynecologic cancers. In this study, we explored fluxomic biomarkers that differentiate responders from non-responders to ICIs in a series of 49 patients with gynecologic cancers, including ovarian, cervical, and endometrial cancers. By applying metabolic enzyme expression as constraints, we utilized an objective-customizable flux balance analysis within a genome-scale metabolic model to predict the metabolic flux differences between responders versus non-responders of ICI treatment. We identified three reactions with consistent differential activity across all ten different optimization objectives: Succinate Dehydrogenase (SUCD1m) in the citric acid cycle, NADH: Guanosine-5-Phosphate Oxidoreductase (r0276) involved in purine catabolism, and Ornithine Transaminase Reversible, Mitochondrial (ORNTArm) in the urea cycle. Additionally, reactions within the folate cycle subsystem, particularly involving MTHFD2, demonstrated significance in distinguishing treatment responses, aligning with previous findings linking MTHFD2 to immune evasion and tumor progression. To further analyze the association between metabolic features and survival outcomes, we implemented machine learning models that integrate multi-omics data. Our model included clinical-pathologic, molecular-genomic features (gene expression, TGF-β score, immune cell abundance from transcriptomic deconvolution), and significant reaction fluxes. Our findings suggest that SUCD1m, MTHFDm and ORNTArm are important metabolic biomarkers that could serve as predictive indicators for ICI response and, if validated in a larger cohort, may guide the development of targeted therapies to enhance treatment efficacy for gynecologic cancer patients. This study highlights the use of genome-scale metabolic modeling to identify clinically relevant biomarkers and improve therapeutic strategies.
15:30-15:50	SPECIMEN powered by refineGEMs for streamlined automated high-quality genome-scale modelling Carolin Brune, Gwendolyn Olivia Döbel and Andreas Dräger Abstract Background: Systems biology aims to provide a systematic, comprehensive understanding of living systems at all scales through context-specific computational models. Metabolism has been identified as a primary aspect for this endeavour because of its well-defined physico-chemical foundations and its involvement in all aspects of life. In many applications, genome-scale metabolic models (GEMs) have demonstrated their usefulness to predict various phenotypic features based on firm principles. Even though GEMs follow relatively simple mathematical structures, their reconstruction and curation are demanding and still involve many manual steps. Results: We introduce SPECIMEN, a collection of workflows designed for automated, standardised strain-specific metabolic modelling and GEM curation. SPECIMEN primarily utilises the toolbox refineGEMs, which provides a database of defined growth media and a plethora of functions for reading, manipulating and curating individual aspects of GEMs, including mass/charge curation and the assignment of SBO terms and EC numbers. RefineGEMs demonstrated its usefulness in reconstructing five C.-striatum models. SPECIMEN combines these functions to provide two principal reconstruction pipelines, both based on previously published methods: the High-Quality Template-Based (HQTB) and the CarveMe-ModelPolisher (CMPB) workflow. CMPB starts from a CarveMe-generated draft model, refines and extends it into a high-quality, strain-specific model, whereas HQTB employs a user-provided template model and genome as its basis for reconstructing a new model. Conclusion: The combination of both tools drastically reduces manual effort while semi-automatically producing directly usable, high-quality GEMs that comply with all current standard operating procedures and best-practice guidelines. Availability: Both Python-based open-source software solutions are freely available at github.com/draeger-lab/SPECIMEN and github.com/draeger-lab/refineGEMs.
15:50-16:00	Temporal competence windows govern four discrete phases of neural progenitor fate commitment in a multi-layer ODE framework of CNS development Batuhan Safa Kar Abstract BACKGROUND How positional signals are translated into discrete neuronal identities across developmental time remains incompletely understood. Transcription factor (TF) competence windows—periods during which progenitors respond to inductive signals—constrain fate decisions, but their system-level dynamics remain uncharacterized. METHODS We developed a six-layer ODE framework modeling CNS progenitor fate commitment along anterior-posterior and dorsal-ventral axes, integrating: (1) a bistable GSE/SVE switch via Lhx3-Phox2b mutual inhibition (Hill n=4), (2) rhombomer boundary sharpening via Hoxb1-Hoxb2 cross-repression (Hill n=6), (3) Notch-Delta lateral inhibition, (4) Eph-Ephrin contact barrier, (5) CN VII axon guidance, and (6) temporal TF competence gating across 18 transcription factors. Validity was assessed against 11 quantitative biological criteria, including SHH dose-response benchmarking against Dessaud et al. (2008). RESULTS Systematic variation of ODE integration time (t_norm=0.03-1.0) revealed four discrete commitment phases. Phase I (t<0.125): complete progenitor indeterminacy. Phase II (t=0.16-0.38): rapid commitment via sequential TF window opening—somatic motor neurons commit before branchiomotor neurons, recapitulating known in vivo temporal ordering. Phase III (t=0.38-0.875): full commitment, zero residual progenitor fraction. Phase IV (t=1.0): partial reversion caused by Chx10 window closure, isolating ARAS progenitors as a structurally irreducible population. GSE-axis parameters showed R²=0.995 concordance with Dessaud dose-response data (EC50 deviation: 2.6%). CONCLUSIONS Temporal TF competence gating—not morphogen signal strength or bistable switch kinetics—is the dominant determinant of commitment timing. The four-phase structure and Hill coefficient independence of commitment rates are emergent properties unreachable by individual subsystem analysis
16:00- 16.40	Coffee Break
16.40-18.00	Session III: Dynamical Modeling of Cellular State Transitions
16:40-17:00	Decoding Anti-TNF therapy outcomes from pre-treatment single-cell data via network-based perturbation modeling Mathilde Meyenberg, Matthew Leipner, Alan James, Mueller-Breckenridge and Jörg Menche Abstract Despite anti-TNF therapies being a cornerstone of IBD management, non-response rates remain as high as 40%. To address the lack of predictive baseline biomarkers, we developed a computational framework that leverages multi-layer network integration and in silico perturbation modeling to predict drug response from pre-treatment single-cell transcriptomics. Using longitudinal data from UC and CD patients treated with adalimumab, we construct cell-type-specific networks by mapping Wilcoxon-ranked genes to a protein-protein interaction (PPI) network. To ensure disease relevance, networks are filtered using LCC-based z-scores, retaining only those where IBD-associated modules show significant connectivity. These layers are integrated into multiplex networks, where cell-specific disease neighborhoods are defined via a multiplex random walk with restart (RWR). These network neighborhoods define subnetworks that reveal distinct differences in connectivity around IBD genes for CD and UC. To model perturbations, layer-specific subnetworks are merged into union graphs representing disease (inflamed cells) and reference states (non-inflamed cells) from pre-treatment data. We simulate therapy by deleting target nodes and edges, evaluating the impact via a graph hopper kernel. This kernel utilizes semantic gene embeddings and topological features to compare graph similarity. By deriving a normalized effect score from the resulting distance matrices, we successfully recover a differential signal of anti-TNF response from baseline data alone. By investigating the cell-specific connectivity around TNF on the multiplex networks, we gain biological insights into how pre-treatment cellular signaling influences clinical outcomes.
17:00-17:20	PIMENTO: a Physics and bIology-inforMed Neural network for dynamic inference of gene regulatory networks to model plant response to biotic stresses Xavier Amorós-Gabarrón, Alessandra Goncalves Ribeiro, Krishna Kumar, Pratila Debnath, Regis Duvigneau, Justyna J. Olas, Lorenzo Sala and Silvia Bottini Abstract In a scenario of climate change, plants are exposed to an increasing variety of pathogens, accompanied by additional hostile conditions. Comprehensive understanding of plant immunity is essential to ensure food production and security. Gene regulatory networks (GRNs) provide a powerful representation of these intricate regulatory systems. However, most inference methods measure only static properties and face the trade-off between scalability and mechanistic interpretability. Here, we propose PIMENTO, a novel hybrid model based on physics-informed neural-networks to infer fully interpretable GRNs from time-series omics data. This approach exploits the ability of neural-networks to approximate high-dimensional systems in combination with explicit ODEs and prior biological knowledge. PIMENTO showed robustness on synthetic data compared to random predictors when inaccurate prior is provided, when trained on as few as 4 timepoints and noise. PIMENTO on 21 time-series (3-7 timepoints, 2- 8 replicates) of tomato (Solanum lycopersicum) plants upon infection by diverse economically important pathogens, with prior network composed of 11822 genes, 277 transcription factors (TFs) and 66343 edges, yielded several TFs with pivotal role in plant defense. Among the top 10, Solyc12g009240-ERF16 and Solyc06g070900-TCP17 were functionally validated using the virus-induced gene-silencing system in tomato, demonstrating their key role in the response to Botrytis cinerea infection, as silencing of either TF resulted in increased susceptibility. Overall, PIMENTO showed good performances regarding scalability, interpretability, predictivity and suitability to model time-series with few timepoints, which will fulfil a gap in the literature. PIMENTO can be used beyond plants, since the model is agnostic of the biological system
17:20-17:40	Mechanistic modeling of rTMS-mediated quiescent cell targeting in glioblastoma Chitransh Dave, Vrutti Mehta, Kenza Benzeroual and Nicolas Gallo Abstract Glioblastoma (GBM) remains highly resistant to therapy in part because hypoxia, impaired perfusion, and treatment-associated cell-state transitions create a microenvironment that favors quiescent, therapy-resistant tumor populations. These latent physiological states are not well captured by models that evaluate treatments in isolation, despite their major influence on therapeutic response. We developed a mechanistic ODE-based computational framework to test whether repetitive transcranial magnetic stimulation (rTMS) can shift this balance by modulating intratumoral perfusion and oxygenation, thereby reducing quiescent-cell burden. The model couples proliferating, quiescent, and necrotic tumor compartments with hypoxia–perfusion dynamics and rTMS as an upstream physiological perturbation. In this framework, rTMS is not treated as a direct cytotoxic agent, but as a driver of downstream changes in perfusion, oxygenation, delivery, and cell-state transitions that alter treatment sensitivity. Simulations across a heterogeneous Monte Carlo virtual patient cohort, combined with scenario exploration and sensitivity analysis, reveal parameter regimes in which modest hemodynamic changes reshape tumor burden and alter combination response. Together, these results establish an in silico test platform for studying neuromodulation-enabled therapeutic synergy in GBM.
17:40-17:55	Closing Remarks Abstract This concluding talk aims to briefly discuss the diversity of topics presented at the “Computational Modeling of Biological Systems” (SysMod) COSI track. This diversity illustrates the importance of the field and the broad range of applications in systems biology and disease. Then, forthcoming meetings of interest will be announced, and the three poster awards will be delivered as a closing event.

Poster presentation:

In Silico Validation of a Dual-Sensing CRISPRa Genetic Circuit for Personalized Diabetic Wound Therapy
Priya Emani

Abstract

Diabetic wounds pose a significant health challenge from the resilience of biofilms and the inflammation of the wound environment. Current antibiotics often lack targeted and adaptive responses. This project introduces a ‘living medicine’ of a smart bacteria with a novel genetic circuit designed to sense key wound microenvironment signals (low pH, high glucose) to respond with localized delivery of anti-biofilm (DNase I) and anti-inflammatory (IL-10) agents, in a dual manner. The genetic circuit was designed using Benchling, with pH-inducible DNase I expression and glucose-sensitive IL-10 expression units regulated by CRISPRa, based on a PTS/Crp system. The design was tested with PhysiCell that simulated a multicellular system of a virtual wound, validating the engineered bacteria’s sensing and secretion logic against the Staph aureus biofilm. The temperature sensitive kill switch (temp >37°C trigger) was modeled for reliability using NetLogo simulations. The engineered bacteria released DNase I predominantly under low pH and IL-10 mainly under high glucose conditions. The quantitative predictions through PhysiCell showed biofilm mass reduction of a potential 50-70%, based on personalization of Hba1c levels. Compared to the 30% Vancomycin biofilm reduction, the genetic circuit outperformed Vancomycin in the different ranges of HbA1c levels. Additionally, the kill switch safety feature demonstrated high predicted reliability in NetLogo models with >97% effectiveness. The in silico simulations validated the design, highlighting specific responses and predicted efficacy that met the engineering goals. The computational results provided promising next steps for in vitro validation and highlighted the potential of personalized synthetic biology for treating diabetic wounds.

Systems Pharmacology Analysis of Off-the-Shelf Peptide Vaccine Combined with PD-1/CTLA-4 Blockade in Fibrolamellar Carcinoma
Osman N. Yogurtcu, Joel Eliason, Aleksander S. Popel, and Hong Yang

Abstract

We developed a preliminary quantitative systems pharmacology (QSP) model of the FLC-Vac vaccine plus nivolumab/ipilimumab for fibrolamellar carcinoma (FLC), qualitatively benchmarked against clinical trial observations, to provide initial estimates of component contributions and scheduling effects. Adapted from a hepatocellular carcinoma platform, this mechanistic ordinary differential equation model was tailored to FLC biology. It incorporates FLC-Vac depots, tumor-draining lymph node antigen handling, CD4+/CD8+ T cell priming, tumor trafficking, checkpoint-antibody pharmacokinetics, suppressive circuits, and an exhausted CD8+ pool. Deterministic simulations mirroring the published protocol compared the full combination against arms missing individual components (vaccine/adjuvant, nivolumab, or ipilimumab) and evaluated concurrent ICI initiation versus a 4-week vaccine lead-in using Julia.
The full combination yielded modest tumor shrinkage (-16.4% best diameter change), outperforming simulated ICI-only conditions, directionally consistent with modest historical ICI activity. Removing the vaccine/adjuvant reduced peak lymph-node pMHC-II (38.8%) and activated CD4+ cells (45.0%). Removing nivolumab increased the week-10 exhausted-to-active intratumoral CD8+ ratio 3.18-fold, worsening tumor response by 16.4 percentage points and accelerating progressive disease. Omitting ipilimumab reduced peak activated lymph-node CD4+ (43.3%) and tumor CD8+ effectors (22.5%). A 4-week vaccine lead-in before ICI worsened tumor response by 6.5 percentage points and attenuated early lymph-node CD8+ priming, qualitatively matching clinical trial observations.
This model supports a mechanistic division of labor: FLC-Vac drives class-II/CD4+ priming, nivolumab restrains local CD8+ exhaustion, and ipilimumab augments CD4+/Treg control to enhance intratumoral CD8+ accumulation. Concurrent ICI initiation proved superior to a vaccine lead-in. These preliminary estimates warrant further refinement via virtual-patient analyses.

DeepTaxa: A Hybrid CNN-Transformer Architecture for Hierarchical Taxonomic Classification of 16S rRNA Sequences
Ahmed Moustafa, Rana Salah Khalel, Khlood Abdelaal, Lobna Ghonaim, and Olaitan I. Awe

Abstract

Accurate taxonomic classification of 16S rRNA gene sequences remains central to microbial ecology, yet existing methods struggle at fine-grained ranks where sequence divergence is minimal and label spaces are large. We present DeepTaxa, a deep learning framework that jointly predicts taxonomy across all seven standard ranks (domain through species) using a hybrid architecture that pairs convolutional neural networks with a BERT-style transformer. The CNN branch extracts local k-mer motifs through multi-kernel convolutions, while the transformer captures long-range contextual dependencies. Input sequences are tokenized using DNABERT-2 byte-pair encoding, which outperforms one-hot nucleotide encoding by over 6 percentage points at the species level.
Trained on Greengenes2 2024.09 (277,336 training / 69,335 test sequences), the hybrid model achieves 92.9% species accuracy with tight cross-seed reproducibility (std < 0.04 pp). Against DADA2’s naive Bayesian classifier on the same test set, DeepTaxa gains +2.1 percentage points at the species level while running roughly 50 times faster on a single GPU. To rigorously evaluate each component, we conducted Optuna-based hyperparameter optimization (20 trials per architecture) so that every variant receives a fair comparison at its best configuration. The optimized hybrid (92.97%) outperforms the optimized CNN-only (92.29%) and BERT-only (86.41%), confirming that both branches contribute meaningfully. The BERT-only model performs catastrophically (4.7%) with the hybrid’s default hyperparameters but recovers to 86.4% with proper tuning, highlighting how sensitive transformers are to learning rate on genomic data. Similarity-stratified testing shows predictable degradation with decreasing train-test identity: 95.3% for high-similarity sequences (>97%), 68.1% for medium (90–97%), and 47.8% for low (<90%). Zero-shot evaluation on in-silico V3-V4 and V4 amplicon extractions shows that shorter fragments lose information as expected (59.5% and 37.4%, respectively). Across all scenarios, the model maintains well calibrated confidence estimates (ECE < 0.03), making predictions suitable for downstream filtering in metagenomic pipelines.

Kinetic modelling of multimeric protein complexes in yeast
Xavier Castellanos-Girouard, Stephen Michnick, and Adrian Serohijos

Abstract

Cells are organized into structured networks of functionally interdependent biomolecules. These networks are integrators of genetic information and environmental context, shaping cellular phenotype. Therefore, accurate models of biological networks and their dynamics is key to understanding the molecular basis of genotype-phenotype relationships.
Protein-protein interaction (PPI) networks are among the most comprehensively mapped biological networks, benefiting from large-scale experimental screens across model organisms. These datasets have subsequently driven the construction of theoretical kinetic and thermodynamic models of PPI networks. Advances in quantitative mass spectrometry have enabled the measurement of interaction stoichiometry (a ratio analogous to fraction of protein bound in a complex). However, no comprehensive kinetic models to date have incorporated this variable.
We previously used measurements of interaction stoichiometry to generate the first proteome-wide estimation of dissociation constants (Kd). Building on this foundation, we now integrate interaction stoichiometry, total protein abundances, and known subunit stoichiometries to estimate intracellular numbers of multimeric protein complexes, free (unbound) protein abundance, and effective equilibrium constants (Kds) for multimeric complexes in the budding yeast Saccharomyces cerevisiae.
We find reasonable concordance between our estimates of complex abundances and those from low-throughput experiments in the literature. Further, building on previous models of reversible heterodimeric PPI networks, we construct a kinetic model of the yeast protein interactome to evaluate how perturbations in specific protein abundances propagate through neighbouring complexes. Finally, we systematically investigate whether gene pairs demonstrating high network sensitivity to abundance changes also exhibit genetic interactions.

Robust and Interpretable Modeling of Dynamic Metabolic States from Longitudinal Multi-Omics under Dietary Perturbation and Aging
Kira Liu, Dudley Lamming, and John Denu

Abstract

Understanding dynamic metabolic states under perturbations and aging is crucial for metabolic adaptation. However, their robust identification from longitudinal multi-omics data remains challenging, due to the high-dimensional, heterogeneous, and sparse nature of multi-omics datasets. We propose a computational framework for dynamic metabolic state discovery that integrates heterogeneous multi-omics data to identify phenotype-relevant states under dietary perturbation and aging.
The framework embeds transcriptomics, compositional histone post-translational modification (PTM) proteomics, and metabolomics into a shared latent space using a latent factor model. We infer state structure with ensemble density peak clustering (eDPC), combined with kernel density estimation, automatic cluster-number selection, and bootstrap resampling to improve robustness. In simulated benchmarks, eDPC achieved 14.6% higher mean ARI and 48.1% lower center estimation error than DPC. For interpretability, we use a time-varying mixed-effects model to characterize directional and temporal molecular patterns and map them back to latent space, enabling each inferred state to be defined by its molecular dynamics and linked to metabolic phenotypes.
We then applied this framework to longitudinal liver multi-omics data from diet-perturbed and aging mouse cohorts, with Western diet control and short-term BCAA restriction followed by Western diet repletion at multiple time points. This design enabled integrated analysis of perturbation response, recovery trajectories, and age-dependent metabolic remodeling.
We uncovered a distinct metabolic state associated with epigenetic memory, persistently elevated energy expenditure, and reduced hepatic lipid accumulation. Our framework enables interpretable multi-omics state discovery and mechanistic insight. Acknowledgment: Data from unpublished work by Calubag et al. (Denu and Lamming labs).

GSFM: A Gene Set Foundation Model Pre-trained on a Massive Collection of Diverse Gene Sets Applied to Gene Set Enrichment Analysis
Daniel Clarke and Avi Ma’Ayan

Abstract

Foundation models have transformed scientific and commercial domains. Trained on massive datasets, foundation models can effectively capture complex patterns within these datasets and convert these patterns into embeddings that can be used for a variety of useful applications. Here we created a gene set foundation model (GSFM) that was trained on millions of unlabeled gene sets from two large gene set databases: Rummagene and RummaGEO. Rummagene automatically extracts gene sets from supplemental tables of publications deposited in PubMed Central (PMC); and RummaGEO has gene sets automatically computed from comparing groups of samples from RNA-seq studies deposited into the gene expression omnibus (GEO) and uniformly aligned by ARCHS4. Several foundation model architectures and training methodologies were benchmarked in the downstream tasks of predicting gene function, gene-disease associations, and protein-protein interactions, as well as gene set enrichment analysis. The selected best GSFM architecture is a denoising autoencoder trained on multi-hot encoded gene sets and it achieves state-of-the-art performance on gene function prediction tasks. GSFM was used to systematically augment gene sets from a collection of sources to create gene pages for all human genes. It also applied to performing gene set enrichment analysis with benchmarks demonstrating superior performance compared with other enrichment analysis algorithms and tools. The GSFM web-interface, model, and source code can be accessed from: https://gsfm.maayanlab.cloud.

Identifying translatable axes of drug resistance in NSCLC via integration of patient and PDX single-cell transcriptomics
Paulina Eberts

Abstract

Understanding drug resistance in NSCLC is critical for improving patient outcomes; however, the sparsity of single-cell data collected from patients on treatment makes it difficult to ascertain cellular state transitions at this resolution. Mechanistic insights therefore rely heavily on patient-derived xenograft (PDX) models, which lack the full set of complex cues present in patients.
Discerning which PDX-derived phenotypes reflect human biology versus model-specific artifacts remains a fundamental challenge limiting translational utility. Here we employed our TransComp-R (Translatable Components Regression) framework to identify gene covariance structures conserved between human and PDX tumor cells. We integrated scRNA-seq data from TKI-treated patients at three timepoints (treatment-naive, minimal residual disease (MRD), progressive disease) with snRNA-seq data from seven different PDX models collected at equivalent stages.
Differentially expressed genes differentiating human treatment populations were used to construct PDX-derived principal components at the pseudobulk level. Human samples were projected into the PDX PC spaces to identify components along which treatment groups separated in both systems with consistent directionality (validated by logistic regression with 5-fold cross-validation).
Principal components meeting these criteria, reflecting shared gene covariance structures, were designated translatable axes. Individual PDX cells were projected along the translatable axes and pseudo-ordered for single-cell trajectory analysis. MRD-enriched transcriptional signatures pre-existed in untreated tumors, suggesting intrinsic fitness enriched under treatment conditions and implicating genes such as S100A9 as key contributors to the transition. This framework bridges PDX and patient biology to expose the earliest, most therapeutically vulnerable stages of drug resistance.

A Novel Computational Framework for Fractal Analysis of Spatial Transcriptomic Data
Elijah Yu and Xinmin Li

Abstract

Fractal analysis is an emerging mathematical technique able to characterize the complexity of biological systems. While traditionally used to analyze morphological imaging data, this study aims to create a computational framework applying the methods and concepts of fractal analysis to 10X Visium transcriptomic outputs. The pipeline was validated using public datasets of glioblastoma and triple-negative breast cancer samples.
Given the discrete nature of Visium spots, individual spots were assigned deconvoluted topics using STdeconvolve, and subsequently categorized as “tumor” or “host” based on top genes and common tumor-associated markers. Then, Python scripts for the box-counting algorithm were used to calculate measures of monofractal tumor boundary complexity through the fractal dimension (Df), while the Chhabra-Jensen algorithm was used to calculate measures of multifractal complexity with the singularity exponent range (Δα) and singularity spectrum (f(α)).
Df and Δα values were interpreted based on transcriptomic and histological profiles of the tissue sections, and cross-referenced with original findings. Our results demonstrated that higher values of Df and Δα illustrate extensive intermingling and complex interactions between host and tumor tissue, while lower values represent homogeneous tumor regions or a defined host-tumor boundary. Compared to spatial metrics like Moran’s I, which measures spatial autocorrelation, Df and Δα values can quantify boundary complexity and diversity of probability distributions, respectively. The proposed framework offers a novel perspective on the biophysical reality of tumor microenvironments and serves as a proof-of-concept for calculation of fractal metrics across a diversity of spatial transcriptomic platforms.

Structural Determinants of Neuromuscular Transmission Through 3D Reaction-Diffusion Modeling
Dweny Geeth

Abstract

Neuromuscular transmission depends on acetylcholine (ACh) diffusing across the synaptic cleft and binding to nicotinic receptors on the postsynaptic membrane. Disruptions in this process can cause diseases such as myasthenia gravis and muscular dystrophies. While fold geometry, receptor distribution, and kinetics have been studied independently, their combined influence on ACh signaling remains poorly understood.
A novel 3D reaction diffusion model of the neuromuscular junction was developed in MATLAB incorporating curved synaptic fold morphology, clustered receptor distributions, and multistate receptor kinetics. On a 7×9×9 voxel grid with Gaussian diffusion, 12 conditions were simulated across four receptor geometries and three firing patterns (regular 50 Hz, regular 100 Hz, burst 100 Hz), tracking ACh concentration, occupancy, and desensitization.
Burst firing at 100 Hz was the only condition approaching EC50 (~100 µM), sustaining concentrations above threshold for 457 µs, while regular stimulation peaked between 58 and 60 µM regardless of frequency. Receptor clustering was the primary structural driver: clustering alone reduced desensitization from 32% to 14% at 50 Hz, preserving an activatable receptor pool, whereas folds produced negligible changes.
Burst firing induced 70% desensitization versus 23 to 32% under regular firing, reflecting a frequency dependent tradeoff between accumulation and receptor availability. Fidelity index was maximal at 50 Hz (1.00), declined at 100 Hz (0.65), and recovered under burst firing (0.78) via temporal summation.
These results identify receptor clustering as the primary structural driver of ACh signaling, with burst firing uniquely enabling functionally relevant accumulation, informing transmission failure in neuromuscular disease.

A Generative Model for the Synthesizing of Single-Cell Proteomic Datasets for Dynamic Modeling of Cell Signaling Pathways.
Carolina Ribeiro, Cristiano Campos, and Marcelo Reis

Abstract

Motivation: Biological processes are dynamic and time-resolved cellular profiling is essential for understanding changes in cell states, signaling pathways and regulatory mechanisms during differentiation, stimulus response, and disease progression. While time-series gene expression data provide rich insights into these dynamics, the scarcity of comprehensive protein-level perturbation data limits systems-level modeling. In particular, single-cell proteomic measurements are often destructive and constrained in depth and coverage, hindering temporal reconstruction and the availability of longitudinal ground-truth datasets for robust algorithm evaluation.
Results: We introduce a reusable pipeline for generating synthetic single-cell proteomic (scProteomic) datasets from ODE-based biochemical models. The pipeline provides modular components for simulating signaling dynamics through numerical integration, enabling controlled variability across cell lines and treatments through randomized initial conditions and parameters. Using an IL-6 signal transduction model in hepatocytes, we generated a dataset to evaluate several random forest configurations. The first model predicts ERK abundance from other molecular markers independently of time. The second model employs a set of time-resolved models trained on resampled data of t0 and generating predictions for the other marks for each time point. A refined version integrates predictions from these temporal models into the first model, and a third jointly trains ERK and other markers. Comparison with the available ground-truth trajectories demonstrates the framework’s ability to simulate and benchmark time-series modeling approaches for scProteomic data.
Availability and Implementation The pipeline is freely available at https://github.com/cristiano-campos/scProteomics-GenerativeModel
Contact: c184298@dac.unicamp.br, msreis@unicamp.br

Image-Based Machine Learning For Alzheimer’s Detection
Varun Menon

Abstract

In this project I examined how well Convolutional Neural Networks (CNNs) can distinguish between the severity of Alzheimer’s disease using MRI brain images. Early diagnosis of dementia is desirable as it leads to better treatment outcomes and allows for earlier intervention, and machine learning techniques provide a useful methodology for improving diagnostic processes. A CNN model was developed and trained using a labeled dataset of MRI brain images sorted into four categories of classification: Non-Demented, Very Mild Demented, Mild Demented, and Moderate Demented. Prior to training, the images were processed and formatted to ensure consistency and improve model performance.
The model was trained over multiple iterations, allowing it to learn patterns and structural differences within the brain scans associated with each stage of the disease. When shown novel images during testing, the model performed with a high level of accuracy. While this indicates that the model can distinguish between the image groups, it may also be due to overfitting of the data used during training or limitations in dataset size and diversity.
All in all, this project demonstrates the usefulness of deep learning algorithms in identifying patterns within neuroimaging data in order to assist in diagnosing Alzheimer’s disease. However, further study would be beneficial to validate these findings and enhance the reliability and real-world applicability of this algorithm.

TEMPEST: A Probabilistic Framework for Longitudinal Analysis of Sparse, Asynchronous Spatial Data in Tumor-Stromal Assembloids
Eashan Monga, Dina Hany, Jake Chang, Gina Bouchard, and Sylvia Plevritis

Abstract

The tumor microenvironment strongly influences cancer progression and therapeutic response, yet capturing spatial, single-cell changes over time remains challenging. Longitudinal spatial analysis of patient-derived assembloids enables spatiotemporally-resolved comparisons between treated and untreated conditions, but such profiling is costly and produces sparse, asynchronous, and variable time-series data due to biological heterogeneity and technical noise. Moreover, existing temporal clustering and deterministic time-series alignment methods do not adequately reconcile asynchronous observations or account for variance in such data. Here, we analyze a lung cancer tumor-stromal assembloid profiled longitudinally with spatial proteomics to identify cellular spatiotemporal clusters that differ by treatment condition. To address these challenges, we introduce TEMPEST, a generative framework based on Bayesian Gaussian processes (GPs) for modeling spatiotemporal dynamics in longitudinal cellular spatial data. To extract distinct biological trends, TEMPEST leverages Gaussian processes to smooth asynchronous biological observations and generate posterior trajectory ensembles that capture uncertainty, applying Dynamic Time Warping (DTW) to align posterior draws and reconcile temporal shifts. In synthetic time-series data with varying noise levels, TEMPEST outperforms standard Bayesian and multi-task Gaussian process approaches in recovering true temporal clusters. Applied to our assembloid dataset, TEMPEST reveals treatment-associated spatial trajectories in cell-type pair colocalization that are not apparent in proportion-based summaries. Notably, we observe increasing spatial segregation among specific cell-type pairs during acute osimertinib exposure, a dynamic missed by proportion-based analyses and other Bayesian methods. These results establish TEMPEST as a robust and generalizable framework for analyzing noisy, asynchronous longitudinal data, particularly for identifying spatiotemporal patterns in low-sample-size biological studies.

Temporal multi-omics modeling resolves sex-specific regulatory mechanisms governing cardiac metabolic adaptation to endurance exercise
Pauline Brochet, Joyce Njoroge, Samuel Montalvo, Malene E. Lindholm, Sasha Gladkikh, Gregory R. Smith, Gina Many, David Jimenez-Morales, Hasmik Keshishian, Nicole R. Gay, Bingqing Zhao, Chia-Jui Hung, Christopher A. Jin, Clarisa Chavez, Daniel Nachun, Elena Zaslavsky, German Nudelman, Hanna Pincas, Jose Juan Almagro Armenteros, Kevin S Smith, Krista M. Hennig, Mary Anne S. Amper, Matthew Wolf, Mital
Vasoya, Nasim Bararpour, Navid Zebarjadi, Nikolai G. Vetr, Roxanne Chiu, Si Wu, Venugopalan D. Nair, Yongchao Ge, Blake B. Rasmussen, Martin J. Walsh, Michael P Snyder, Stephen B Montgomery, Stuart C. Sealfon, William E Kraus, Zhen Yan, Euan Ashley, Matthew T. Wheeler, and Daniel H. Katz

Abstract

Endurance exercise improves cardiac metabolism and reduces cardiovascular risk, yet the molecular programs that drive cardioprotection remain incompletely understood. Here, we applied a systems biology approach to the MoTrPAC rat cardiac dataset, integrating eight omic layers measured after 1-8 weeks of endurance training in males and females. Using a temporal multi-omic factor analysis framework (MOFA/MEFISTO) that models time as a continuous covariate, we identified 7 latent factors capturing coordinated temporal molecular variation across multi-omic layers. Factor 2 emerged as the most robust multi-omic signature (transcript-protein weight correlation r=0.49, p≤2.2e-16). This sex-consistent factor revealed a progressive shift toward enhanced glycolysis and reduced lipid transport, correlating with systemic VO2max improvements (r>0.8 in both sexes).
Conversely, Factor 1 identified an early sex-divergent epigenetic signal. Females exhibit chromatin closure at 1w while males show opening at 4w (P≤0.05) at the promoters of critical glucose-glycogen regulators (Pdk4, Ppp1r3d, Ppara). Epigenetic motif analysis identified the metabolic transcription factor KLF15 as the top candidate driver of this divergence. Furthermore, Factor 6-associated proteins were enriched in fatty-acid (FA) metabolism (FDR≤0.05), with a maximal expression at 2w in males, but 4w peak in females. Kinase enrichment confirmed these divergent strategies through GYS1 regulation, where male-specific dephosphorylation at 2w (Factor 7) promotes its glycogen storage function. These results identify a temporal divergence in metabolic substrate preference, where early female adaptation is characterized by oxygen-efficient glucose utilization while males favor fastest FA adaptation. Ultimately, this analysis establishes a cross-omic mechanistic chain linking early epigenetic regulation to downstream signaling and proteomic dynamics.

Temporal competence windows govern four discrete phases of neural progenitor fate commitment in a multi-layer ODE framework of CNS development
Batuhan Safa Kar

Abstract

BACKGROUND
How positional signals are translated into discrete neuronal identities across developmental time remains incompletely understood. Transcription factor (TF) competence windows—periods during which progenitors respond to inductive signals—constrain fate decisions, but their system-level dynamics remain uncharacterized.
METHODS
We developed a six-layer ODE framework modeling CNS progenitor fate commitment along anterior-posterior and dorsal-ventral axes, integrating: (1) a bistable GSE/SVE switch via Lhx3-Phox2b mutual inhibition (Hill n=4), (2) rhombomer boundary sharpening via Hoxb1-Hoxb2 cross-repression (Hill n=6), (3) Notch-Delta lateral inhibition, (4) Eph-Ephrin contact barrier, (5) CN VII axon guidance, and (6) temporal TF competence gating across 18 transcription factors. Validity was assessed against 11 quantitative biological criteria, including SHH dose-response benchmarking against Dessaud et al. (2008).
RESULTS
Systematic variation of ODE integration time (t_norm=0.03-1.0) revealed four discrete commitment phases. Phase I (t<0.125): complete progenitor indeterminacy. Phase II (t=0.16-0.38): rapid commitment via sequential TF window opening—somatic motor neurons commit before branchiomotor neurons, recapitulating known in vivo temporal ordering. Phase III (t=0.38-0.875): full commitment, zero residual progenitor fraction. Phase IV (t=1.0): partial reversion caused by Chx10 window closure, isolating ARAS progenitors as a structurally irreducible population. GSE-axis parameters showed R²=0.995 concordance with Dessaud dose-response data (EC50 deviation: 2.6%).
CONCLUSIONS
Temporal TF competence gating—not morphogen signal strength or bistable switch kinetics—is the dominant determinant of commitment timing. The four-phase structure and Hill coefficient independence of commitment rates are emergent properties unreachable by individual subsystem analysis

Identifying clinically relevant cell state interactions in the tumor microenvironment of IDH-mut glioma.
Arashdeep Singh

Abstract

Tumor microenvironment (TME) comprises a diverse milieu of cell types that occupy heterogeneous transcriptional states across tumors. Interactions among these cell states play a central role in shaping tumor progression and therapeutic response. However, systematic identification of functional cell-state interactions (CSIs) remains challenging, owing to the limited availability of scRNA-seq cohorts with clinical annotation and the lack of cellular resolution in bulk RNA-seq datasets.
Here, we present CSI-TME, a computational framework that generalizes the concept of gene-level interactions, such as synthetic lethality, to the level of cell states, enabling the inference of prognostic CSIs directly from large-scale bulk transcriptomic cohorts. Applying CSI-TME to IDH-mutant gliomas, we uncover a highly reproducible cell-state interaction network (CSIN) that is predominantly pro-tumorigenic and exhibits distinct activation patterns between IDH-mutant astrocytoma and oligodendroglioma. Malignant cell states within this network recapitulate multiple neuronal lineage programs, including astrocyte-like and oligodendrocyte progenitor–like states, and reveal key interactions between glioma stem-like cells and T cells.
The CSIN stratifies patient response to immune checkpoint blockade, underscoring its clinical relevance. Approximately 20% of identified CSIs are supported by direct ligand–receptor interactions and show spatial co-localization in spatial transcriptomic data. Notably, we identify a robust pro-tumorigenic interaction between tip-like endothelial cells and hypoxic malignant cells, mediated by multiple ligand–receptor pairs. In contrast, anti-tumor CSIs associated with oncogenic alterations are preferentially active in early disease stages, suggesting a transient tissue homeostatic response.
Collectively, CSI-TME provides a scalable and clinically grounded framework to uncover prognostic cell-state interactions and nominate therapeutic ligand–receptor targets, offering new insights into how coordinated cell-state dynamics shape the TME in IDH-mutant glioma.

Percolation and lifestyle transition in microbial metabolism
Rydberg Supo and Dennis Vitkup

Abstract

Microbial species exhibit a remarkable diversity in their metabolic properties, genome composition, and ecological distribution. However, it is currently not well understood how the structure of metabolic networks defines and reflects the lifestyle and phenotypic properties of diverse bacterial species across the tree of life. By analyzing thousands of genome-scale metabolic models of bacteria, we found a percolation-like transition characterized by a sharp increase in metabolic network functional connectivity. This transition – primarily associated with the completion of central carbon metabolism and the TCA cycle – occurs at a threshold of approximately 800 metabolic reactions or about 2000 protein-coding genes. Strikingly, species with metabolic network sizes below the transition are typically obligate symbionts, while species above it, are predominantly free-living generalists. The observed percolation transition is also reflected in multiple other genomic properties, such as a substantial decrease in the fraction of regulatory genes below the transition and increased evolvability for new metabolic phenotypes above the transition. Furthermore, we found that the distribution of bacterial genome sizes from unbiased environmental metagenomic sequencing also reveals two regimes corresponding to the observed transition. Overall, our work identifies two qualitatively different regimes in microbial metabolism and lifestyles, characterized by distinct structural and functional properties of their metabolic networks.

Tissue- and system-level discovery of gero-protective perturbation candidates using transcriptomic signature enrichment
Philipp Trollmann, Rocio Rodriguez Quiroz, Paul Okoro ,and Preshita Dave

Abstract

Aging is a gradual and irreversible biological process which not only drives the decline in tissue function across organ systems, but also significantly increases the risk of various aging-related diseases. To systematically investigate interventions that may reverse tissue-specific aging signatures, we derived tissue-specific aging profiles from GTEx, Tabula Sapiens, and Tabula Senis and compared them with signatures derived from the large-scale perturbational datasets Tahoe-100M, Perturb-Seq, DrugMatrix and SciPlex. By standardizing these signatures as OmicSignature objects and applying gene set enrichment analysis (GSEA), we classified perturbations as “gero-protective” or “gero-advancing” based on their similarity or anti-similarity with age-associated transcriptional changes. Candidate signatures were prioritized by computing the mean effect, the p-value weighted sum of effects, and the average directionality of effects across tissues. We identified several systemic gero-protective candidates. One of the most notable gero-protective drugs identified was fenofibrate, which exhibited a consistent effect direction across all three aging datasets and across top-contributing tissues, in agreement with prior literature. Other identified candidates, such as bezafibrate, enalapril, and pantoprazole, further support the analysis through existing or emerging evidence of protective effects. These results highlight a promising framework for the identification of systemic gero-effects across diverse aging and perturbational resources. Future work will include integrating these hits with drug mechanism annotations and cross-validating findings using GSEA leading-edge genes and Perturb-Seq knockdowns in order to prioritize consistent candidates for follow-up experimental validation. This systematic approach provides a clear strategy for identifying novel pre-clinical anti-aging therapies.

SXXX
XXX

Abstract

Data selection significantly impacts automated reconstruction of genome-scale metabolic networks
Blaine Bates and Jason Papin

Abstract

Genome-scale reconstructions of metabolic networks are powerful tools for studying metabolism at a systems level. The process of generating reconstructions involves several steps (e.g., gene annotation, compiling a draft network, gap-filling, etc.), all of which rely heavily on prior genetic and/or biochemical data. This reconstruction data can be obtained from multiple sources, and each source has its own advantages and limitations which affect the reconstruction process; however, the impact of data selection on the quality of reconstructed networks has not been systematically evaluated. To address this gap, we generated different sets of reconstruction data with varying levels of annotation accuracy or taxonomic coverage. We used each set of data to produce metabolic networks for several strains of bacteria. We compared these networks to manually curated ones for the same strains to assess reconstruction quality. Despite having different methods, two reconstruction tools, CarveMe and Reconstructor, perform similarly when they use the same reconstruction data. Additionally, while data with misannotated genes consistently worsened reconstruction quality, the magnitude of this impact varied considerably from one organism to another. Finally, adjusting which species were represented in the reconstruction data significantly affected reconstruction. When the data included species that were closer to the target organism, both precision and recall improved. Conversely, greater taxonomic coverage had little impact on recall and decreased precision. These results emphasize the importance of data selection and demonstrate principles to improve automated reconstruction of metabolic networks.

Precision Mental Health Analytics: Scalable Brain-Based Models for Population Insights
Anushree Bhople

Abstract

Mental health diagnostics often rely on symptom based assessments, limiting the discovery of predictive markers. This paper utilizes normative modeling and cutting-edge machine learning techniques to decode individual heterogeneity, predicting brain architecture and function across regions. The goal is to revolutionize mental health diagnostics by offering personalized insights for improved, reliable, and fair patient care. Leveraging extensive reference samples, this research aims to impact population health by unraveling nuanced mental health patterns. The innovative algorithms developed here will address the complexity of mental health data, potentially uncovering novel markers and paving the way for tailored interventions and treatments.

Spatiotemporal modeling of developmental architecture guides engineering scalable hPSC-derived esophageal mucosa.
Sadhana Gaddam, Ying Yang, Ian Glass, and Anthony Oro

Abstract

Human induced pluripotent stem cell-derived tissue engineering is a promising approach for patient-specific cell therapies, genetic diseases, and tumor pathogenesis and helps better understanding how cells coordinate at the tissue level through spatial organization and cell-cell signaling. We previously established a scalable manufacturing protocol for an ectoderm-derived stratified epithelium for cell replacement therapy but there is a clinical need to manufacture tissue replacements for endodermal-derived tissues like the esophagus. Even though endoderm-derived esophageal epithelium shares broadly similar structure to ectoderm-derived cells the signaling interactions differ, making manufacturing difficult. To bridge this gap, here we integrate single-cell and spatial profiling to build a spatiotemporal, multi-omics atlas of human esophageal development. This atlas delineates the evolving cellular composition of the developing esophageal epithelium and its surrounding stroma. We then use a machine-learning strategy to infer and rank combinations of candidate human developmental signals to derive in vitro specification of esophageal basal progenitors from pluripotent stem cells. This framework helps accelerate human tissue manufacturing to support regenerative therapies, genetic defects, and chronic wounds. Because esophageal mucosa stability depends on the local stromal niche, we have extended our work to focus on understanding the tissue architecture at the gastroesophageal junction using spatial transcriptomics. This work will help identify how epithelial and stromal niches align in this junction and how we can harness this information to investigate disease states such as Barrett’s esophagus and other metaplastic condition associated with an increased risk of esophageal adenocarcinoma.

Expressed Gene Landscape (EGL) Analysis: Systems-level Evaluation of Lung Cell Phenotypes and Druggable Targets
Bingfang Xu, Ajith Pankajam, Kumar Arvind, Anne Deslattes-Mays, Matthew Diller, Raymond LeClair, Beverly Peng, Noam Rotenberg, William Spear, Zhizheng Wang, Yun Zhang, and Richard Scheuermann

Abstract

A cell’s phenotype, and thereby its function, is dictated by the subset of genes and proteins it expresses. Recent advances in single-cell technologies allow us to quantify cell phenotype at the resolution of individual cells. Here, we present Expressed Gene Landscape (EGL) analysis, a framework that not only profiles the gene-expression-defined phenotype of individual cell types but also systematically evaluates phenotypic variation across cell types within a tissue.
EGL adopts gene-specific probabilistic base models to estimate the likelihood of gene expression in each cell type. After fine-tuning the base models and applying a data-driven probability threshold, EGL identifies all genes expressed in each cell type. To enable systems-level analysis, EGL generates a binary gene-by-cell-type matrix, with hierarchically organized cell types as columns and genes as rows. By defining input genes using Gene Ontology categories, EGL compares diverse biological functions across cell types. EGL is also integrated with an agentic AI workflow that uses large language models (LLMs) to guide input selection and aid in result interpretation.
Applying EGL to lung single-cell datasets, we constructed gene expression landscapes for key biological processes across lung cell types, defined gene expression landscapes in healthy and diseased lungs, and generated druggable gene expression landscapes. Within each landscape, EGL revealed three patterns of expression: ubiquitously expressed genes, cell-type-selective genes, and genes enriched in specific cell-type classes.
Together, EGL provides a comprehensive view of the functional characteristics of each cell type and serves as a tool for systems-level evaluation of cell phenotypes and therapeutic targets.

Mutation‑Induced Repulsive Hotspots Reshape Nanobody Binding to the Omicron SARS‑CoV‑2 RBD
Mert Golcuk, Fareeda E. Abu-Juam, Derman Basturk, Ayten Dilara Gursel, Clara Xazal Buran, Reyhan Metin Akkaya, Ahmet Yildiz and Mert Gur

Abstract

Why some SARS-CoV-2 Spike Protein targeting nanobodies lose potency against Omicron while others retain measurable binding remain unresolved at atomic resolution. We carried out all-atom simulation study of 13 nanobodies bound to the Omicron receptor-binding domain (RBD), totaling 24 microseconds in length, and combined PCA-based binding landscapes, covariance-derived interface fingerprints, and unbinding work requirements obtained through low-speed steered molecular dynamics to address this open question. Most complexes remained confined to a dominant bound basin, but preservation of the wild-type binding pose varied substantially across epitopes. NB21 and SB14 were clear outliers, sampling multiple low-occupancy states consistent with pronounced binding-pose plasticity. Across the panel, favorable contacts repeatedly converged on hydrophobic anchor patches within the receptor-binding motif centered near V445-F456 and F490-Y501, with nanobody-specific hydrogen bonds and salt bridges providing additional stabilization. Compared to wild-type, Omicron substitutions at 478, 484, 493, 498, and 501 rewired local contact networks, shifted bound orientations, and introduced mutation-adjacent repulsive couplings decreasing binding stability. Mechanical unbinding further ranked ACE2 as the most robust binder (~45 kcal/mol), whereas nanobodies required lower work to detach (~16-42 kcal/mol) with NM1230 and SB23 being the strongest binders among them. Our results suggest that Omicron escape is driven not only by loss of favorable interactions, but also by mutation-induced reorientation, broadened pose ensembles, and newly introduced repulsion (e.g. same charge repulsions). More broadly, this framework converts long all-atom molecular dynamics trajectories into compact, interpretable interaction fingerprints that pinpoint conserved anchoring features and mutation-sensitive liabilities, offering actionable design rules for broader, variant-resilient nanobodies.

Balanced Polymorphism Arising from Temporal Offsets in Temporally Varying Selection
Xavier Larason and Davorka Gulisija 

Abstract

 Genetic polymorphism plays a crucial role in adaptation to changing environments, yet the mechanisms that maintain polymorphism under temporally varying selection loci remain poorly explored. While previous theoretical models have traditionally assumed synchronous selection across loci or space, natural conditions often impose mismatches in the timing of selection between loci or across space. Here, we examine how timing mismatches in the onset of temporally varying selection across genetic loci and/or space influence the maintenance of genetic variation at one or two selected loci. Using computational simulations and a one or two-locus Wright-Fisher two-deme model, with and without conditions for maintenance of variation, we introduce phase shifts in selection between two loci and between two demes and evaluate their effects on polymorphism persistence across modes (duration and magnitude) of selection and migration rates. Our results show that timing mismatches modulate heterozygosity levels compared to synchronous selection. These findings suggest that varying selection offsets can act as a stabilizing force in varying environments, promoting the long-term maintenance of genetic diversity. This research provides novel insights into the evolutionary dynamics of genetic polymorphism and its implications for adaptation in changing environments. 

Modeling Cellular State Transitions: An RNA Velocity–Guided Neural Stochastic Framework for Forward Simulation
Bob Zhao

Abstract

Single-cell RNA sequencing (scRNA-seq) can represent each cell as a high-dimensional gene expression vector and enables computation of RNA velocity, which estimates short- term directional change. However, existing methods primarily infer fate probabilities rather than simulate explicit future trajectories. In this project, I model cellular transitions as a stochastic process in gene expression space and develop an RNA velocity–guided stochastic simulation framework to simulate forward evolution. Transitions are proposed using local similarity and velocity alignment, and accepted via a Metropolis–Hastings inspired criterion. Terminal states are treated as absorbing states. This approach generates explicit simulated trajectories from a single snapshot,shifting scRNA-seq analysis from static inference toward dynamic prediction of cellular evolution. 

Modeling Macrophage Polarization in Wounds: A Heterogeneity Perspective
Prateek Gupta and Doraiswami Ramkrishna

Abstract

Macrophages are specialized immune cells with very high plasticity and extreme diversity. Equipped with super-sensing capabilities, they can rapidly sample their milieu and polarize into functionally distinct subsets. This work studies their remarkable heterogeneity through the lens of a population balance framework and shows how a distribution of cell states emerges from intracellular metabolic dynamics. Under the postulate that each macrophage cell purposefully allocates its limited internal resources to drive the correct metabolic pathway, we recover phenotypic heterogeneity even under a fixed cytokine stimulus. This addresses a fundamental question: although LPS/IFNγ and IL4/IL13 are known to push macrophages toward opposing M1- and M2-like phenotypes, respectively, how do macrophages resolve conflicting cytokine cues in the wound milieu and adopt a phenotype? We further study the consequences of this postulate on the number densities of different macrophage subsets and show how they evolve as a wound resolves. We also demonstrate how macrophage populations use biochemical noise to their advantage and accelerate the process of wound healing. Finally, we reproduce experimental flow-cytometry trends, including the relaxation and convergence of FACS-sorted iNOS^low and iNOS^high macrophage subpopulations toward an overlapping distribution over time. The modeling framework presented herein is modular and extensible to any immune cell and can seamlessly integrate detailed immune profile data (e.g., scRNA-seq, flow cytometry etc.). Moving forward, such model features are desirable as we embrace the complexity of human disease in the single-cell era. 

Modeling and Control of Tumor–Immune Dynamics in a Nonlinear Systems Framework
Seyedehzahra Paylakhi and Robbee Wedow

Abstract

Understanding and controlling tumor–immune interactions is a central challenge in computational oncology. This study presents a computational framework based on Integral Sliding Mode Control (ISMC) applied to a nonlinear three-dimensional cancer model representing interactions among tumor cells, active immune cells, and resting immune cells. The approach is designed to investigate how control-theoretic strategies can be used to regulate tumor growth dynamics and enhance immune response within biologically realistic system behavior. The proposed framework aims to suppress tumor burden while promoting activation of immune cells and maintaining stable immune system dynamics. By incorporating ISMC, the model achieves robust regulation under system nonlinearities and uncertainties, which are characteristic of tumor–immune interactions. Comparative analyses with established control strategies, including proportional–integral–derivative (PID), Lyapunov-based control, and conventional sliding mode control, demonstrate improved stability and adaptability in regulating cell population dynamics. Simulation results illustrate that the ISMC-based approach enables precise modulation of tumor–immune trajectories and provides a flexible platform for exploring treatment strategies. This work highlights the potential of integrating control-theoretic methods into computational biology to study complex disease dynamics and to inform the design of adaptive therapeutic interventions. The framework can be extended to incorporate patient-specific parameters and experimental data, supporting future applications in personalized cancer modeling and treatment optimization. 

BMBC-Sim: An open source simulation package for coupled biomechanical and biochemical events
Michael Innerberger, Norma Perez Rosas, Ahmed El Hady, Stephan Preibisch and Kayvon Pedram. 

Abstract

 Biomechanical and biochemical phenomena are often coupled in biological systems. Examples include cell migration, muscle contraction, integrin-mediated adhesion, and nuclear mechanotransduction, among others. There are currently no open source biological simulation packages capable of modeling coupled biochemical and biomechanical events. Motivated by our interest to explore the influence of extracellular calcium dynamics on putative shape changes in extracellular spaces, we developed Biomechanics and Biochemistry Simulator (BMBC-Sim), an open source, Python-based simulation environment. BMBC-Sim enables users to define complex geometries, define material properties, define species and interactions among those species, and couple the chemical and mechanical events occuring within the system. This is achieved through delineation of partial differential equations (PDEs) that describe system behavior and solving of those PDEs using finite elements, implemented through NGSolve. BMBC-Sim interfaces with Astropy to handle physical unit conversions and with ParaView for visualization and plotting. BMBC-Sim is still under development, and below we showcase its capabilities so far by replicating modeling results from the calcium biology literature. 

Simulation-Guided Deep Learning for Target-Layer Oxygenation Inference from NIRS Signals
Jinho Park, Dohum Kim, Thien Nguyen and Amir Gandjbakhche 

Abstract

 Accurate estimation of deep tissue oxygenation is essential for understanding physiological processes and assessing conditions such as placental dysfunction. However, NIRS signals consist of mixed contributions from multiple tissue layers, making it challenging to accurately infer oxygenation of a specific target tissue layer. We propose a simulation-guided deep learning framework to infer oxygen saturation in a target tissue layer from multi-wavelength, multi-distance NIRS measurements. A two-layer tissue model was constructed to represent superficial and deep biological structures, and Monte Carlo simulations were used to generate a large-scale dataset under varying tissue thicknesses and oxygenation conditions, incorporating a multi-distance source–detector configuration (1–6 cm separations) and three wavelengths (730, 800, and 850 nm) to reflect realistic measurement settings. A one-dimensional convolutional neural network (1D-CNN) was trained to estimate target-layer oxygenation from optical measurements and superficial layer thickness. The model was trained using an adaptive optimization method and mean squared error loss, effectively capturing the nonlinear relationship between mixed optical signals and underlying physiological parameters, and demonstrating accurate and robust prediction performance across varying tissue conditions. This study shows that integrating physics-based simulation with data-driven modeling enables target-specific, depth-resolved inference of tissue oxygenation from non-invasive measurements. The proposed approach provides a scalable framework for solving inverse problems in biological signal analysis and has potential applications in non-invasive monitoring of placental oxygenation and other clinically relevant physiological processes. 

TwinCell: Large Causal Cell Model for Reliable and Interpretable Therapeutic Target Prioritisation
Jean-Baptiste Morlot, Yann Abraham, Elie Hatem and Thomaz Luscher Dias

Abstract

Drug development is a lengthy process requiring crucial decisions, starting with selecting the right target for a specific indication. Choosing the wrong target can cause significant delays or even failure. Therefore, target identification is a critical area in drug discovery, driving continuous computational biology advancements. However, challenges like cellular heterogeneity and disease context persist. Recent advances in single-cell and spatial molecular biology have generated massive datasets capturing compound effects, often used to train models for perturbation prediction. However, these models frequently fail to generalize or struggle to outperform simpler linear methods. We introduce TwinCell, a new approach that we define as Large Causal Cell Models. TwinCell leverages foundational single-cell models to learn the underlying biology from in vitro perturbational data. It can then predict which driver genes are most likely to be connected with disease-specific differentially regulated genes. To ensure a fair and accurate assessment, we also developed TwinBench, a new benchmark framework that specifically addresses the common “popularity bias” challenge found in recommendation algorithms. We applied TwinCell to both in vitro and in vivo datasets. Our results demonstrate superior performance compared to established methods and show strong generalization to new targets and cellular states not included in the model training. TwinCell and TwinBench represent a major advancement toward creating reliable, interpretable “virtual cells” for target identification, effectively bridging the data from high-throughput in vitro experiments with clinical insights. 

Multi-omics integrative clustering revealed three independent molecular subtypes of poor prognosis and drug resistance in acute myeloid leukemia
Flora Mikaeloff, Nona Struyf, Mattias Vesterlund, Huthayfa Mujahed, Trung Nghia Vu, Georgios Mermelekas, Albin Österroos, Anna Bohlin, Sofia Bengtzén, Yudi Pawitan, Rozbeh Jafari, Lukas Orre, Brinton Seashore-Ludlow, Janne Lehtiö, Päivi Östling, Sören Lehmann, Olli Kallioniemi and Tom Erkers 

Abstract

 Background: Acute myeloid leukemia (AML) is a prognostically unfavorable cancer due to its heterogeneity, significant relapse rate and drug resistance. We aimed to identify data-driven clusters to improve AML risk prediction and identify potential treatments.
Methods: AML patients (n =110) with complete RNA-seq, MS-proteomics, DNA methylation and mutational screening profiles were clustered using robust consensus clustering. Clusters were characterized by comparing clinical parameters, drug sensitivity from functional ex vivo drug testing data and finding cluster-specific pathways. Clusters were validated in external datasets.
Results: Four cancer subtypes (C1-C4) were identified based on muti-omics clustering. C4 showed the best prognosis, younger age and a favorable ELN compared to the other clusters. C1 and C2 had a bad prognosis which was partially explained by their mutation pattern : high-frequency mutation of NPM1-FLT3 in C1 and NPM1-FLT3-DNMT3A in C2. C1 was sensitive to venetoclax but resistant to FLT3 inhibitors due to the large proportion of immature cells and up-regulation of stemness markers and pathways. C2 had the opposite drug profile and showed a much larger proportion of differentiated and immune cells. C3 had the worst prognosis and had a large proportion of adverse ELN2017. C2 and C3 showed independent up-regulation of calcium pathways. Clusters were validated in external datasets. Interestingly, relapse samples showed an enrichment of C3. Drug combinations and predicted drugs were computed to suggest cluster-specific potential treatments.
Discussion/conclusion: We have identified omics subtypes that can be applied in other cohorts, suggest potential treatments and improve multi-omics precision medicine in the future. 

GeneSNAKE: a Python package for simulation of gene regulatory networks and perturbation-induced expression data
Erik Sonnhammer

Abstract

 Understanding how genes interact with and regulate each other is a key challenge in systems biology. One of the primary methods to study this is through gene regulatory networks (GRNs). The field of GRN inference however faces many challenges, which necessitate effective tools for evaluating inference methods. For this purpose, data that corresponds to a known GRN, from various conditions and experimental setups is necessary, which is only possible to attain via simulation. Today, most existing tools for GRN-based simulation are limited either in network or data properties, with few or no options to modify these properties.
To address these limitations we present GeneSNAKE, a Python package designed to allow users to generate biologically realistic GRNs and expression data for benchmarking purposes. GeneSNAKE improves on previous work by providing a unique combination of modules, allowing users to control a wide range of GRN and data properties. It provides full control of the noise level, several noise models, full control of the perturbation design, and a wide range of pre-defined perturbation schemes.
For benchmarking, GeneSNAKE offers a number of functions both for comparing network similarity, and properties in data and GRNs. These functions can further be used to study properties of biological data to produce simulated data with more realistic properties. GeneSNAKE is an open-source, comprehensive simulation and benchmarking package with powerful capabilities that are not combined in any other single package, and thanks to the Python implementation it can be extended and modified by users. 

Exploring Resource Constraints for Modeling Pairwise Microbial Metabolic Interactions in the Hydra-associated microbes 
Natchapon Srinak, Peter Deines, Christoph Kaleta and Jan Taubenheim 

Abstract

 Microbe-microbe metabolic interactions are primary determinants of community composition and ecosystem function. Despite advances in omics technologies, resolving the mechanistic basis of these interactions remains a significant challenge. In this study, we combined experimental and computational approaches to investigate the interactions of the microbial community associated with Hydra. We conducted cell-free supernatant (CFS) experiments to characterize pairwise interactions and their impact on microbial growth. Our experimental results revealed that Curvibacter-CFS generally inhibited the growth of most taxa except Pelomonas, while Acidovorax-CFS consistently promoted the growth of several species, including Pseudomonas, Undibacteria, and Duganella. To elucidate the underlying metabolic mechanisms, we performed in metabolic modeling simulations of both CFS environments and pairwise microbial communities, implemented with metabolic models using classical stoichiometric and Resource Allocation Constraints (RACs). Stoichiometric Flux Balance Analysis (FBA) omits enzymatic capacity, potentially leading to inaccurate predictions of nutrient consumption and byproduct secretion. Models using RACs are more difficult to reconstruct and rely on more assumptions, leading to potentially wrong simulation prerequisites. Here we compare both modeling approaches and their capability to predict microbial interactions and to recapitulate growth rates from a defined experimental system. This study provides a foundational step toward more reliable mechanistic modeling of complex microbe-microbe interactions, bridging the gap between simplified metabolic flux and the biological phenotype. 

A Mechanistic Modeling Framework Integrating Bayesian Optimization to Reveal Critical Mechanogenic Regulations in the Drosophila Wing Disc
Emerald Win 

Abstract

 Tissue morphogenesis emerges from coordinated biochemical and mechanical regulation, yet the quantitative rules linking these processes remain elusive. We used the Drosophila wing disc as a model system to develop an integrative platform combining experiments with computational modeling. Our framework couples Decapentaplegic (Dpp) signaling to intracellular Rho1 and Cdc42 dynamics through experimentally informed activation and inhibition motifs, including mutual antagonism and integrin-mediated basal activation. By leveraging Bayesian optimization to calibrate a multiscale reaction-diffusion model, we successfully reproduced in vivo spatial distributions of these mechanogens under both wild-type and mutant conditions. Our simulations identified mutual inhibition as a critical requirement for robust patterning during tissue development. This approach demonstrates how coupling mechanistic simulations with data-driven inference reveals the hidden regulatory logic of epithelial morphogenesis, offering a scalable strategy for modeling complex biochemical–mechanical feedbacks.

Integration of aged brain multi-omics reveals cross-system mechanisms underlying Alzheimer’s disease heterogeneity
Ricardo Vialle, Lucas Scheidemantel, Katia de Paiva Lopes, Chris Gaiteri, Vilas Menon, Philip De Jager, Julie Schneider, Aron Buchman, Yanling Wang, Shinya Tasaki, Roberto Raittz and David Bennett

Abstract

 The molecular correlates of Alzheimer’s disease (AD) are increasingly being defined by omics. Yet, the findings from different data types or cohorts are often difficult to reconcile. Collecting multiple omics from the same individuals allows a comprehensive view of disease-related molecular mechanisms, while addressing conflicting findings derived from single omics. Such same-sample multi-omics can reveal, for instance, when changes observed in the transcriptome share distinct but coordinated signals in epigenetics and proteomics, relationships otherwise unclear. Here, we apply a data-driven multi-omic framework to integrate epigenomic, transcriptomic, proteomic, metabolomic, and cell-type-specific population data from up to 1,358 aged human brain samples from the Religious Orders Study (ROS) and Rush Memory and Aging Project (MAP). We demonstrate the existence of sprawling cross-omics cross-system biological factors that also relate to AD phenotypes. The strongest AD-associated factor (factor 8) involved elevated immune activity at the epigenetic level, decreased expression of heat shock genes in the transcriptome, and disrupted energy metabolism and cytoskeletal dynamics in the proteome. We also showed immune-related factors (factors 2 and 3) with discordant enrichments, reflecting reactive-like glial subpopulations and protective contributions from surveillance microglia. Finally, unsupervised clustering of participants revealed eleven molecular subtypes of the aging brain, including three clusters strongly associated with AD but displaying distinct molecular signatures and phenotypic characteristics. Our findings provide a comprehensive map of molecular mechanisms underlying AD heterogeneity, highlighting the complex role of neuroinflammatory processes, and yielding potential novel biomarkers and therapeutic targets for precision medicine approaches to AD treatment. 

Towards a Catalogue of Protein Functional States for Dynamic Regulatory Network Modeling
Szabolcs Cselgő Kovács, Erzsébet Fichó, István Zoltán Reguly and Attila Csikász-Nagy

Abstract

 The decision-making capability of regulatory networks is embedded in the crosstalk of protein functional regions, which serve as logical switches that drive the downstream progression of cellular responses. Computational methods have demonstrated that PTM crosstalk patterns are conserved and decryptable. One of the most comprehensive databases is Reactome, which integrates multiple data sources into unified regulatory interactions. However, building genome-scale regulatory models remains challenging because data sources often provide binary interaction data, while Reactome compresses interacting entities into implicit complex assembly rules that are suitable only for qualitative modeling. Here, we propose a protein state catalogue database for Saccharomyces cerevisiae that reorients regulatory interactions from the perspective of the participating proteins and considers protein states as building blocks for model construction. We utilized Reactome V95 as a primary source to combinatorially extract complex assembly rules into a pool of explicit interactions and assigned structural annotations to each protein. The protein state catalogue contains protein state observations for 1322 yeast proteins that participate in 15 789 explicit alternative interactions across 194 regulatory pathways. Although 484 proteins have PTM site annotations, only 132 participate in interactions that lead to their state transitions. The state catalogue can serve as a valuable basis for systematically constructing regulatory models, training machine-learning models, and gaining a deeper understanding of protein functional crosstalk in broader cellular regulation mechanisms.
Supported by the URSP-CDP-24 University Research Scholarship Program – Cooperative Doctoral Programme of the Ministry for Culture and Innovation from the source of National Research, Development and Innovation Fund. 

A Scalable Framework for Comparing Centromere Sequence Similarity Using Alignment and k-mer-Based Methods
Christina Mulch

Abstract

 Centromeric regions are highly repetitive and structurally complex, limiting comparative analyses across genomes and samples. The advent of telomere-to-telomere (T2T) assemblies provides an opportunity to systematically investigate sequence similarity in these previously inaccessible regions. Here, we present a computational framework, currently under active development, for quantifying and comparing centromere-proximal sequence similarity across genomes and strains.
Our approach integrates alignment-based and k-mer-based strategies to capture both large-scale homology and fine-scale sequence variation. Centromere-proximal regions are defined using terminal genomic windows from chromosome assemblies, and all-vs-all alignments are performed using minimap2 with repeat-aware parameters. In parallel, we compute k-mer profiles using Jellyfish and estimate sequence similarity using Mash. We further identify chromosome- and strain-specific sequence signatures by quantifying unique k-mers within centromeric regions and comparing their distribution across samples.
Preliminary results in mouse (Mus musculus) demonstrate substantial shared sequence content among centromere-proximal regions, alongside distinct sets of unique k-mers that differentiate chromosomes and strains. These complementary signals enable the identification of conserved repeat structure as well as strain-specific divergence within centromeres.
This framework provides a scalable and extensible tool for comparative analysis of repetitive genomic regions across samples. Ongoing development focuses on improving computational efficiency, refining similarity metrics, and extending the approach to additional genomes and sequencing platforms. This work lays the foundation for systematic investigation of centromere variation, genome structure, and evolutionary dynamics in complex repetitive sequence contexts.

In silico comparison of three genome-scale models (GEMs) of Bacillus subtilis using Flux Balance Analysis (FBA)
Gabriel Granados

Abstract

 Abstract Genome-scale metabolic models (GEMs) are key tools for studying microbial metabolism through in silico simulations such as Flux Balance Analysis (FBA). Although several GEMs exist for Bacillus subtilis, they differ in structure, annotation quality, and predictive performance. This work presents the first curated repository compiling B. subtilis GEMs and introduces GEMcompare, a computational pipeline to prepare and compare metabolic models. Three GEMs (iYO844, iBB1018, and iBsu1147R) were evaluated through structural analysis, model quality assessment, FBA simulations under different carbon sources and comparisons with experiments from the literature. Additional comparisons included central carbon flux distributions and essential gene prediction. Also, a experimental validation in glucose, lactose and arabinose showed that models are useful to predict general in vitro behaviors but not precise ones. iYO844 provided the best predictions at higher glucose uptake rates, while iBB1018 showed the most conservative flux behavior, particularly in the TCA cycle. iBsu1147R exhibited better performance in essential gene prediction. Overall, this study enables researchers to find B. subtilis GEMs in a single place, compares three of them and establishes a comparison pipeline, and highlights the complementary strengths of existing GEMs for future systems biology and strain engineering applications. 

Modeling Alzheimer’s Disease Through Navigation Circuit Degeneration and Brain Rhythms
Pratham Balaji 

Abstract

Spatial disorientation is an early and debilitating symptom of Alzheimer’s disease (AD), yet the circuit-level mechanisms underlying navigational deficits remain poorly understood. I developed brain-inspired computational models of two core navigation systems: a ring attractor network of head-direction cells and a toroidal attractor network of grid cells. These continuous attractor networks maintain stable activity bumps that encode heading and position through structured recurrent connectivity. To simulate AD-related neurodegeneration, I progressively removed neurons and synaptic connections and quantified degradation in bump stability and decoding accuracy. Neuronal loss destabilized attractor states, producing systematic errors in direction and position estimates. To evaluate whether such circuit-level instability is reflected in human data, I analyzed resting-state EEG recordings from control and AD subjects. Using band-power features and logistic regression, I achieved 80% classification accuracy, with relative theta power at O2 emerging as the most informative feature. These findings suggest that disruption of continuous attractor dynamics may provide a mechanistic link between neurodegeneration and spatial disorientation and highlight potential EEG biomarkers of circuit instability in AD. 

Integrating single-nucleus RNA-seq with genome-scale metabolic models to predict dementia biomarkers
Seo Young Kim, Junhyeok Jeon and Hyun Uk Kim

Abstract

Dementia lacks accessible biomarkers that reflect disease mechanisms and support early diagnosis. In this study, we developed a systems biology framework to predict metabolite biomarkers for dementia by integrating single-nucleus RNA sequencing (snRNA-seq) data with genome-scale metabolic models (GEMs). Public and in-house postmortem brain datasets were analyzed to classify 1,977,975 high-quality nuclei into major brain cell types. Among the identified cell types, astrocytes, excitatory neurons, inhibitory neurons, microglia, and oligodendrocytes were selected for downstream analysis. Pseudo-bulk expression profiles were generated and integrated into the human generic GEM Human1 to reconstruct 3,622 patient-specific GEMs across Alzheimer’s disease (AD), Parkinson’s disease (PD), Lewy body dementia (LB), and normal controls. The cell-specific GEMs were simulated to quantify disease-associated metabolic activity. Significant metabolites were identified using the Wilcoxon rank-sum test, Spearman correlation analysis, and log2 fold change. A total of 831 metabolites initially showed significant alterations in the AD and PD+LB cohorts, with prominent disease-associated changes observed in microglia, astrocytes, and oligodendrocytes; these metabolites were further subjected to screening. Candidate biomarkers included betaine, spermidine, N-acetylgalactosamine, and S-adenosylmethionine for AD, and agmatine, arachidonate, sphingosine, spermidine, and hypoxanthine for PD+LB. These results show that integrating snRNA-seq with GEMs can systematically identify cell type-specific metabolic alterations and support biomarker discovery in dementia.

From Multi-modal Data to Cross-scale Virtual Kidney Modeling
Lan Jiang and Qifei Wang

Abstract

Current virtual cell research largely depends on dynamic perturbation data, which is often limited in scale, confined to a single modality, and overlooks cellular interactions—while obtaining such data from primary tissues remains technically challenging. To address these gaps, we constructed SHAPE, a dataset integrating over 12 million cells with paired transcriptomic and regulatory element profiles from the same cells, illuminating intracellular gene regulatory networks. Focusing specifically on kidney data within SHAPE, we developed PRISM-K, a cross-scale modeling framework. PRISM-K first builds a cell-scale foundation model by integrating paired multi-modal data to capture high-resolution regulatory logic. It then incorporates an interpretable individual-level encoder that integrates cellular context with systemic associations, generating a robust phenotypic space at the individual level. Finally, we introduce the Individual Health Deviation Score as a quantitative readout for in silico target screening. Together, PRISM-K establishes a cross-hierarchy digital twin that bridges molecular regulation, cellular behavior, and organismal states, offering a scalable strategy for modeling patient-specific responses. SHAPE and PRISM-K together lay the essential data and modeling groundwork for next-generation AI Virtual Cells and multi-scale “Virtual Individuals” in precision medicine.

Using mutual information-based network-inference algorithms to understand retinal regeneration in zebrafish
Gemma van der Hurk

Abstract

Retinal regeneration is a promising strategy for restoring vision in individuals affected by ocular disease. While humans cannot replace deceased visual neural cells, many other species, such as members of the teleost fish family, possess extraordinary regenerative capacity. The Teleostei Dario rerio model organism, zebrafish, shares a crucial retinal glial cell type with mammals, Müller glia, which responds to retinal injury in the former by deprogramming into a stem-like state then reprogramming into injured cell types, but instead forms a scar in mammals. The final reprogramming step is not well-understood, therefore further studies into the gene regulatory networks that program this process in zebrafish hold great promise for eventual transfer to the mammalian system. By applying network inference algorithms based on mutual information theory and transfer entropy to single-cell RNA sequencing data, alongside functional studies of identified target genes and cell-lineage tracing in zebrafish, this work aims to shed further light on this remarkable regenerative capability and contribute to the development of therapeutics.

Transcriptomic Profiling of Fanconi Anemia Subtypes Reveals Splicing and Expression Dysregulation
Hua Tan, Shivatheja Soma, Valer Gotea, Frank Donovan, Kinjal Bhadresha, Oliva Alston, Jeremy Amen, Arleen Auerbach, Agata Smogorzewska, Laura Elnitski and Settara Chandrasekharappa

Abstract

Fanconi anemia (FA) is a genetically and phenotypically heterogeneous disorder caused by mutations in at least 23 genes within the DNA repair pathway. Beyond their established roles in maintaining genomic stability, the broader molecular consequences of FA gene deficiency remain incompletely understood. In this study, we performed comprehensive transcriptomic profiling of 189 cell lines to characterize both gene expression and alternative splicing dysregulation in FA-deficient cells across ten FA subtypes. Global analyses revealed that transcriptomic patterns can distinguish tissue types and experimental conditions, while also uncovering both shared and distinct molecular features across FA genetic subtypes. Restoration of FA gene function led to widespread changes in gene expression and splicing, highlighting coordinated yet partially independent regulatory layers. Differentially expressed and alternatively spliced genes converged on key biological processes, including cell cycle regulation, extracellular matrix (ECM) organization, and apoptosis, while also showing pathway-specific distinctions, with expression changes enriched in immune and inflammatory signaling and splicing alterations enriched in RNA processing. Notably, FA subtypes exhibited both common and exclusive transcriptomic signatures, suggesting diverse functional contributions to disease pathology. These findings indicate that FA deficiency disrupts cellular homeostasis through combined effects on transcriptional and post-transcriptional regulation. Importantly, the results provide insight into potential therapeutic strategies, including targeting immune signaling, ECM remodeling, and splicing regulation. Overall, this study highlights the dual impact of gene expression and alternative splicing dysregulation in FA pathogenesis and offers a broader framework for understanding the molecular complexity of this disease.

Key Dates

April 9, 2026
Abstract submission deadline
May 5, 2026
Abstract acceptance notification
May 7, 2026
Late poster submissions deadline
May 14, 2026
Late poster acceptance notifications
Sunday-Thursday July 12-16, 2025
ISMB/ECCB conference
July 14, 2026
SysMod meeting

Registration and fees

Please register for the SysMod meeting through the ISMB conference registration .

Accomodation

ISMB 2026 hotel bookings will be handled directly through the Washington Hilton using special booking links. To help prevent fraudulent reservations, your personalized hotel link will be included only in your registration confirmation and your Checklist emails. A limited student room block is available at a discounted rate on a first-come, first-served basis. Once the student block is full, student attendees will need to use the standard booking link provided by the hotel.
Please register for the conference before booking your accommodations. .Please see the ISMB website for more information.

Standard accommodation rate $279 per night plus taxes
Student-only accommodation rate $175 per night plus taxes
Deadline for Discounted Rate: Friday, June 19, 2026

More information

For more information, please contact the SysMod coordinators 🔗.

2026 SysMod meeting @ ISMB

2026 SysMod annual meeting

Integrating systems biology and bioinformatics

July, 2026 | Washington, DC

Prof. Shayn Peirce-Cottler,Department of Biomedical Engineering,University of Virginia

Overview

Topics

Keynote speakers

Schedule

Poster presentation:

Key Dates

Registration and fees

Accomodation

More information

Prof. Shayn Peirce-Cottler,
Department of Biomedical Engineering,
University of Virginia