What this is
- (quanTifiEd immuNe-aging dySregulation index) quantifies immune-aging dysregulation from pathway enrichment analyses.
- It uses a Large Language Model to classify enriched pathways into five categories related to immune aging.
- provides a normalized score that reflects the magnitude and distribution of aging-associated processes.
Essence
- transforms pathway enrichment outputs into a quantitative measure of immune-aging dysregulation, enabling comparative analysis across biological contexts.
Key takeaways
- aggregates pathway signals into a single dysregulation score, facilitating easier interpretation and comparison of immune aging dynamics across datasets.
- Distinct dysregulation profiles were observed in various conditions, with Alzheimer's disease modules showing high inflammaging signatures, while radiation response datasets were dominated by DNA damage signals.
- The framework demonstrated high reproducibility and agreement between LLM-derived annotations and human consensus scores, indicating its reliability in assessing immune-aging dysregulation.
Caveats
- 's framework is based on selected biological processes, which may not capture all aspects of aging-related dysregulation.
- The model's validation was limited to a small selection of reference conditions, necessitating further testing across diverse biological scenarios.
- LLM stochasticity may introduce variability in values, which could affect reliability without broader human consensus validation.
Definitions
- TENSE: A framework for quantifying immune-aging dysregulation by summarizing pathway enrichment results into a single score.
- DIRES: A scoring scheme categorizing biological processes related to immune aging: DNA damage, DNA repair, epigenetic drift, inflammaging, and nucleic acid sensing.
Simplified
Introduction
Temporal accounting of biological processes is an essential aspect of system coordination for living organisms, from the cellular level (Gliech and Holland, 2020). Biological time for an organism can be defined as a set of iteratively organized biological processes that maintain cellular and organismal homeostasis throughout its lifespan. Biological time is inescapably linked with physical time (Lestienne, 1988). As organisms progress along physical time, the fidelity of biological processes declines in the human body, defining the process of aging (Tenchov et al., 2024). The loss of biological process fidelity reflects the decay of all cellular systems, from nuclear processes to cellular metabolism and transcellular communication (López-Otín et al., 2023).
One of the more prominent drivers of aging is the progressive accumulation of DNA damage, reflecting a decline in the DNA damage response efficiency (Gorbunova et al., 2007). However, declining repair efficacy leads to the accumulation of additional damage in a feed-forward manner (Schumacher et al., 2021). As a central process in aging, unmitigated DNA damage initiates another feedforward process: inflammation mediated by nucleic acid sensors. This induction, along with potential causes of dysregulated mitigation processes, drives age-related inflammation (Zhao et al., 2023). The loss of genomic stability due to the failure of repair mechanisms thus marks a salient point of no return, where time-dependent disorder prevails over its mitigation (Laffon et al., 2021; Vavougios et al., 2022; López-Gil et al., 2023).
In addition to the idea of inflammaging (Fülöp et al., 2019), these concepts suggest an immune arrow of time—a biological clock that indicates the gradual decline in the accuracy of essential homeostatic processes, observed, as increasing inflammatory tonicity.
The implied parallel between additive disorder in biological systems and aging can also be used to describe the critical role of inflammation in the maturation and development of stem cells (Tatullo, 2024), as well as the critical role of the inflammatory milieu for tissue-specific processes such as synaptogenesis and neurogenesis (Paul et al., 2021). In short, the immune arrow of time may also accelerate purposefully due to physiological processes, as well as secondary to pathophysiological conditions.
An informative approach to identifying systems-level dysregulation is transcriptomic profiling to extract disease-associated gene signatures (Ganesan et al., 2025; Moore et al., 2025; Chen et al., 2026) and the corresponding enriched pathways (Liu et al., 2026; Wang et al., 2026; Song et al., 2026). Gene signature methods quantify the coordinated activity of gene networks, generating weighted scores that can predict outcomes in specific disease contexts (Sparks et al., 2024). Similarly, pathway enrichment analyses identify biological processes that are statistically overrepresented within a dataset, enabling the detection of pathways potentially involved in the underlying biological condition (Vavougios et al., 2022). Similarly, current state-of-the-art applications of LLMs focus on extracting knowledge from gene sets and annotating their function (Wang et al., 2025). While these approaches provide valuable insights into molecular activity and pathway involvement, they primarily address specific research questions within defined biological contexts. Consequently, a methodological gap remains for metrics capable of integrating system-level genomic signals into a quantitative estimate of the magnitude and distribution of dysregulation across predefined biological axes.
In this study, we propose an alternative strategy for annotating pathway enrichment results to extract biologically meaningful information on dysregulation and disease-associated processes. As a proof of concept, we introduce the TENSE (quanTifiEd immuNe-aging dySregulation indEx) scoring framework, designed to quantify immune aging–associated regulatory disorder in biological systems. TENSE functions as an integrative, normalized, systems-level metric that summarizes pathway-derived dysregulation into a single interpretable index, enabling comparative assessment of immune-aging dynamics across datasets and experimental contexts.
TENSE is designed to transform pathway enrichment outputs into a structured representation of immune-aging dynamics by mapping enriched pathways to a set of mechanistically defined biological axes. These axes correspond to key processes implicated in immune aging, namely DNA damage (D), DNA repair (R), epigenetic drift (E), inflammaging (I), and nucleic acid sensing (S), which together form the DIRES scoring scheme. Recent advances in artificial intelligence and computational biology have enabled increasingly sophisticated interpretation of high-dimensional data. These applications include machine learning–based modeling, pathway analysis, and network-level integration approaches. Broadly, these developments illustrate the capacity of AI-driven methods to extract structured insights from complex datasets across diverse scientific domains (Raza and Hanif, 2026; Raza et al., 2026; Hanif et al., 2026; Raza et al., 2025a; Raza et al., 2025b).
Large language models (LLMs) are increasingly integrated into bioinformatics workflows, including assisting with knowledge curation of the Reactome database (Wu et al., 2025). LLMs have been shown to retrieve semantic knowledge to predict regulatory relationships based on provided gene networks, albeit with varying performance (Azam et al., 2024).
In the TENSE workflow, we capitalize LLM capabilities for semantic analysis and classification of enriched pathways into mechanistic aging axes; we then employ them to compute a systems-level dysregulation metric. The LLM core serves as a semantic classifier within this framework, rather than functioning as a biological database. It interprets pathway descriptions and categorizes them into established mechanistic groups: DNA damage, DNA repair, epigenetic drift, inflammaging, and nucleic acid sensing. Pathway nomenclature often contains biological significance that is not consistently captured by simple keyword matching or static database mapping. For instance, pathways such as “HDR through MMEJ” or “OAS antiviral response” necessitate contextual interpretation to appropriately associate them with DNA repair or nucleic acid-sensing processes. The integration of an LLM semantic processing layer, GPT-4.1-NANO, enables scalable and reproducible semantic classification of enriched pathways while retaining mechanistic clarity. It is important to note that the LLM layer GPT-4.1-NANO does not generate biological conclusions; its role is limited to pathway annotation, while the TENSE metric is calculated according to defined quantitative rules. This rigidity also serves to develop TENSE in a relatively controlled data environment while the intended core functionalities are assessed.
TENSE combines the magnitude of enrichment and the pathway contributions along each axis to deliver a normalized quantitative assessment of immune-aging–related dysregulation. With this approach, pathway enrichment outcomes are viewed not simply as lists of significant pathways, but as an integrated metric that represents both the extent and spread of biological disruption among aging-associated processes.
We explore TENSE scoring by reanalyzing relevant published studies and procuring several cases, with a primary focus on Alzheimer’s disease. We demonstrate the utility of the TENSE and its associated measures in annotating pathway enrichment data and the biological context of each studied condition.
Methods
Theoretical framework
To define TENSE, we consider factors that drive the accumulation of disorder in a biological system as a function of age, herein a cell or tissue, building toward incremental genomic instability (Lemus et al., 2026). For the development of this model, major drivers of aging, adjacent to genomic instability, were considered and modeled as dimensionless variables; Specifically, we considered the equilibrium between DNA damage and repair (Gorbunova et al., 2007; Schumacher et al., 2021; Copp et al., 2023; Vlachogiannis et al., 2023; Whittemore et al., 2019), nucleic acid sensing mechanisms (Paul et al., 2021; Stillman et al., 2024; Dorrity et al., 2024; Hatch and Hetzer, 2015; Miller et al., 2021; Taffoni et al., 2021), inflammaging (Franceschi et al., 2017; Fulop et al., 2023) and epigenetic drift (Bertucci-Richter and Parrott, 2023; Issa, 2014; Bertucci-Richter et al., 2024; Ma et al., 2018).
Although the relationships among these biological processes are complex and involve multiple feedback and regulatory interactions, for the purpose of computing TENSE, their contributions to cumulative dysregulation are modeled as additive. This assumption reflects both the simplifying representation of system-level dysregulation and the statistical framework of pathway enrichment analysis, in which biological pathways are evaluated as independent enrichment events (Nguyen et al., 2019). Consequently, enriched pathways are treated as independent proxies for the underlying biological processes implicated in disease pathogenesis.
Thus, we define the Immune-Entropy Quotientas a function of aging-related processes: H t ( )
Where:
The normalized Immune-Entropy quotient would then be calculated with the following formula, when: D t I t R t E t S t ( ) , ( ) , ( ) , ( ) , ( ) is 1
To calculate normalized H(t) from enrichment datasets, we replaceproxy scores derived from gene set overrepresentation analyses. To reformulate the function using pathway proxies, we consider the following selection criteria: D t I t R t E t S t ( ) , ( ) , ( ) , ( ) , ( )
Where:
We again constrainby applying a 1/k for each D, I, R, E, S count, to account for multiple memberships in one of the five DIRES categories, leading to standardized values between 0 and 1 for the calculation of TENSE. For example, a pathway scoring in D, R, and S would have a count of 1. n i
To provide a standardized estimate of the Entropy Quotient, henceforth the TENSE metric, we divide by. E ¯ D E ¯ I E ¯ R E ¯ E E ¯ S + + + +
To obtain normalized TENSE, producing a range of values between 0 and 1, we divide by the maximum TENSE = 5. Implementation of TENSE and DIRES scoring via an LLM-powered KNIME workflow.
To calculate TENSE in an automated manner, a hybrid semantic – deterministic workflow is designed based on the previously presented theoretical framework and implemented via an LLM-powered KNIME analytical workflow.1 The overall architecture uses pathway enrichment analysis via Enrichr to produce combined enrichment scores that adjusts for potential bias in favoring larger length gene sets in gene set libraries (Chen et al., 2013). Thus, Enrichr combined scores provide weighted contributions to pathways by incorporating both statistical significance and enrichment magnitude. Notably, while Enrichr combined scores provide a useful measure of pathway relevance within a given analysis, they are not inherently comparable across datasets. The TENSE framework addresses this limitation by normalizing pathway contributions, enabling comparison of the relative structure of dysregulation across biological contexts. Within the TENSE framework, these scores are utilized to maintain the relative importance of pathways within enrichment analyses, followed by normalization across DIRES categories. Consequently, TENSE characterizes the distribution of dysregulation in each dataset rather than depending on direct comparison of raw enrichment scores from independent analyses.
Pathways included in each DIRES category were selected for their established mechanistic links to biological processes relevant to immune aging, such as genomic instability, DNA repair capacity, epigenetic regulation, chronic inflammatory activation, and nucleic acid sensing. Enrichment analyses utilized Enrichr,2 which offers curated pathway annotations derived from recognized biological databases including Reactome, KEGG, and Gene Ontology. These curated pathways enable enriched terms to be mapped to mechanistically interpretable biological processes, a requirement of the DIRES framework underlying the TENSE metric. Conversely, clustering approaches like MCODE (Zhou et al., 2019) in Metascape identify densely connected gene clusters based on network topology. However, while useful for exploratory network analysis, these clusters are dependent on individual datasets and may not align with predefined biological mechanisms.
Because the TENSE framework requires consistent mechanistic axes across datasets to quantify immune-aging–associated dysregulation, curated pathway annotations were preferred over topology-based clustering methods. Subsequently, enriched terms and combined scores are then analyzed by an LLM-powered KNIME workflow to provide DIRES and TENSE scoring. KNIME is an open-source workflow management platform with a graphical user interface that enables users to design and implement sophisticated data science pipelines (Song et al., 2026). In recent updates, KNIME has incorporated AI nodes to facilitate the integration of large language models, further empowering designed workflows. Figure 1 presents an overview of the study design.
An overview of the study design, including the calculations for TENSE (quantified immune-aging dysregulation index) and DIRES (DNA damage, inflammaging, DNA repair, epigenetic drift, sensing mechanism). The workflow accepts input in the form of significantly enriched Reactome pathways from RNA sequencing experiments designed to address specific research questions. The TENSE and DIRES engine is a hybrid semantic – deterministic workflow in which natural language processing is used to score Reactome pathways according to the DIRES scheme (semantic layer) and then calculate TENSE (deterministic layer). Created in BioRender () by GV and licensed under CC BY 4.0. https://biorender.com
Semantic layer: LLM-powered scoring of enriched biological processes
For the semantic layer of the workflow, natural language processing of enriched processes was performed via an API node using GPT-4.1-nano. GPT-4 has been previously utilized for processing the Reactome DB (Tiwari et al., 2023) and biomedical literature (Poretsky et al., 2025), and thus the GPT-4.1-nano3 was selected. GPT-4.1-nano is one of the default choices provided by the OpenAI LLM node in KNIME. The main advantages it provides are cost-effectiveness and speed according to its specifications.4 The use of GPT-4.1-nano is intended as a proof of concept and relies on at least one use case involving GPT-4 processing of Reactome data. The application of other LLMs, newer versions of GPT-4.1-nano (e.g., GPT-5-nano), as well as manual pathway annotation and TENSE calculation, is equally feasible.
In KNIME, GPT-4.1-nano is accessed via a unique API key, which is applied to an authenticator node and connected to an OpenAI LLM selector node. For this proof-of-concept application, GPT-4.1-nano was used with default settings, with a sampling temperature of 0.2 and a maximum response length of 200 tokens. Notably, sampling temperatures typically range from 0.0 to 1 (and above), with higher values allowing for creative responses, whereas lower temperatures, e.g., 0.0 to 0.3, correspond to answers with greater precision and are more appropriate when exact answers are required (Hong et al., 2025).
To score enriched pathways by DIRES, LLM prompter nodes were connected serially to GPT-4.1-nano and subsequently prompted via Expression Nodes to score significantly enriched terms (provided by input nodes; defined by a false discovery rate <0.05). The prompts are available in the online version of the workflow and from. An example of a D node prompt is the following, delivered to an LLM prompter node via a KNIME Expression Node: Supplementary File 1
The output provided at the end of this process is a table containing the DIRES scores before weighting for multi-category membership (Table 1).
This scoring scheme is performed independently for each D, I, R, E, S category and serially, D ➔ I ➔R ➔E ➔S, using a dedicated LLM prompter node and prompt. This design choice was intended to provide greater precision for each category while providing a dissectible architecture for troubleshooting and further development. The final DIRES scoring table is then submitted to the deterministic layer for further calculations.
| Term | D_count | Ι_count | R_count | E_count | S_count |
|---|---|---|---|---|---|
| Immune System | 0 | 0 | 0 | 0 | 0 |
| Innate Immune System | 0 | 1 | 0 | 0 | 0 |
| Neutrophil Degranulation | 0 | 1 | 0 | 0 | 0 |
| Keratinization | 0 | 0 | 0 | 0 | 0 |
| Cytokine Signaling in the Immune System | 0 | 1 | 0 | 0 | 0 |
| Signaling by Interleukins | 0 | 1 | 0 | 0 | 0 |
| RAC1 GTPase Cycle | 0 | 0 | 0 | 0 | 0 |
Deterministic layer: calculating TENSE via DIRES
The deterministic layer utilizes multiple KNIME nodes to calculate TENSE. To improve interpretation, we also provide normalized values for TENSE and for each DIRES category. This design choice enables easier interpretation as TENSE can score from 0 to 1 following normalization (i.e., 0–100% dysregulation), rather than 0–5 as a standardized value. To normalize TENSE’s values, we divide the derivative score from steps 1–5 by 5, i.e., the maximum standardized value of TENSE.
Specifically, the following step-wise procedure is followed to calculate TENSE from the DIRES pivot table.
As an overview of the workflow, the following process is followed to calculate DIRES and TENSE:
When none of the DIRES categories have enriched pathways, all components of the TENSE score become zero, resulting in a zero denominator. In the KNIME workflow, this leads to a missing value. To ensure clarity and consistency, these cases are deliberately assigned a TENSE value of 0. This approach indicates that there is no DIRES-associated dysregulation in the analyzed pathways, instead of representing a numerical calculation from non-zero values. Therefore, a TENSE score of 0 should be understood as an absence of immune-aging–related dysregulation within the boundaries of this framework.
Quality control and accuracy
Eidetic layer: restricting the black box effect of the LLM-scoring process
For quality control and to reduce black-box effects, we established a parallel explanatory layer that provides insight into the LLM scoring process and decision-making. This explanatory layer uses the same prompts, albeit instructed to provide a rationale rather than a binary classifier exclusively. This layer can be connected to the main classifier to provide insight into the DIRES scoring process.
As an example, for the first term in Table 1, the Eidetic Layer provides the following exchange when checking the term “Immune System” for membership in D:
GPT-4.1-nano response
We subsequently assess the rationale for each answer to ensure consensus between manual scoring by the author and GPT-4.1-nano. Notably, this example, “Immune System,” is a broad term that does not specify processes that map to inflammaging and, by default, is not intended to score in the I category to avoid artificial inflation.
Accuracy and precision metrics with reference to a human consensus
Theoretical approach
To address LLM non-determinism, i.e., the variability between repeated responses for the same prompt (Shusterman et al., 2025), we performed multiple (n = 100) sequential runs of a single dataset with at least one valid term for each category in DIRES.
We then extract TENSE values for each run and calculate mean TENSE±SD and 95% Confidence interval (CI). Subsequently, we score the dataset using DIRES as a human-consensus and provide a human-consensus TENSE value.
Subsequently, we use the latter value to calculate the following metrics:
Implementation dataset
To assess TENSE’s accuracy and precision, we utilize a transcriptomic module from Sparks et al. (2024) study. In brief, Sparks et al. (2024) performed weighted gene co-expression network analyses (WGCNA) of whole-blood transcriptomics from 228 patients of 22 monogenic immune diseases and 42 age- and sex-matched healthy participants. WGCNA identified 12 transcriptional modules that are differentially enriched for immune processes and cells. Among available TMs, we select TM1, which is enriched for Type I interferon responses. The rationale for this selection is that Type I interferon signaling and related pathways are expected to map broadly to DIRES categories (Cao, 2022; Papatriantafyllou, 2013). This will ensure that a non-zero TENSE score can be calculated and used to identify oscillations caused by LLM stochasticity.
TENSE is calculated as described above via a looped version of the TENSE workflow, utilizing the Counting Loop Start and Loop End nodes for n = 100 iterations of the workflow; These workflow iterations furthermore correspond to 5 prompt iterations for each of DIRES categories per significantly enriched pathway (npathway = 38), for a total of 100x5x38 = 1,540 total prompt iterations utilizing GPT-4.1-nano via LLM prompter nodes. (Supplementary File 1). TENSE scores produced via this process (Mean TENSE LLM = 0.242 ± 0.015, 95% CI:0.240–0.244) demonstrated high concordance with human consensus (BM = −0.006; RBM = −2.7%) and stable performance across repeated runs (SD = 0.015; CV = 4.24%; Table 2). The TENSE values produced by the workflow across 100 runs are available as Supplementary material 1 (Figure 2).
Subsequently, we leveraged Reactome Pathways from TM10, a transcriptional module enriched for housekeeping cell signaling pathways, which, by design, are not expected to score in DIRES. The workflow loop produced a consistent TENSE value of 0 across 100 independent workflow iterations, indicating no spurious DIRES activation and demonstrating high classifier specificity.
TENSE value calculations over 100 iterations of the KNIME workflow on the Interferon Signaling transcriptional module.
| Metric | Value |
|---|---|
| Reference TENSE | 0.248 |
| LLM mean TENSE | 0.242 |
| Standard deviation | 0.015 |
| 95% confidence interval (Half-Width) | 0.002 |
| Bias of the mean (BM) | −0.006 |
| Relative BM | −0.027 |
| Coefficient of variation (CV) | 0.042 |
Interoperability and modularity
To ensure TENSE’s application beyond the manual version of the workflow, we implemented Anthropic’s Claude Haiku 4.5 (Available as CLAUDE-HAIKU-4-5-20251001 from the KNIME LLM nodes) as an alternative LLM classifier for the semantic layer at the same sampling temperature T = 0.2, as previously described. The Haiku model was selected as relatively equivalent to GPT-4.1-nano in terms of cost efficiency for API calls, accuracy, and speed (Available from: https://platform.claude.com/docs/en/about-claude/models/overview↗; Accessed on March 9th, 2026).
TM9, enriched for RNA and RNA-related processes, was utilized as the comparison dataset. Genes comprising the module underwent the process described previously to produce a Reactome term list and associated Enricher-derived metrics.
Using fixed prompts for both LLM classifiers, Claude Haiku produced comparable results (Table 3 and Figure 3), indicating that the TENSE / DIRES workflow is interoperable between LLMs.
Claude Haiku produced several formatting errors that skewed DIRES; specifically, instead of binary classification, response cells contained text with descriptive answers. An example follows:
For the calculation of TENSE using Claude Haiku 4.5, formatting errors were treated as missing values; GPT-4.1-nano did not produce format errors. Despite fail-safes and consistent performance by GPT-4.1-nano, these format errors in Claude Haiku 4.5 response structure may plausibly reflect the result of prompt structure and its bias toward GPT-4.1-nano nativity. Further refinement with Claude was not pursued beyond the feasibility phase, in favor of GPT-4.1-nano’s performance on the given task and, by extension, API cost-efficiency.
To assess the robustness of the workflow in scoring a predetermined biological concept, we implemented an alternative/perturbed DIRES scheme where:
We then reanalyzed TM10, enriched for cellular signaling, and with a zero TENSE score. The perturbed TENSE score was 0.246 (24.6%), with cellular signaling cascades inflating the novel R node. We conclude that the workflow is adequately modular in the sense that it can serve biological concepts beyond those intended by TENSE’s design, e.g., function as a cell signaling/metabolic/stress classifier based on the design of the semantic layer.
The DIRES radar plot on the RNA metabolism transcriptomic modules by Chen and colleagues; note that modules Yellow and Pink produce a zero TENSE score. The “_rlt” suffix denotes the utilization of relative scores (i.e.,, i ∈ D, I, R, E, S) for each DIRES category. n i
| Reactome term | mODEL | D_rlt | I_rlt | R_rlt | E_rlt | S_rlt | TENSE |
|---|---|---|---|---|---|---|---|
| RNA metabolism | GPT-4.1-nano | 0.044 | 0.013 | 0.033 | 0.013 | 0.002 | 0.025 |
| RNA metabolism | claude-haiku-4-5-20251001 | 0.022 | 0.03 | 0.041 | 0.009 | 0.001 | 0.041 |
Access and availability
The current version of the TENSE classifier workflow is available via the KNIME community hub under/gvavou. An individual API key, or manual DIRES annotation of the datasets to be analyzed, is required to run each iteration of the workflow.
Results
Use cases for TENSE and DIRES annotation in the literature
Following the construction of the TENSE engine, we examine the utility of both TENSE and DIRES scores by annotating published datasets in various settings as use cases. We examine the TENSE/DIRES annotations in prior studies to determine whether they adequately capture biological information and offer further insight, as well as a novel means of presenting that information. Raw data usage from referenced studies, Enrichr access to Reactome,KNIME workflow utilization,and API Calls to GPT-4.5-nano (via KNIME AI integration) were performed on March 9th, 2026. 0005 0006
Use case 1: annotating Alzheimer’s disease immune aspects on neuron and microglial gene signatures
Background and rationale
For the first use case, we will annotate cell-specific gene signatures associated with Alzheimer’s disease diagnosis and pathology. We will furthermore examine the concordance between Neuron- and microglia-specific gene signatures across two distinct studies (Chen et al., 2022; Morabito et al., 2020). Specifically, we utilized a Neuronal-derived consensus gene module from medial temporal gyri from Alzheimer’s disease compared to control, as reported by Chen et al. (2022). Briefly, Chen and colleagues performed weighted gene co-expression network analysis (WGCNA) on 10,000 highly variable genes, identifying eight co-expression gene modules, four of which (yellow, brown, pink, and turquoise) showed to correlate with AD pathology.
DIRES and TENSE annotations on Chen et al.’s dataset were compared to annotations on corresponding neuron-enriched (Consensus Module 1; CM1) and microglia-enriched (Consensus Module 8; CM8) AD-associated gene modules reported by Morabito et al. (2020). In their study, Morabito et al. (2020) reported on WCGNA-derived consensus gene modules that were correspondingly microglia- and neuron-specific, and associated with Alzheimer’s disease.
Finally, we aimed to determine whether TENSE and DIRES annotations would be conceptually concordant with data reported by Das et al. (2024) by comparing the response transcriptomes between Αβ plaques vs. neurofibrillary tangles, which revealed inflammation as the most salient component.
Approach and results
We performed enrichment analysis for each pair of gene modules using Enrichr and identified significantly over-represented pathways from Reactome. We replicated the steps described previously. Subsequently, we obtain the following values for TENSE:
The DIRES radar plot on the four AD-related transcriptomic modules by Chen and colleagues; note that modules Yellow and Pink produce a zero TENSE score. The “_rlt” suffix denotes the utilization of normalized relative scores (i.e.,, i ∈ D, I, R, E, S normalized bounded by 0,1) for each DIRES category. n i
The DIRES radar plot for the “Aβ plaques vs. neurofibrillary tangles” gene signature extracted from Das and colleagues, indicating a predominantly inflammatory component. The “_rlt” suffix denotes the utilization of normalized relative scores (i.e.,, i ∈ D, I, R, E, S normalized bounded by 0,1) for each DIRES category. n i
The DIRES radar plot on the AD-related brown module by Chen and colleagues, and the corresponding AD microglial module by Morabito and colleagues. The “_rlt” suffix denotes the utilization of normalized relative scores (i.e.,, i ∈ D, I, R, E, S normalized bounded by 0,1) for each DIRES category. n i
The DIRES radar plot on the AD microglial gene module by Chen and colleagues, and the corresponding AD microglial module by Morabito and colleagues. The “_rlt” suffix denotes the utilization of normalized relative scores (i.e.,, i ∈ D, I, R, E, S normalized bounded by 0,1) for each DIRES category. n i
Use case 2: annotating the differential transcriptomic response of radiation-resistant and radiation-sensitive cells exposed to 4 Gy radiation at 4- and 12-h post-exposure
Background and rationale
Radiation induces both genotoxic stress and inflammation in affected cells; contrary to annotated data from AD, data from cellular irradiation inform us on the temporally resolved effects of a well-defined exposure on the transcriptome. To approach this research question, we utilized data available from Borràs-Fresneda et al. (2016). Their study reported differences in gene expression between radiosensitive (RS) and radioresistant (RR) cell lines at various time points after irradiation at a dose of 4 Gy.
Approach and results
To obtain DIRES and TENSE annotations, we replicate the steps described previously. Subsequently, we obtain the following values for TENSE:
Figures 8, 9 display Radar plots for DIRES scores for the radiosensitive and radioresistant lines, respectively. DIRES annotation suggests that the radiosensitive line exhibits an inflammatory component at 4 h, which is absent at 12 h post-exposure. Conversely, the radioresistant line’s transcriptome indicates invariant activation of repair mechanisms at both time points, and the absence of a detectable inflammatory component.
DIRES annotation outlines a sustained inflammatory and DNA damage component for radiosensitive cells, whereas radioresistant cells display an invariant DNA repair component. Differences in TENSE scores indicated a more dynamic utilization of cellular processes in radiosensitive cells and a robust yet invariant utilization for radioresistant cells, polarized toward repair.
These findings are in agreement with the findings of the original study, where differences in DNA repair capacity and mitigation of radiation-induced inflammation Yahyapour et al. (2018) between the two cell types was outlined. With our approach, both TENSE and DIRES serve to quantify these changes beyond the study of differentially expressed genes or pathways of interest. The concept of ΔTENSE, i.e., the difference between TENSE values at baseline and post-exposure, is useful for distinguishing between stable and dynamic processes.
The DIRES radar plot demonstrates a transient inflammatory component evident at 4 h and absent at 12 h, while the DNA damage component remains active. The “_rlt” suffix denotes the utilization of normalized relative scores (i.e.,, i ∈ D, I, R, E, S) for each DIRES category. Gy: Gray, RS: Radiosensitive; RR: Radioresistant. n i
The DIRES radar plot demonstrates an invariant polarization toward DNA damage pathways at 4 h and 12 h. The “_rlt” suffix denotes the utilization of relative scores (i.e.,, i ∈ D, I, R, E, S) for each DIRES category. Gy: Gray, RR: Radioresistant. n i
Use case 3: TENSE and DIRES as descriptors of immune gene network signatures
Background and rationale
Conserved gene networks have previously been linked to impaired regulation of the host response to infection, and their activity is often captured by distinct gene signatures that link gene activity with clinical outcomes. One such signature, the 42-gene Severe-or-Mild (SoM) signature, was found to be associated with several risk factors for poor outcomes from infection, mortality, and capture the response to treatment Ganesan et al. (2025). Similar approaches posit that deviations from health, including aging and disease, have a common systemic immune dysregulation footprint. The 150-gene signature Immune Health Metric (IHM) score represents one such effort to capture this shared influence (Sparks et al., 2024). Specifically, the IHM correlates with aging in healthy individuals and provides differential values between healthy individuals and those with monogenic diseases. In this use case, we explore TENSE and DIRES as descriptors of SoM and IHM scores. To calculate TENSE and DIRES, we assume that the SoM and IHM genes are differentially expressed, map enriched processes to DIRES, and calculate TENSE for each SoM and IHM score.
Approach and results
As described in use case 1, replicating the procedures established to calculate DIRES and TENSE, we determine the following TENSE values. Notably, as IHM is calculated as the difference between the geometric means of two gene sets (Sparks et al., 2024), (herein referred to as IHM-A and IHM-B), we calculate TENSE and DIRES for each.
The DIRES radar plot demonstrates an invariant polarization toward inflammaging for the IHM-B and SoM signatures, whereas IHM-A signatures do not score in any DIRES categories, resulting in a TENSE = 0. The “rlt” suffix denotes the utilization of relative scores (i.e.,, i ∈ D, I, R, E, S) for each DIRES category. IHM: Immune ealth etric; SOM: evere-or-mild index. n i
Discussion
This study introduces TENSE (quanTifiEd immuNe-aging dySregulation indEx), a framework that turns pathway enrichment results into a quantitative, system-level measure of immune-aging–related dysregulation. Instead of generating typical lists of significant pathways for qualitative interpretation, TENSE integrates signals from multiple pathways into key biological categories, summarizing them as an understandable dysregulation profile. As a result, it offers a higher-level overview of enrichment results, making it easier to compare biological situations based on the degree and type of aging-related regulatory imbalance present. The classification used in TENSE proved reasonably reliable, giving consistent results across repeated tests and closely matching human expert assessments. Similar outcomes were observed with alternative large language models such as Claude Haiku, suggesting that the underlying classification task is well-structured and produces stable results across different tools. Each of the five DIRES axes was found in at least one of the analyzed datasets, indicating that TENSE captures a variety of dysregulation patterns rather than being limited to just one category. Overall, these findings support TENSE as a reproducible and easy-to-interpret approach for converting complex enrichment outputs into meaningful estimates of immune-aging dysregulation.
Critical analysis of TENSE use cases
Application of the TENSE framework across transcriptional modules and experimental contrasts revealed distinct patterns of aging-associated immune dysregulation enrichment patterns across biological contexts. Within Alzheimer’s disease (AD) modules derived from WGCNA analysis, TENSE values were primarily driven by the inflammaging axis, with the strongest contributions observed in the CM8 microglial module (TENSE = 0.208/20.8%) and the brown module (TENSE = 0.131/13.1%). This pattern is consistent with extensive evidence implicating microglial activation and chronic innate immune signaling as key contributors to AD pathology. Conversely, neuronal modules such as CM1 exhibited relatively low TENSE values, with minor contributions from both inflammaging and nucleic acid sensing pathways, suggesting that immune-aging–related dysregulation is less pronounced within neuronal transcriptional programs. Similarly, the comparison between neurofibrillary tangles and plaques demonstrated moderate TENSE values largely attributable to inflammaging, a finding corroborated by the role of inflammatory signaling in AD-associated transcriptional dysregulation.
Outside the AD context, distinct enrichment patterns emerged. In the radiation response experiments, TENSE values were dominated by the DNA damage axis, particularly in the radiosensitive condition at 12 h (RS_12h; TENSE = 0.364/36.4%). This observation is consistent with the induction of genomic instability and tandem DNA damage–response pathways following irradiation, representing expected transcriptional responses to ionizing radiation (Miyake et al., 2019). Conversely, the sepsis-associated SOM gene signature yielded the highest TENSE value observed in the dataset (0.4/40%), driven entirely by the inflammaging axis, highlighting the extensive inflammatory activation characteristic of systemic infection. The Immune Health Metric gene signatures further illustrated this contrast: while IHM_A showed no detectable DIRES contributions, IHM_B exhibited a strong inflammaging signal (TENSE = 0.217/21.7%), suggesting that subsets of immune health–associated gene programs may capture inflammatory activation states. Notably, ΔTENSE for the two IHM signatures is negative, a finding that is in accordance with SoM and IHM scores, as expected to be inversely correlated (Ganesan et al., 2025): DIRES annotation suggests that high IHM reflects a shift from IHM-B contributors, i.e., Inflammaging and Nucleic Acid Sensing (higher TENSE values) to IHM-A contributors, primary nucleic acid sensing mechanisms, and those associated epigenetic drift at a quiescent state (lower TENSE values).
Finally, the TM10 RNA metabolism module exhibited low but distributed contributions across multiple DIRES axes, including DNA damage, DNA repair, epigenetic drift, and inflammaging, producing a comparatively modest TENSE value. This pattern suggests diffuse engagement of multiple regulatory processes rather than dominance of a single dysregulated mechanism.
Collectively, these results illustrate how the TENSE framework can distinguish between biological contexts characterized by inflammation-dominated dysregulation (e.g., microglial activation and sepsis) and those dominated by genomic stress responses (e.g., radiation exposure). Such mechanistic differentiation is difficult to infer directly from lists of enriched pathways, highlighting the utility of summarizing pathway enrichment outputs into structured, interpretable dysregulation profiles. Table 4 summarizes DIRES and TENSE scoring across the use cases.
| Use cases | D_rlt | I_rlt | R_rlt | E_rlt | S_rlt | TENSE |
|---|---|---|---|---|---|---|
| AD CM1 neuronal | 0 | 0.029 | 0 | 0 | 0.013 | 0.022 |
| AD CM8 microglial | 0 | 0.208 | 0 | 0 | 0 | 0.208 |
| AD module brown | 0 | 0.131 | 0 | 0 | 0 | 0.131 |
| AD module turquoise | 0.005 | 0.038 | 0 | 0 | 0.009 | 0.012 |
| AD NFT vs. plaques | 0 | 0.111 | 0 | 0 | 0.009 | 0.096 |
| AD Pink | 0 | 0 | 0 | 0 | 0 | 0 |
| AD Yellow | 0 | 0 | 0 | 0 | 0 | 0 |
| IHM_A | 0 | 0 | 0 | 0 | 0 | 0 |
| IHM_B | 0 | 0.217 | 0 | 0 | 0 | 0.217 |
| RR_12 h | 0.2 | 0 | 0 | 0 | 0 | 0.2 |
| RR_4 h | 0.2 | 0 | 0 | 0 | 0 | 0.2 |
| RS_12 h | 0.364 | 0 | 0 | 0 | 0 | 0.364 |
| RS_4 h | 0.044 | 0.044 | 0 | 0 | 0 | 0.044 |
| SOM | 0 | 0.4 | 0 | 0 | 0 | 0.4 |
| TM10 RNA metabolism | 0.044 | 0.013 | 0.033 | 0.013 | 0.002 | 0.025 |
Limitations, caveats, and further development
In this study, the TENSE and, by extension, DIRES are defined as an alternative to arbitrary annotation of pathway enrichment data and forming decisions about the contribution and directionality of biological disorder in various settings. Rather than a comprehensive annotation of every possible process, it adopts a utilitarian approach, annotating hubs of disorder as aging progresses and providing a summary estimate that translates dysfunction as measurable relative disorder. The use cases explored indicate that TENSE and DIRES align with the biology and novel findings of the original studies, showing concordance in annotating data stemming from similar biological contexts and capturing the directionality of biological processes responding to exposures.
TENSE should be viewed within the appropriate context and limitations. First, it should be acknowledged that aging is a complicated process and even more so at the molecular level. The pathways selected to construct the TENSE workflow are hypothesis-driven rather than data-driven themselves. This issue is addressed by selecting hub processes that reflect molecular events that are not specifically modeled. For example, mtDNA leakage in the cytoplasm may not be directly modeled as a term in the formula; however, it serves as a proxy in the Nucleic Acid Sensing term. Similarly, genotoxic stress from various agents or metabolic/proteostatic effects is not directly modeled by a dedicated term but would correspondingly score as a proxy in the D, R, and I categories – as shown in the relevant use case 2.
Another caveat is that the theoretical formulation of TENSE involves inherent skew. For instance, the TENSE formulation processes D, I, R, E, S as additive and equivalent contributors. This potential caveat is addressed by substituting each constituent with its standardized equivalent and finally standardizing the output to its maximum value. This, in turn, yields interpretable values: higher TENSE values reflect greater relative utilization of the transcriptome, with DIRES revealing the relative contribution of each component. LLM stochasticity may produce TENSE values that are generally reliable based on our own iterative validation; however, the human consensus annotation herein is performed by the authors for proof of concept. A broader concept that utilizes multiple raters, e.g., a Delphi consensus and reporting inter-rater agreement statistics, can further bolster human-reference reliability and provide a more direct comparator to LLM stochasticity.
Similarly, it is worth reiterating, as a potential caveat, that TENSE does not capture the detailed contributions of each possible contributor to aging, but rather summarizes several salient hub processes. However, this limitation can be further overcome by adding additional hub functions in the DIRES schema, e.g., mitochondrial function, and adjusting the TENSE formulation correspondingly. As a modular workflow, TENSE relies on a semantic layer that can be adapted to a specific concept. Rather than simply representing the weighted standardized sum of activity scores and imposing the arithmetic mean as a constraint for each DIRES term, it aims to reduce skew by biologically similar overrepresented processes. We have opted to constrain potential scoring inflation of DIRES terms by infection terms that report on redundant gene networks, such as ribosomal proteins. Hence, these pathways do not score explicitly in DIRES unless referring to specific inflammatory processes during the course of infection.
Another potential caveat of the TENSE framework is that its initial validation was restricted to a small selection of reference conditions, specifically one interferon-associated module and one housekeeping module. Although this demonstrates that the method can work, it does not confirm its applicability to a wide range of biological scenarios. To address this issue, future studies should include a more diverse set of baseline and homeostatic conditions, enabling a thorough evaluation of the framework’s robustness and accuracy, especially in situations close to floor effects or lower TENSE states. The TENSE workflow should also be considered within the context and constraints of overrepresentation analyses, in that they represent enrichment or depletion of gene sets against our knowledge base, rather than in vivo evidence of pathway activity. As such, high-throughput datasets containing, e.g., methylomics, differential gene expression, and DNA damage marker data can provide alternative formulations of TENSE and benchmarking closer to the theoretical model, rather than the current enrichment pattern-based formulation.
Lastly, possible permutations that utilize the same theoretical model are possible. For instance, another pathway-centric approach could incorporate Shannon Entropy (Shannon, 1948), which estimates pathways related to aging and standardizes their impact over the background in sequencing experiments (Carels, 2024; Zenil et al., 2018). Furthermore, it should be noted that it was not conceptualized as a substitute or surrogate to existing predictors of biological age. Following Horvath’s clock (Horvath, 2013), several approaches utilizing data from epigenomics (Porter et al., 2021, Hannum et al., 2013, Prosz et al., 2024), metabolomics (Huang et al., 2025), transcriptomics (Zakar-Polyák et al., 2024) and proteomics (Argentieri et al., 2024) based approaches, among others (Galkin et al., 2020) have provided deeper insight into deriving accurate biological age estimates (Jansen et al., 2021). By design, TENSE and DIRES were constructed to capture a distinct biological signal from aging clocks. While a biological clock would be utilized to estimate biological age, TENSE would be utilized to estimate the effect of an exposure that maps to age-related processes, or the skew of any of these processes (e.g., inflammation and/or DNA damage) in a biological context (e.g., disease).
TENSE can thus complement these approaches, functioning more like a biological compass and accelerometer, rather than a biological clock.
Conclusion
The TENSE and DIRES scoring schema represented a novel, reductionist approach in annotating pathway enrichment data and an alternative to arbitrary or hypothesis-driven interpretation. The responsiveness of both TENSE and DIRES to complex disease processes, cell-specific gene modules, and complex exposures, including ionizing radiation, across diverse biological contexts, suggests further potential usage beyond aging-related disease. Using the provided cases, we highlighted the potential of TENSE as a reproducible and interpretable method for summarizing complex pathway enrichment outputs into biologically meaningful estimates of immune-aging–related dysregulation. In this sense, TENSE transforms pathway enrichment outputs from descriptive lists into quantitative, mechanistically interpretable estimates of system-level dysregulation.
Acknowledgments
The authors would like to express their gratitude to the Cyprus Academy of Sciences, Letters, and Arts for their support during the preparation of this manuscript. GV would like to express his gratitude to Taisia Kyriakou for her insight and support during the preparation of this manuscript. Figures were created through BioRender.
Funding Statement
The author(s) declared that financial support was not received for this work and/or its publication.
Data availability statement
The original contributions presented in the study are included in the article/; further inquiries can be directed to the corresponding author/s. Supplementary material
Ethics statement
Ethical approval was not required for the study involving humans in accordance with the local legislation and institutional requirements. Written informed consent to participate in this study was not required from participants or their legal guardians/next of kin in accordance with national legislation and institutional requirements.
Author contributions
GV: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing. GH: Funding acquisition, Writing – original draft, Writing – review & editing.
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
The author GV declared that they were an editorial board member of Frontiers, at the time of submission. This had no impact on the peer review process and the final decision.
Generative AI statement
The author(s) declared that Generative AI was not used in the creation of this manuscript.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Supplementary material
The Supplementary material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frai.2026.1732901/full#supplementary-material↗
References
Associated Data
Supplementary Materials
Data Availability Statement
The original contributions presented in the study are included in the article/; further inquiries can be directed to the corresponding author/s. Supplementary material