Drug repositioning offers an efficient route to discover new therapeutic indications for existing drugs. However, current computational drug repositioning models often face challenges related to data scarcity, heterogeneity, and therefore limited generalizability. To address these limitations, this study introduces DMAPLM, a multimodal pretrained framework for predicting drug-disease associations for further drug repositioning screening. DMAPLM leverages a lightweight dual-encoder architecture, utilizing ChemBERTa-2 for molecular encoding of drug SMILES strings and BioBERT for semantic encoding of multi-field disease texts. The framework explicitly aligns drug and disease representations through contrastive learning and employs attention-weighted pooling to emphasize informative molecular substructures. A Random Forest classifier is finally used for association prediction based on the enhanced multimodal features. We compile a new and comprehensive benchmark dataset for performance evaluation. Extensive experiments demonstrate that DMAPLM significantly outperforms six state-of-the-art baseline models, achieving an AUROC of 0.8919 and AUPR of 0.9116 under five-fold cross-validation, representing an improvement of up to 9%. Furthermore, DMAPLM exhibits robust performance in challenging cold-start scenarios, highlighting its practical utility for identifying novel drug-disease relationships. Case studies along with molecular docking analysis confirm the interpretability and biological meaningfulness of our predictions. Our study provides a powerful and interpretable approach for computational drug repositioning.
Cerebral hemodynamics is tightly regulated by arteriolar vasodynamics. In this study, a systems biology approach was employed to investigate how the interplay between passive, myogenic, neurogenic, and astrocytic responses shapes arteriolar vasodynamics in small rodents. A model of neurovascular coupling is proposed in which neurons inhibit and dampen the myogenic response to promote vasodilation during activation, and facilitate the myogenic response to promote rapid vasoconstriction immediately post-activation. In this model, inhibition of the myogenic response is mediated by the hyperpolarization of smooth muscle and endothelial cells. Dampening and facilitation of the response are mediated by neuronal production of nitric oxide and release of neuropeptide Y, respectively. We also introduce a model for gliovascular coupling, in which astrocytes periodically inhibit the myogenic response upon detecting an increase in myogenic activity through interactions between their endfeet and arterioles. Our simulations suggest that in the resting state, delays in myogenic autoregulation can intrinsically generate low-frequency (∼0.1 Hz) oscillations in vessel diameter (vasomotion), in the absence of extrinsic neurogenic or systemic rhythmic inputs. In the active state, these oscillations are disrupted by the neurogenic and astrocytic responses. The biophysical model of arteriolar vasodynamics presented in this study lays the foundation for quantitative analysis of cerebral hemodynamics for cerebrovascular health diagnostics and hemodynamic neuroimaging.
Ras proteins are prominent oncogenes, with KRas mutations found in approximately 80% of cancer cells harboring Ras mutations. The mechanism by which Ras mutations cause cancer remains unclear. Human Son of Sevenless (SOS) promotes the GDP-to-GTP exchange in the inactive GDP-bound Ras (RasGDP) by interacting with RasGDP conformation, thereby leading to the development of human cancer. Elucidating the Ras-SOS interaction mechanism can guide the drug design for Ras and SOS proteins. Based on our previously sampled special structure KRasGDP·Mg2+S1.2, this study constructs a functional ternary complex (KRasGDP·Mg2+)·SOS1·(KRasGTP·Mg2+). Furthermore, the KRas-SOS1 interactions regulated by the KRas G12D mutation and the SOS1 inhibitor BI-3406 that reportedly exhibits synergistic effects with G12D-mutant Ras inhibitors, are investigated through molecular dynamics (MD) simulations. The findings reveal that the G12D mutation and BI-3406 both affect the KRas-SOS1 interaction via the Switch-II (SW2) region of KRas. The negatively charged Asp12 has a repulsive effect on KRas, particularly on SW2, altering the interfacial electrostatic landscapes and diminishing the binding affinities by approximately 25 kcal/mol for both KRasGDP·Mg2+ and KRasGTP·Mg2+. BI-3406 forms a hydrogen-bond bridge between SW2 and SOS1 in wild type (WT) KRas, interrupting the interactions among the N-terminal residues of SW2 and SOS1. Moreover, BI-3406 was found here to attenuate the binding affinity of both WT and G12D-mutant KRasGDP·Mg2+ to SOS1, Interestingly, BI-3406 hardly affects the binding affinity of WT KRasGTP·Mg2+, while enhances the binding affinity of G12D-mutant KRasGTP·Mg2+. The change of binding affinity makes the catalytic pocket of SOS1 prefer to KRasGTP·Mg2+ and inhibits the growth of G12D-mutant KRas-driven tumors. These mechanistic insights provide valuable information for designing SOS1-co-targeting inhibitors to potentiate antitumor efficacy against G12D-mutated KRas.
Software containerization has become a cornerstone of modern computational biology, enabling researchers to package code, dependencies, and execution environments in portable and reusable units. Containers support reproducibility, facilitate collaboration, and lower barriers to deploying complex computational workflows across heterogeneous systems. At the same time, inappropriate or superficial use of containers can undermine these benefits, leading to brittle environments, security risks, or false confidence in reproducibility. In this article, we present nine practical and actionable tips for using software containers effectively in computational biology research. Rather than focusing narrowly on container syntax or tooling, we address conceptual decisions that arise throughout the research lifecycle: when containerization is appropriate, how to balance reproducibility with flexibility, how to manage dependencies and data, and how to share containers responsibly. These tips are intended for researchers with varying levels of experience, from those adopting containers for the first time to those maintaining mature, containerized workflows.
Prioritizing a reliable list of cancer-associated epigenetic regulators (cERs) is critical for cancer diagnosis and discovery of drug targets. While various cERs have been proposed to play important roles as cancer drivers, we anticipate that further cERs can be identified through computational analyses. In this study, we introduce a semi-supervised machine-learning approach based on tri-training model, termed Cancer-ASsociated Epigenetic Regulator identification (CASER). CASER integrates a wide range of multi-omics-derived features, including mutational, genomic, epigenetic, and transcriptomic data, to prioritize cERs as well as the four functional subtypes of cERs. When evaluated against an independent gene set, CASER demonstrates superior predictive performance compared to various other supervised machine-learning and deep semi-supervised models. CASER identified novel cERs that demonstrated cancer-driving potential and essentiality for cell survival. These novel cERs were comparable to established cancer driver genes and outperformed existing approaches for cER prediction. CASER identified dozens of novel cERs, of which six candidate cERs were shown to have roles in altering cell proliferation in four cancer cell lines. Furthermore, the prioritized cERs, particularly dual-role cERs, are more associated with anti-cancer medicines, underscoring their potential as therapeutic targets in cancer. Our study can offer valuable insights of cERs for future functional studies, advancing the understanding of their role in cancer biology.
Multiplets arise when multiple cells are captured within the same droplet during single-cell sequencing, producing hybrid molecular profiles that can distort downstream analyses. Detecting multiplets in single-nucleus ATAC-seq (snATAC-seq) data is particularly challenging due to the sparsity and overdispersion of chromatin accessibility measurements. Moreover, computational approaches that jointly leverage evidence across multiple features and data modalities are highly desirable for multiplet detection. We introduce SEBULA, a semi-parametric empirical Bayes framework for multiplet detection in snATAC-seq data. SEBULA models the singlet background directly from observed chromatin accessibility signals using fragment-level information from snATAC-seq data. This approach avoids reliance on synthetic doublets and produces classification probabilities that enable direct false discovery rate control. We further extend SEBULA to integrate complementary evidence from additional features and modalities, such as simultaneously measured gene expression profiles. Across simulations and seven multimodal datasets with hashing-based ground truth, SEBULA demonstrates improved sensitivity and specificity compared with existing snATAC-seq methods. The evidence integration framework achieves comparable or superior performance relative to state-of-the-art multiomic approaches while maintaining computational efficiency.
Histone deacetylases (HDACs) regulate neuroprotection; however, Trichostatin A (TSA), an HDAC inhibitor, lacks clear molecular mechanisms and core targets in Alzheimer's disease (AD), limiting clinical translation. This study aimed to decipher TSA's AD-regulating network, screen core genes, and support AD early diagnosis and multi-target therapies. TSA targets were computationally predicted. Five GEO AD datasets were analyzed for differential genes and core modules, and 130 machine learning algorithms were employed to identify core genes. Functional annotation, immune cell analysis, and single-cell expression profiling were conducted. Molecular docking and 100 ns molecular dynamics simulations verified TSA-protein interactions. 949 potential TSA targets were identified, overlapping with AD differential genes and enriching key pathways such as GABAergic synapse and tau phosphorylation. Eight machine learning-identified core genes (EFNA1, GABRB2, GABARAPL1, EGR1, CDK5, KCNC2, MET, GRIA2) exhibited a distinct AD expression pattern: synergistic downregulation of protective genes and unique upregulation of pathological EFNA1. These genes are implicated in neurotransmission, synaptic plasticity, tau clearance, and immune-neural crosstalk. Molecular dynamics simulations suggested TSA may not stably bind these candidates, implying its regulation relies on epigenetic mechanisms via HDAC1-3/6 inhibition, potentially restoring gene network balance and disrupting neuroinflammation-neurodegeneration cycles. Complex regulatory modes and cell type-specific expression were also observed. This study provides preliminary insights into TSA's putative mechanisms in AD intervention, highlighting the eight candidate core genes' potential diagnostic and therapeutic value as AD biomarkers, supporting TSA's multi-target therapy. All findings are computationally derived and require experimental verification.
Kinetic metabolic models provide invaluable insights into cellular metabolism, supporting applications in synthetic biology, metabolic engineering, and systems biology. However, reproducibility and utility of these models hinge on clear and rigorous documentation, standardized annotation, and accessible visualization. This paper presents a workflow for building, annotating, visualizing, and sharing kinetic metabolic models. Our method integrates community standards and open-source tools to ensure reproducibility, interoperability, and user accessibility. This procedure enables researchers to produce reusable and well-documented kinetic models, advancing their role as powerful tools in metabolic research.
Agent-based modeling (ABM) is a powerful tool for understanding self-organizing biological systems, but it is computationally intensive and often not analytically tractable. Equation learning (EQL) methods can derive continuum models from ABM data, but they typically require extensive simulations for each parameter set, raising concerns about generalizability. In this work, we extend EQL to Multi-experiment equation learning (ME-EQL) by introducing two methods: (i) one-at-a-time ME-EQL (OAT ME-EQL), which learns individual models for each parameter set and connects them via interpolation, and (ii) embedded structure ME-EQL (ES ME-EQL), which builds a unified model library across parameters. We demonstrate these methods by learning continuum models from a noisy birth-death mean-field model and from an on-lattice agent-based model of birth, death, and migration with spatial structure, often used to investigate cell biology experiments. We show that both methods significantly reduce the relative error in recovering parameters from agent-based simulations, with OAT ME-EQL offering better generalizability across parameter space. Our findings highlight the potential of equation learning from multiple experiments to enhance the generalizability and interpretability of learned models for complex biological systems.
Glomerulonephritis (GN) is an immune-mediated kidney disorder that causes glomerular injury, progressive renal dysfunction, and end-stage kidney disease. Traditional treatments such as corticosteroids and immunosuppressants are limited by variable efficacy and severe adverse effects, highlighting the need for novel therapeutic targets and personalized strategies. We performed a systematic multi-omics Mendelian randomization (MR) analysis applying established proteomic and transcriptomic quantitative trait loci (pQTL/eQTL) resources to genome-wide association studies (GWAS) of four GN subtypes: acute, chronic, IgA nephropathy, and membranous nephropathy. Bayesian colocalization was used to strengthen causal inference, while independent replication and meta-analysis were conducted using the FinnGen cohort. Mouse knockout phenotypes, drug reposition, and computational pharmacology algorithm were applied to evaluate translational potential. Proteomic-wide MR revealed MTR as protective in chronic GN and HCK as a risk factor for membranous nephropathy, whereas CD302 and CDKN1B showed protective effects. Transcriptomic-wide MR identified candidate genes across GN subtypes: RECQL, BRSK2, and MGP in acute GN; AFM, CFHR5, and EPHB2 in chronic GN; IL6R, MBL2, and PRSS3 in IgA nephropathy; and TIMP4, HCK, and PEAR1 in membranous nephropathy. Bayesian colocalization analysis provided strong support for shared causal variants (PPH4 > 0.8) for HCK, CD302, TIMP4, PEAR1, PARP1, and FHIT. Replication and meta-analysis in the FinnGen cohort provided additional consistency across datasets, while downstream translational annotations highlighted IL6R, MBL2, C5, and CD55 as potential hub targets within immune and complement-related pathways. This integrative multi-omics study provides novel insights into the genetic architecture and therapeutic landscape of GN, identifying potential therapeutic targets that may inform precision nephrology and drug repurposing. Notably, most targets supported by colocalization, mouse knockout phenotypes, and drug repurposing evidence were predominantly identified in membranous nephropathy, suggesting a particularly tractable genetic and therapeutic architecture for this subtype.
Long-term transplant success is limited by allograft rejection, a complex process traditionally studied on an organ-specific basis. To establish a unified framework beyond organ-specific studies, we performed a network-based systems biology analysis of transcriptomic data from 672 liver, kidney, and heart transplant biopsies to identify a conserved, pan-organ molecular framework of rejection. By constructing and comparing organ-specific gene co-expression networks, we identified a consensus, six-module immune cascade that captures the hierarchical nature of the alloimmune response. In addition, we also uncovered a highly conserved 24-gene cell cycle signature consistently upregulated in rejecting allografts, implicating cellular proliferation as a core feature of rejection pathology. From this framework, we derived a 172-gene immune signature and applied machine learning models to assess its predictive performance, achieving accuracy comparable to established benchmarks. We further refined this to a minimal, high-performance 20-gene immune signature (AUC > 0.96). Both the immune and cell cycle signatures demonstrated robust, pan-organ utility when independently validated in a lung transplant cohort (n = 243). Collectively, these findings define a pan-organ molecular framework for rejection and highlight cell cycle dysregulation as a conserved hallmark, offering a foundation for standardized, cross-organ diagnostic platforms to improve allograft surveillance and patient outcomes.
Advances in functional genomic technology, notably CRISPR using Cas9 or Cas12, now allow for large-scale double perturbation screens in which pairs of genes are inactivated, allowing for the experimental detection of genetic interactions (GIs). However, as it is not possible to validate GIs in high-throughput, there is no gold standard dataset where true interactions are known. Hence, we constructed a Double-CRISPR Knockout Simulation (DKOsim), which allows users to reproducibly generate synthetic simulation data where the single gene fitness effect of each gene and the interaction of each gene pair can be specified by the investigator. We adapted Monte-Carlo randomization methods to extend single knockout simulation methods to double knockout designs, which simulate the gene-gene interactions between all possible combinations of the input genes. Using DKOsim, we generated simulated datasets that closely resemble real double knockout CRISPR datasets in terms of Log Fold Change (LFC), GI distribution, and replicate correlation. We further inferred optimal CRISPR library designs by systematically investigating critical experimental parameters including depth of coverage, guide efficiency, and the variance of initial guide distribution. This simulation scheme will help to identify optimal computational methods for GI detection and aid in the design of future dual knockout CRISPR screens.
Humans just don't fall asleep like a log - or step-function. Rather, the sleep-onset period (SOP) exhibits dynamic and non-monotonous changes of electroencephalogram (EEG) with high, and so far poorly understood, intra- and inter-individual variability. Computational models of the sleep regulation network have suggested that the transition to sleep can be viewed as a noisy bifurcation at a saddle node which is determined by an underlying control signal or "sleep drive". However, such models do not describe how internal control signals in the SOP can produce rapid switches between stable wake and sleep states, nor how these state-space changes are translated in the macroscopic EEG. Here, we propose a minimally-parameterized stochastic dynamical model, in which one slowly-varying control parameter drives the wake-to-sleep transition while exhibiting noise-driven bistability. We provide a procedure for estimating the parameters of the model given single observations of experimental sleep EEG data, and show that it can reproduce a wide variety of SOP phenomenology. Using the model to analyze a pre-existing sleep EEG dataset, we find that the estimated model parameters correlate with subjective sleepiness reports. These results suggest that the bistable characteristics of the SOP can serve as biomarkers for tracking intra- and inter-individual variability of sleep-onset disorders.
Humans and other animals learn the value of candidate actions by interacting with their environment, which invariably requires the exertion of effort. Dopamine has been implicated in both effort and reward learning, but little is known about how these processes interact. In this double-blind study, healthy young adults (N = 42) were randomized to receive either high-dose sulpiride (a post-synaptic D2-receptor antagonist) or placebo. Participants then completed a novel two-armed bandit task, in which they weighed the effort costs associated with each option against their expected rewards. Overall, learning accuracy was lower on sulpiride compared to placebo. Computational modeling revealed that this was driven by the capacity of effort to significantly modulate learning rates on placebo but, critically, not on sulpiride. Simulations showed that the capacity of effort to modulate learning rates plays an adaptive role by improving performance in agents whose learning would otherwise be compromised by low motivation. Together, these data provide causal evidence that dopamine supports the relationship between effort and learning, and reveal a novel role for dopamine in shaping how humans learn from the consequences of their actions.
Neural synchronization is central to cognition. However, incomplete synchronization often produces chimera states, where coherent and incoherent dynamics coexist. Recent studies have suggested that these chimera states could be important in human cognitive organization. In particular, chimera states have been suggested as a regulator of cognitive integration and regulation with varying quality as humans age. While previous studies have explored such chimera states using networks of coupled oscillators, it remains unclear why neurons commit to communication or how chimera states persist. Here, we investigate the coevolution of neuronal phases and communication strategies on directed, weighted networks where interaction payoffs depend on phase alignment and may be asymmetric due to unilateral communication. The graph structure enables us to apply a game-theoretic model of Kuramoto-like oscillators to brain connectomes, and the asymmetry captures biochemical differences between communicative and non-communicative neurons. Combined, these two generalizations enable us to apply the computationally-tractable game-theoretic model of Kuramoto models to realistic brain networks and analyze the role of connectome structure on neuron communication. We find that both connection weights and directionality influence the stability of communicative strategies-and, consequently, full synchronization-as well as the strategic nature of neuronal interactions. Applying our framework to the C. elegans connectome, we show that emergent payoff structures, such as the staghunt game, control population dynamics. We demonstrate that weighted, directed connectivity in the Caenorhabditis elegans (C. elegans) connectome is sufficient to generate robust chimera states modulated by payoff asymmetries. Our computational results demonstrate a promising neurogame-theoretic perspective, leveraging evolutionary graph theory to shed light on mechanisms of neuronal coordination beyond classical synchronization models.
Computational models of macromolecules have many applications in biochemistry, but physical inaccuracies limit their utility. One class of models uses energy functions rooted in classical mechanics. The standard datasets used to train these models are limited in diversity, pointing to a need for new training data. Here, we sought to explore a new paradigm for training an energy function, where the Rosetta energy function was used to design de novo proteins. Experimental results on these designs were then used to identify failure modes of design, which were subsequently used as a "guiding principle" to retrain the energy function. Specifically, we examined a diverse set of de novo protein designs experimentally tested for their ability to stably fold, identifying unstable designs that were predicted to be stable by the Rosetta energy function. Using deep mutational scanning, we identified single amino-acid mutations that rescued the stability of these designs, providing insight into common failure modes of the energy function. We identified one key failure mode, involving steric clashing in protein cores. We identified similar overpacking when using Rosetta to refine high-resolution protein crystal structures, quantified the degree of overpacking, and refit a small set of energy-function parameters to better recapitulate native-like packing. Following fitting, we largely eliminated the failure mode in the refinement task, while retaining performance on other benchmarks, resulting in an updated version of the Rosetta energy function. This work shows how learning from protein designs can guide energy-function development.
Adaptive behavior depends on the brain's capacity to vary its activity across multiple spatial and temporal scales. Yet, how distinct facets of this variability evolve from childhood to older adulthood remains poorly understood, limiting mechanistic models of neurocognitive aging. Here, we characterize lifespan neural variability using an integrated empirical-computational approach. We analyzed high-density EEG cohort data spanning 111 healthy individuals aged 9-75 years, recorded at rest and during a passive and an attended auditory oddball stimulation task. We extracted scale-dependent measures of EEG fluctuation amplitude and entropy, together with millisecond-resolved phase-synchrony networks in the 2-20 Hz range. Multi-condition partial least squares decomposition analysis revealed two independent lifespan trajectories. First, slow-frequency power, variance, and complexity at longer timescales declined monotonically with age, indicating a progressive dampening of low-frequency fluctuations and large-scale coherence. Second, the temporal organization of phase-synchrony reconfigurations followed an inverted U-shaped trend: young adults exhibited the slowest yet most diverse switching-characterized by low mean but high variance and low kurtosis of jump lengths at 2-6 Hz, and the opposite pattern at 8-20 Hz-whereas children and older adults showed faster, more stereotyped dynamics. To mechanistically account for these patterns, we fitted a ten-node phase-oscillator model constrained by the human structural connectome. Only an intermediate, metastable coupling regime qualitatively reproduced the empirical finding of maximally heterogeneous synchrony dynamics observed in young adults, whereas deviations toward weaker or stronger coupling mimicked the children's and older adults' profiles. Our results demonstrate that development and aging entail changes in the switching dynamics of EEG phase synchronization by differentially sculpting stationary and transient aspects of neural variability. This establishes time-resolved phase-synchrony metrics as sensitive, mechanistically grounded markers of neurocognitive status across the lifespan.
Identifying cancer driver genes (CDGs) remains a central challenge in cancer genomics, as frequency-based mutation approaches often miss rare but functionally important regulators. We present PICDGI, a computational framework that predicts driver-like regulatory genes by integrating dynamic gene-gene interaction modeling with single-cell RNA sequencing (scRNA-seq) data. Rather than relying on DNA mutation calls, PICDGI infers functional driver activity from time-resolved expression patterns and latent regulatory influence among genes during tumor progression. Methodologically, PICDGI employs a time-varying state-space model with variational Bayesian inference and Markov Chain Monte Carlo (MCMC) sampling to estimate evolving gene interaction effects. The posterior distributions capture both the magnitude and uncertainty of each gene's inferred regulatory influence. From these, PICDGI derives a driver coefficient that quantifies the strength and reliability of each gene's contribution to progression-associated expression dynamics, enabling the prioritization of impactful regulators over neutral passengers. Applied to lung adenocarcinoma (LUAD) scRNA-seq data, PICDGI recovered known oncogenes and tumor suppressors and nominated novel candidate drivers, including JPH1 and CHEK1, which are implicated in calcium signaling, mitochondrial regulation, and DNA repair. These genes exhibit trajectory-aligned activity consistent with tumor evolution and immune-modulatory processes. Comparative analysis using Moran's I statistics in Monocle 3 showed that PICDGI-prioritized genes display stronger progression-associated dynamics than genes selected by spatial autocorrelation alone. We further validated PICDGI on an independent pediatric acute myeloid leukemia (AML) scRNA-seq cohort, where it consistently recovered known drivers and relapse-associated regulatory programs under fixed model parameters. By integrating interaction-informed dynamic modeling with single-cell resolution data, PICDGI provides a generalizable and biologically grounded framework for identifying rare and context-specific regulatory drivers of cancer progression, with broad applicability across tumor types.
Accurate prediction of B-cell epitopes plays a key role in facilitating advancements in vaccines, therapeutics, and diagnostics. In contrast to labor-intensive experimental approaches, computational strategies provide a more economical and efficient means of identifying potential epitopes. Existing methods are often limited by their reliance on experimentally resolved protein structures or by the use of lower-accuracy predicted structures. Sequence-based approaches, while fast, largely fail to capture the 3D spatial context essential for conformational epitopes. With the breakthroughs achieved by AlphaFold3 in predicting protein structures, we present MsgaBpred, the model to apply AlphaFold3-derived structures to B-cell epitope identification. Given only a protein sequence, our model employs a multi-scale graph convolutional network and additive attention to capture complex structural dependencies without relying on experimentally determined structures. The multi-scale design allows for effective modeling of both local and global contexts by aggregating information across different neighborhood ranges. Additionally, we leverage ESM-C, a more expressive protein language model than ESM-2, to enhance feature representation for B-cell epitope prediction. Extensive evaluations across multiple benchmark datasets demonstrate that MsgaBpred achieves competitive and robust performance; notably, it yields a statistically significant improvement in AUC compared to existing state-of-the-art methods. Moreover, the modular and scalable architecture of MsgaBpred holds promise for broader applications, including the structural analysis of other biomolecular entities such as nucleic acids and carbohydrates.
Intracellular calcium ions (Ca2+) exhibit diverse dynamical behaviors linked with cellular physiological states related to health and disease. While deterministic models predict how biochemical parameters create distinct dynamical regimes - steady states, oscillations, bursting, chaos, and multiple periodicity - real biological systems are inherently stochastic due to finite molecular populations. Previous studies using conventional statistical measures demonstrated that increasing intrinsic fluctuations render these dynamical states increasingly indistinguishable, particularly for chaotic and multiple-periodicity patterns. This raises whether parameter-dependent organizational principles persist under realistic noise levels to remain biologically meaningful and computationally detectable. We address this using a large-kernel convolutional neural network (LKCNN) designed to capture global dynamical features across noise levels. Using chemical Langevin equations to generate synthetic training data with realistic intrinsic fluctuations, the LKCNN achieves ~90% accuracy in classifying eight distinct dynamical states despite noise levels that visually obscure distinctions. Validation with experimental Ca2+ data from pancreatic [Formula: see text]-cells as well as other cells, including WT-HEK293, STIM-KO, and ORAI TKO, achieves 96.8% accuracy, confirming generalizability beyond synthetic datasets, substantially outperforming conventional baselines (Support Vector Machine and Random Forest), which achieve only 54.0% and 51.6% accuracy respectively on the same experimental data. These results demonstrate that deterministic organizational signatures persist through realistic biological noise, suggesting parameter-dependent dynamical structures represent robust principles governing cellular function. Our findings establish that sophisticated pattern recognition can bridge theoretical deterministic dynamics and noisy biological reality, offering a framework for extracting meaningful dynamical information from inherently stochastic oscillatory biological processes.