An NLRP3 inflammasome–anchored Astragalus mechanistic prior yields a mortality-associated transcriptomic signal in ICU sepsis: a secondary analysis with exploratory cross-cohort assessment

Abstract

Objective and design

We developed ASTRA, a literature-informed, author-defined Astragalus mechanistic prior, and examined whether integrated and node-level whole-blood transcriptomic scores were associated with 28-day mortality in ICU sepsis.

Methods

The primary analysis included 479 adults with sepsis from GSE65682; an exploratory cross-cohort directional assessment included 51 Day-1 patients with septic shock from GSE95233. Mean-Z scores were evaluated using age-adjusted logistic regression with false-discovery-rate correction. Sensitivity analyses included singscore, restricted cubic splines, label permutation, leave-one-gene-out analysis, correlations with IL1B and IL6 mRNA, and adjustment for transcriptome-derived myeloid-cell composition.

Results

The integrated ASTRA score showed a nominal inverse association with mortality but did not survive false-discovery-rate correction. The NLRP3-related three-gene score showed the strongest association in GSE65682 (OR per 1-SD increase, 0.70; 95% CI 0.57–0.86; q = 0.0072) and remained directionally stable in leave-one-gene-out analyses. It correlated positively with IL1B and IL6 mRNA and with MCP-counter monocytic-lineage and neutrophil scores. Joint MCP-counter adjustment attenuated the association to an OR of 0.79 (95% CI 0.58–1.06), whereas xCell neutrophil adjustment strengthened it to an OR of 0.64 (95% CI 0.50–0.83), indicating method-sensitive dependence on estimated cell composition. In GSE95233, the age-adjusted NLRP3 estimate was directionally concordant but imprecise (OR 0.68; 95% CI 0.37–1.23).

Conclusions

ASTRA provides a reproducible framework for mechanism-guided transcriptomic scoring in sepsis. The NLRP3-linked signal represents an observational whole-blood transcriptomic state that partly tracks estimated myeloid-cell composition and concurrent inflammatory transcription. It does not establish protective inflammasome activity, Astragalus efficacy, or pharmacological target engagement.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s00011-026-02352-0.

Introduction

Sepsis is a life-threatening syndrome caused by a dysregulated host response to infection, in which excessive inflammation, immune suppression, oxidative stress, mitochondrial injury, endothelial and epithelial barrier dysfunction, and impaired tissue repair interact to promote organ failure and death [1–3]. Blood transcriptomic studies have shown that sepsis is not a single molecular state but comprises heterogeneous host-response patterns and genomic endotypes associated with diagnosis, prognosis, and immune dysregulation [4–6]. However, translating high-dimensional transcriptomic signals into interpretable, mechanism-focused, patient-level readouts remains difficult. General-purpose pathway enrichment can identify broad biological themes, but overlapping pathway boundaries, redundancy among inflammatory modules, and limited pharmacologic traceability often constrain its usefulness when the objective is to connect a predefined mechanistic hypothesis with clinical outcomes [7–11].

Among immune-inflammatory circuits implicated in sepsis, the NLRP3 inflammasome is particularly relevant because it integrates upstream priming signals, including TLR4–NF-κB activation, with cellular stress cues such as mitochondrial reactive oxygen species, ionic flux, and metabolic injury [12, 13]. Mitochondrial dysfunction is also a central feature of critical illness and sepsis, contributing to impaired bioenergetics, redox imbalance, immune-cell dysfunction, and organ injury, while mitochondrial quality control and biogenesis have been linked to recovery and survival in critical illness [1, 14–16]. These observations support a host-response framework in which inflammatory stress pathways and mitochondrial homeostasis/repair pathways should be interpreted jointly rather than as isolated modules. A transcriptomic scoring strategy that captures these axes at the single-patient level may therefore provide a useful bridge between pathway biology and outcome-associated immune signatures.

Radix Astragali (Huangqi), the root of Astragalus mongholicus Bunge [synonym: Astragalus membranaceus (Fisch.) Bunge], is widely used as an adjunctive botanical medicine in East Asian clinical practice, and its major constituents include saponins, flavonoids, and polysaccharides with reported immunomodulatory, anti-inflammatory, antioxidative, barrier-protective, and mitochondrial effects [17–19]. Experimental studies of Astragalus constituents, including astragaloside IV, calycosin, formononetin, and Astragalus polysaccharides, have linked this botanical to pathways relevant to sepsis biology, including TLR4–NF-κB signaling, NLRP3 inflammasome regulation, MAPK activation, AMPK–SIRT1/PGC-1α signaling, NRF2–HO-1 antioxidative responses, autophagy/mitophagy, epithelial or endothelial barrier integrity, and mitochondrial homeostasis [17, 20–22]. Nevertheless, these mechanistic observations remain dispersed across experimental systems and disease contexts, and have rarely been converted into a transparent, computable prior that can be tested in human sepsis transcriptomes.

To address this gap, we constructed ASTRA as a literature-informed, author-defined Astragalus mechanistic prior that organizes Astragalus-related mechanistic concepts into predefined immune-inflammatory and homeostatic-repair nodes and their representative genes. We then applied this prior to the whole-blood ICU sepsis transcriptome cohort GSE65682 to compute integrated and node-level pathway scores and examine their associations with 28-day mortality. Specifically, we aimed to: (1) organize literature-informed Astragalus-related mechanistic concepts into a transparent node–gene framework; (2) assess whether the resulting gene sets are computable in a human sepsis whole-blood transcriptomic background; and (3) evaluate whether integrated and node-level scores, with attention to the NLRP3 inflammasome module, are associated with 28-day mortality in a hypothesis-generating secondary analysis.

Materials and methods

Study design

We conducted a hypothesis-generating secondary transcriptomic analysis using a literature-informed, author-defined Astragalus mechanistic prior and a publicly available ICU sepsis cohort. The analytical workflow consisted of five sequential steps: structured database retrieval and rule-assisted title/abstract-level candidate-record prioritization; literature-informed mapping of Astragalus-related mechanistic concepts to predefined immune-inflammatory and homeostatic-repair nodes and representative genes; construction of integrated and node-level gene sets; single-sample transcriptomic scoring in whole-blood sepsis profiles; and association testing with 28-day mortality. The study was designed to evaluate whether this mechanistic prior could generate interpretable patient-level host-response scores in sepsis, rather than to establish clinical validity or therapeutic efficacy.

Clinical transcriptomic dataset

Whole-blood transcriptomic and clinical data were obtained from the publicly available Gene Expression Omnibus dataset GSE65682 [23], generated by the Molecular Diagnosis and Risk Stratification of Sepsis consortium. This cohort includes adult critically ill patients with blood microarray profiles and 28-day outcome information and has been used in prior studies of ICU infection, sepsis diagnosis, and genomic sepsis endotypes [5, 23–25]. We used the harmonized HGNC symbol-level expression matrix and matched phenotype file, including 479 patients with complete 28-day survival status: 365 survivors and 114 non-survivors. Probe identifiers were mapped to HGNC gene symbols using the hgu219.db annotation package. Probes without an assigned gene symbol were removed, and when multiple probes mapped to the same gene symbol, their median expression value was used to obtain a single gene-level measurement. Only gene sets with at least two genes overlapping the expression background were retained for downstream scoring. Because the dataset is de-identified and publicly available, no additional ethics approval or informed consent was required for this secondary analysis.

Literature retrieval, candidate-record prioritization, and construction of the ASTRA prior

We searched PubMed, Embase, and Web of Science from database inception to 13 October 2025. The search strategy combined botanical terms, including Astragalus membranaceus, Astragalus mongholicus, Huangqi, and Radix Astragali; constituent terms, including astragaloside IV, calycosin, formononetin, and Astragalus polysaccharides; and prespecified mechanistic terms related to immune-inflammatory and homeostatic repair pathways. The mechanistic search domain included TLR4–NF-κB signaling, the NLRP3 inflammasome, MAPK/p38/JNK/ERK signaling, JAK/STAT3, PI3K–AKT–mTOR, AMPK–SIRT1/PGC-1α, NRF2–HO-1, autophagy/mitophagy, epithelial or endothelial barrier integrity, mitochondrial homeostasis, and short-chain fatty acid receptor pathways. Full database-specific search strategies are provided in Supplementary Material 1.

After the original DOI-, PMID-, and title-based deduplication procedure, 4,692 bibliographic records remained. These records underwent two rule-assisted title/abstract-level prioritization stages. The first stage removed 35 meeting abstracts, letters, editorials, or retracted records, leaving 4,657 records for further prioritization. The archived second-stage fields recorded the presence of an eligible Astragalus constituent or preparation, a prespecified mechanistic term, an abstract-level indication of experimental validation, and whether the report was limited to docking or network pharmacology. The archived output contained 2,222 candidate-record rows. These rows constituted a broad source-discovery pool and were not treated as 2,222 unique full-text-included publications or as 2,222 independent experimentally extracted evidence observations.

The final computable ASTRA prior comprised 10 broad mechanistic categories, 20 predefined nodes, and 37 representative genes, as recorded in the archived node-mapping and Gene Matrix Transposed files. The genes were used as a parsimonious transcriptomic representation of the prespecified mechanistic nodes rather than as exhaustive pathway membership. The archived workflow did not preserve a complete publication-by-gene record of the original representative-gene selection process or a formally discriminative evidence-grading system. Accordingly, ASTRA is interpreted in this study as a literature-informed, author-defined mechanistic prior rather than as a systematic quantitative synthesis of experimental evidence. The resulting resources included one integrated Astragalus-derived gene set and 20 node-specific gene sets. The complete gene composition, node membership, and transcriptomic coverage of the locked prior are provided in Supplementary Table S6. Of the 37 integrated genes, 33 were represented in GSE65682 and 36 in GSE95233; the three-gene NLRP3 node was fully represented in both datasets.

The three-gene NLRP3 node was not selected from the 37-gene integrated set on the basis of its association with mortality. Node membership had been prespecified in the locked node-level GMT resource before outcome modeling. The NLRP3 node comprised NLRP3, PYCARD, and CASP1, representing the sensor, adaptor, and effector components of the canonical inflammasome complex, respectively. All three genes were represented in both GSE65682 and GSE95233. The node score was calculated as the mean of the gene-wise standardized expression values. This parsimonious node representation was author-defined and was not derived from publication-frequency ranking, network-centrality analysis, or outcome-driven feature selection.

Gene-set coverage assessment

Before association testing, each integrated or node-level gene set was intersected with the GSE65682 expression background. For each set, we calculated the total number of genes in the prior, the number of overlapping genes in the transcriptomic dataset, and the overlap proportion. Gene sets with fewer than two overlapping genes were considered insufficiently represented and were excluded from formal scoring and outcome modeling. This threshold was used to reduce instability from extremely small gene sets while preserving mechanistic interpretability.

Single-sample transcriptomic scoring

The primary scoring method was mean-Z scoring. For each gene, expression values were standardized across samples. For each patient and each gene set, the mean of standardized expression values across the overlapping genes was then calculated:

where denotes the mean-Z score for patient  iand gene set  G, is the expression value of gene  gin patient  i, and and are the cohort-level mean and standard deviation of gene  g, respectively.

As a sensitivity method, we applied singscore, a rank-based single-sample gene-set scoring approach that is less dependent on expression magnitude and more robust to sample-level distributional differences [26]. For comparability across methods, all resulting scores were standardized across patients to mean 0 and standard deviation 1. Effect estimates are therefore interpreted as the change in odds of 28-day mortality per 1-SD increase in the transcriptomic module score. Because these scores were derived from mRNA expression and did not encode experimental activating or inhibitory directions as signed weights, higher values indicate greater relative transcript abundance rather than protein activity, inflammasome assembly, pathway flux, pyroptotic activity, Astragalus-like inhibition, or pharmacological target engagement.

Outcome

The primary outcome was 28-day mortality, coded as 1 for death and 0 for survival. Outcome status was complete for all included patients. Time-to-event analyses were not performed because the primary analysis used the binary 28-day outcome available in the harmonized phenotype file.

Primary analysis

We fitted age-adjusted logistic regression models to examine associations between standardized pathway scores and 28-day mortality. The integrated Astragalus-derived score and all eligible node-level scores were evaluated using the same modeling framework:

where denotes 28-day mortality for patient  i, denotes the standardized pathway score, and denotes age. Odds ratios were calculated as and reported with 95% confidence intervals. An OR below 1 indicates that a higher transcriptomic score is associated with lower odds of 28-day mortality, whereas an OR above 1 indicates higher odds of 28-day mortality. The model was intentionally limited to age adjustment because key clinical variables, treatment exposures, and sampling-window variables were unavailable or incompletely harmonized in the public dataset. Accordingly, the analysis should be interpreted as a mechanism-guided association study rather than a fully adjusted prognostic model.

Multiplicity control and reporting

Multiplicity was addressed using the Benjamini–Hochberg false-discovery-rate procedure across the prespecified primary comparison family, consisting of the integrated score and all eligible node-level scores evaluated with the primary mean-Z method [27]. Two-sided P values and FDR-adjusted q values are reported. Associations with q < 0.05 were considered statistically significant after FDR correction; associations with nominal P < 0.05 but q ≥ 0.05 were interpreted as nominal and hypothesis-generating. This rule was applied consistently in the main text and supplementary tables. The integrated score and the NLRP3 inflammasome node were emphasized because they reflected the overall mechanistic prior and the sentinel immune-inflammatory node, respectively; however, statistical interpretation followed the same FDR framework.

Restricted cubic spline analysis

For the sentinel NLRP3 inflammasome node, we examined the functional form of the association with 28-day mortality using restricted cubic splines in the age-adjusted logistic model. Four knots were placed at empirical quantiles of the standardized NLRP3 score, and the curve was referenced to a score of 0, corresponding to the cohort mean. We reported the overall association P value and the nonlinearity P value, following standard approaches for spline-based dose–response assessment [28, 29]. This analysis was used to evaluate whether the observed association was approximately linear over the observed score range and was not used for model selection.

Sensitivity and robustness analyses

We performed several sensitivity analyses to evaluate the stability of the main findings. First, we compared mean-Z and singscore estimates across the integrated and node-level gene sets. Method agreement was assessed by comparing log odds ratios from the two scoring methods and by visualizing whether estimates fell within a practical equivalence band of , corresponding approximately to a ± 22% difference on the odds-ratio scale. This equivalence band was used descriptively to assess magnitude concordance rather than as a non-inferiority test.

Second, for the NLRP3 node, we conducted label-permutation analyses to evaluate whether the observed association was likely to arise under random outcome allocation. Outcome labels were permuted 10,000 times while preserving the observed pathway scores and age values. In each permutation, the same age-adjusted logistic model was refitted, and the likelihood-ratio chi-square statistic was recalculated. The empirical P value was computed as:

where and  kis the number of permutations with a test statistic greater than or equal to the observed statistic. As an additional robustness check, we repeated the permutation test within age deciles to preserve the coarse age distribution. Monte Carlo uncertainty was estimated as:

Third, we performed a leave-one-gene-out sensitivity analysis to determine whether the NLRP3-node association was dependent on any single component. NLRP3, PYCARD, and CASP1 were excluded in turn, and each reduced two-gene score was reconstructed using the same gene-wise standardization, averaging, and 1-SD rescaling procedure as the original three-gene node. Each reduced score was evaluated using the same age-adjusted logistic-regression model as the primary analysis. Because this analysis examined one prespecified node and was conducted as a targeted robustness assessment, no additional multiplicity adjustment was applied.

Fourth, as a reproducibility check, we inspected the archived implementation of the evidence-weighted NLRP3 node score. The archived weighting scheme assigned identical or near-identical weights to NLRP3, PYCARD, and CASP1, and the resulting standardized score was effectively indistinguishable from the unweighted score (Pearson r = 1.00 at the reported precision). In addition, the available documentation did not permit full reconstruction of the original evidence-grading logic required to support differential weighting. Accordingly, the evidence-weighted score did not provide an independent robustness test and was not interpreted as separate supporting evidence.

Transcript-level correlation analysis To examine whether the mortality-associated NLRP3 transcriptomic score tracked concurrent inflammatory transcription, we assessed its correlations with IL1B and IL6 mRNA abundance in the 479-patient GSE65682 cohort. The analysis retained the locked standardized three-gene NLRP3 score used in the primary analysis (Original_3_gene_z; NLRP3, PYCARD, and CASP1). Spearman correlations were used as the primary analyses, with 95% confidence intervals estimated from 5,000 nonparametric bootstrap replicates and P values adjusted across the two transcripts using the Benjamini–Hochberg method. Pearson correlations and age-adjusted partial Spearman and Pearson correlations were examined as sensitivity analyses.

Estimated cell-composition sensitivity analysis

To assess whether the locked whole-blood NLRP3 transcriptomic score tracked variation in leukocyte composition, we estimated transcriptome-derived myeloid-cell scores in the same 479-patient GSE65682 cohort using MCP-counter version 1.2.0 and xCell version 1.1.0. MCP-counter monocytic-lineage and neutrophil scores were treated as the primary estimated-composition covariates. The xCell neutrophil enrichment score was examined as a method-specific sensitivity measure. The xCell monocyte score was retained for descriptive assessment only because of marked practical zero inflation.

Associations between the locked NLRP3 score and the estimated cell scores were assessed using Spearman correlations, with 95% confidence intervals obtained from 5,000 nonparametric bootstrap replicates. The original age-adjusted mortality model was then refitted after separate adjustment for the MCP-counter monocytic-lineage and neutrophil scores, joint adjustment for both MCP-counter scores, and adjustment for the xCell neutrophil score. In an additional sensitivity analysis, the NLRP3 score was residualized against both MCP-counter scores and re-standardized before age-adjusted logistic regression.

Changes relative to the original model were summarized as the proportional change in the absolute log-OR magnitude. ORs and their 95% CIs were obtained from the fitted logistic-regression models. Uncertainty in the attenuation estimates was characterized using 5,000 nonparametric bootstrap replicates. Variance-inflation factors were examined in the joint MCP-counter model to assess multicollinearity. To exclude direct gene-set overlap as an explanation for the xCell results, the selected xCell monocyte and neutrophil signatures were inspected for direct inclusion of NLRP3, PYCARD, or CASP1.

Complete blood counts and differential leukocyte measurements were unavailable. Moreover, the NLRP3 and cell-composition scores were derived from the same bulk expression matrix. These analyses therefore assessed statistical dependence on transcriptome-derived cell-composition estimates rather than measured leukocyte counts, causal mediation, or cell-intrinsic NLRP3 activity.

Exploratory cross-cohort directional assessment in GSE95233

For exploratory cross-cohort assessment, we used GSE95233 [30], an independent GPL570 whole-blood transcriptomic dataset of patients with septic shock. After excluding healthy controls and later time-point samples, the analysis included 51 Day-1 patients with available 28-day survival status, comprising 34 survivors and 17 non-survivors (Supplementary Material 4; Table S3). Age was extracted from the GEO series-matrix metadata and matched by GSM accession. Gene-set scores were calculated using the same mean-Z procedure as in GSE65682 and standardized to a 1-SD scale. Node-level estimates were interpreted only when at least two prespecified genes were detected; single-gene nodes were retained solely for coverage auditing, and nodes with no detected genes were classified as non-estimable. Associations with 28-day mortality were evaluated using age-adjusted logistic regression, with unadjusted models retained as descriptive sensitivity analyses. Given the limited number of deaths, the analysis was interpreted as an exploratory assessment of directional concordance rather than formal external validation or prognostic-model evaluation. Detailed preprocessing, sample selection, gene-set coverage, and results are provided in Supplementary Material 4 and Tables S3–S5.

Missing data

The primary outcome and age were complete among all 479 patients included in GSE65682. Age and 28-day survival status were also complete among the 51 Day-1 GSE95233 patients included in the exploratory cross-cohort assessment. Therefore, no imputation was performed in either cohort. Gene-set scores were computed only from genes present in each expression matrix; missing genes from the locked prior were handled through the predefined overlap rule rather than imputation.

Software & reproducibility

All analyses were conducted in R version 4.5.0. Gene-set scoring, logistic regression, FDR adjustment, restricted cubic spline analysis, permutation testing, and plotting were implemented using scripted and version-controlled workflows. The integrated and node-level GMT files, harmonized result tables, and analysis scripts are provided in the accompanying data package to support reproducibility. Package versions and executable scripts are provided in the code directory.

Results

Construction of the ASTRA mechanistic prior

The searches yielded 7774 database records. The original deduplication procedure removed 3082 rows, leaving 4692 bibliographic records. A first rule-assisted screen removed 35 records on the basis of publication type or retraction status, and 4657 records entered the second title/abstract-level prioritization stage. The archived output retained 2222 candidate-record rows for mechanistic source discovery. This number represents the row count of the archived prioritization output rather than 2222 unique publications, full-text-included studies, or independent experimental evidence observations (Fig. 1).

Fig. 1.

Literature retrieval, rule-assisted candidate-record prioritization, and construction of the locked ASTRA node–gene prior. Legend Database searching identified 7774 records. The archived deduplication procedure removed 3082 rows, leaving 4692 bibliographic records. After exclusion of 35 records by publication-type or retraction rules, 4657 records entered rule-assisted title/abstract-level mechanistic prioritization. The archived output retained 2222 candidate-record rows for broad mechanistic source discovery. These rows were not subjected to complete full-text eligibility assessment or study-level experimental evidence extraction and therefore do not represent 2222 unique included studies or independent evidence observations. Literature-informed node mapping and author-defined representative-gene selection produced a locked prior comprising 37 representative genes across 20 predefined mechanistic nodes. Gene-set composition was fixed before mortality-association modelling

The resulting prior organized Astragalus-related mechanisms into 10 broad mechanistic categories and 20 predefined nodes. These nodes represented immune-inflammatory and homeostatic repair pathways, including TLR4–NF-κB signaling, the NLRP3 inflammasome, MAPK-related stress signaling, AMPK–SIRT1/PGC-1α signaling, NRF2–HO-1 antioxidative response, autophagy/mitophagy, barrier integrity, and mitochondrial homeostasis. The locked prior was converted into one integrated ASTRA gene set comprising 37 unique representative genes and 20 node-level gene sets in GMT format. The accompanying candidate-record tables, node-mapping resources, gene-set files, and reproducible scripts are provided in the public data package.

Coverage of ASTRA gene sets in the GSE65682 sepsis transcriptome

The locked integrated ASTRA prior contained 37 unique representative genes distributed across 20 mechanistic nodes, corresponding to 40 node–gene memberships because CLDN1, OCLN, and TJP1 were assigned to both the barrier and tight-junction nodes. The NLRP3 node was a prespecified three-gene set comprising NLRP3, PYCARD, and CASP1; it was not derived by selecting the strongest mortality-associated genes from the integrated prior. Thirty-three of the 37 integrated genes were represented in the GSE65682 expression matrix, whereas 36 were represented in the Day-1 GSE95233 matrix. All three NLRP3-node genes were available in both datasets and were included in the corresponding node scores. The complete gene composition, node membership, and dataset-specific coverage of the locked prior are reported in Supplementary Table S6.

We next evaluated whether the integrated and node-level gene sets were computable in the GSE65682 whole-blood transcriptomic background. The integrated score was calculated using the 33 detected genes. Eleven of the 20 node-level sets met the prespecified eligibility threshold of at least two detected genes and were retained for association analyses: NLRP3, NF-κB, TLR4, SCFA–FFAR2, AKT, JAK, autophagy, NRF2, mitochondrial homeostasis, JNK, and PI3K. The remaining nine nodes—STAT3, p38, mTOR, AMPK, SIRT1, HO-1, HCAR2, barrier, and tight junction—contained fewer than two detected genes and were excluded from formal node-level outcome modeling to reduce instability from very small scores (Fig. 2A).

Fig. 2.

Transcriptomic coverage and node–gene architecture of the locked ASTRA prior. LegendPanel A shows the overlap between the locked ASTRA gene sets and the GSE65682 whole-blood expression matrix. The integrated prior contained 37 genes, of which 33 were detected and used to calculate the integrated score. For each node-level set, bar labels indicate the number of detected genes relative to the total number of genes assigned to that node. Eleven nodes with at least two detected genes were retained for formal association analyses; nodes with fewer than two detected genes were displayed for transparency but excluded from node-level outcome modeling. Panel B shows the mapping from broad mechanistic categories to the 20 predefined nodes and 37 representative genes. Link widths represent node–gene membership rather than evidence strength, effect magnitude, or pathway activity. CLDN1, OCLN, and TJP1 are linked to both the barrier and tight-junction nodes

The node–gene map further showed that ASTRA was not represented as a single undifferentiated botanical signature. Instead, the 37 genes formed a prespecified, literature-informed architecture spanning distinct but partially overlapping immune-inflammatory, metabolic, redox, autophagy, mitochondrial, and barrier-related nodes. The Sankey diagram displays the mapping from broad mechanistic categories to nodes and representative genes, including genes assigned to more than one node (Fig. 2B). This locked architecture provided the basis for the integrated and eligible node-level transcriptomic scores used in the subsequent analyses.

Association between ASTRA transcriptomic scores and 28-day mortality

In the primary mean-Z analysis, the integrated Astragalus-derived transcriptomic score showed a nominal inverse association with 28-day mortality in the age-adjusted logistic model. Each 1-SD increase in the integrated score was associated with lower odds of death at 28 days, but the association did not remain significant after FDR correction across the primary comparison family (OR = 0.80, 95% CI 0.65–0.98; P = 0.031; q = 0.124; Fig. 3A and Table S1). The rank-based singscore analysis showed the same direction of association, supporting directional consistency, although the integrated score was not considered FDR-significant.

Fig. 3.

Associations between ASTRA transcriptomic scores and 28-day mortality. Legend Forest plots show odds ratios for 28-day mortality per 1-SD increase in standardized transcriptomic scores from age-adjusted logistic regression models. Panel A shows the integrated Astragalus-derived score evaluated with mean-Z and singscore. The integrated score showed a nominal inverse association in the primary mean-Z model but did not remain significant after FDR correction. Panel B shows node-level associations using mean-Z scoring among gene sets with at least two overlapping genes in GSE65682. The NLRP3 inflammasome module was the only node that remained significant after FDR correction. NF-κB showed a nominal inverse association but was not FDR-significant. Panel C shows node-level associations using singscore, supporting broad directional concordance for the NLRP3 module while indicating method dependence for weaker nodes. OR < 1 indicates lower odds of 28-day mortality per 1-SD higher transcriptomic score

At the node level, the NLRP3 inflammasome module showed the strongest and most statistically robust association with 28-day mortality. In the primary mean-Z model, each 1-SD increase in the NLRP3 transcriptomic score was associated with lower odds of 28-day mortality, and this association remained significant after Benjamini–Hochberg correction across the primary comparison family (OR = 0.70, 95% CI 0.57–0.86; P = 6.02 × 10⁻⁴; q = 0.0072; Fig. 3B and Table S1). NF-κB also showed a nominal inverse association (OR = 0.78, 95% CI 0.63–0.96; P = 0.020; q = 0.122), but did not survive FDR correction. TLR4 and the remaining eligible node-level scores were not significant after FDR correction.

The singscore analysis broadly supported the direction of the main findings. The NLRP3 module remained inversely associated with 28-day mortality at the nominal level under singscore, whereas several other nodes showed weaker or method-dependent estimates (Fig. 3C). Because singscore and mean-Z use different information from the expression matrix—rank order versus standardized expression magnitude—these results were interpreted as method-comparison evidence rather than independent validation.

Importantly, all gene-set scores were transcriptomic module-level readouts. Therefore, the observed inverse association for the NLRP3 module should not be interpreted as direct evidence of reduced protein-complex inflammasome activation. Instead, it indicates that the predefined NLRP3-related transcriptomic module, as captured in whole blood at the available sampling point, was associated with lower 28-day mortality in this cohort.

Sensitivity and robustness analyses for the NLRP3 module

Inspection of the archived evidence-weighted implementation showed that NLRP3, PYCARD, and CASP1 received identical or near-identical weights. The resulting standardized score was effectively indistinguishable from the unweighted score (Pearson r = 1.00 at the reported precision). Because the weighting scheme lacked meaningful within-node discrimination, this analysis was not treated as an independent robustness test.

Leave-one-gene-out analyses supported the stability of the parsimonious NLRP3 node. After excluding NLRP3, PYCARD, or CASP1 in turn, all reduced two-gene scores retained inverse associations with 28-day mortality, with ORs ranging from 0.68 to 0.74. The reduced scores remained highly correlated with the original three-gene score (Pearson r = 0.889–0.899), indicating that the observed association was not dependent on retaining any single gene (Supplementary Table S7 and Supplementary Figure S2).

Method-concordance analysis compared log odds ratios derived from mean-Z and singscore across the eligible node-level sets (Fig. 4A). Most estimates fell within the same directional quadrant or within the prespecified descriptive equivalence band of |Δlog(OR)| ≤ 0.20. The NLRP3 module was located in the concordant inverse-association quadrant, indicating that the direction of association was stable across scoring approaches. NF-κB and TLR4 showed weaker inverse associations, whereas a minority of nodes showed method-dependent or directionally inconsistent estimates, supporting a cautious interpretation of non-NLRP3 findings.

Fig. 4.

Method agreement and functional-form assessment for ASTRA node-level scores. LegendPanel A shows method agreement between mean-Z and singscore estimates. The X-axis and Y-axis represent log odds ratios from age-adjusted logistic models using mean-Z and singscore, respectively. The dashed diagonal line indicates equality between methods, and the shaded diagonal band indicates the descriptive equivalence range of |Δlog(OR)| ≤ 0.20, corresponding approximately to a ± 22% difference on the odds-ratio scale. Points in the lower-left quadrant indicate inverse associations under both scoring methods. Panel B shows restricted cubic spline analysis for the standardized NLRP3 mean-Z score. The curve is referenced to a score of 0, corresponding to the cohort mean. The model showed an overall association with 28-day mortality without evidence of nonlinearity, supporting an approximately linear association within the observed score range. Alt text Scatter plot comparing node-level log odds ratios between mean-Z and singscore, and restricted cubic spline curve showing the association between the NLRP3 transcriptomic score and 28-day mortality

Restricted cubic spline analysis of the standardized NLRP3 mean-Z score showed evidence of an overall association with 28-day mortality without evidence of nonlinearity (P_overall = 0.017; P_nonlinear = 0.986; Fig. 4B). The fitted curve was consistent with an approximately linear inverse association over the observed score range. This supports the appropriateness of reporting the primary model as an OR per 1-SD increase in the NLRP3 transcriptomic score, while not implying a causal dose–response relationship.

Permutation analyses further supported the empirical robustness of the NLRP3 association. In 10,000 label permutations using the same age-adjusted logistic framework, the NLRP3 mean-Z association was unlikely under random outcome allocation, with empirical P values of 3.0 × 10⁻⁴ in the unstratified permutation analysis and 8.0 × 10⁻⁴ in the age-stratified permutation analysis. Under singscore, the corresponding empirical P values were 2.42 × 10⁻² and 2.04 × 10⁻², respectively (Table S2 and Figure S1).

Taken together, the primary association model, method-concordance analysis, spline diagnostics, and permutation testing indicate that the NLRP3 transcriptomic module showed the most stable mortality-associated signal among the ASTRA-derived nodes in GSE65682. These findings should be interpreted as hypothesis-generating and require further assessment in larger independent sepsis cohorts and, ideally, orthogonal protein or functional inflammasome readouts.

Inflammatory-transcript and estimated cell-composition analyses

In the 479-patient GSE65682 cohort, the locked NLRP3 transcriptomic score was positively correlated with both IL1B expression (Spearman ρ = 0.417, bootstrap 95% CI 0.337–0.493; BH-adjusted P = 1.39 × 10⁻²¹) and IL6 expression (ρ = 0.459, bootstrap 95% CI 0.383–0.531; BH-adjusted P = 4.33 × 10⁻²⁶). Pearson and age-adjusted partial-correlation analyses were directionally consistent (Supplementary Table S8 and Supplementary Figure S3).

The locked NLRP3 score was also positively correlated with the MCP-counter monocytic-lineage score (Spearman ρ = 0.396, bootstrap 95% CI 0.313–0.476) and neutrophil score (ρ = 0.536, bootstrap 95% CI 0.463–0.603). Joint adjustment for both MCP-counter scores attenuated the mortality association from an OR of 0.70 to 0.79 (95% CI 0.58–1.06), corresponding to a 33.8% point attenuation on the absolute log-OR scale. However, uncertainty around this change was substantial (bootstrap 95% CI − 44.0% to 93.2%). In contrast, adjustment for the xCell neutrophil score strengthened the inverse estimate to an OR of 0.64 (95% CI 0.50–0.83), consistent with a method-specific statistical suppression pattern. None of NLRP3, PYCARD, or CASP1 was directly included in the selected xCell monocyte or neutrophil signatures. Collectively, these findings indicate that the whole-blood NLRP3 association partly tracked estimated myeloid-cell composition and was sensitive to the composition-estimation method, but they neither establish that the association was entirely attributable to cell composition nor demonstrate independence from cell composition (Supplementary Table S9 and Supplementary Figure S4).

Exploratory cross-cohort directional assessment in GSE95233

After excluding healthy controls and later time-point samples, the GSE95233 analysis included 51 Day-1 patients with septic shock, comprising 34 survivors and 17 non-survivors. Age was available for all patients and ranged from 25 to 85 years. Gene coverage was complete for the NLRP3 node (3/3 genes) and nearly complete for the integrated ASTRA signature (36/37 genes) (Supplementary Material 4; Tables S3 and S4).

In age-adjusted analyses, each 1-SD increase in the NLRP3 score was associated with lower odds of 28-day mortality, although the estimate was imprecise (OR 0.68, 95% CI 0.37–1.23; P = 0.200). The corresponding unadjusted estimate was OR 0.67 (95% CI 0.37–1.21; P = 0.187). The integrated signature showed a weaker inverse association after age adjustment (OR 0.83, 95% CI 0.46–1.50; P = 0.539), with an unadjusted OR of 0.81 (95% CI 0.46–1.45; P = 0.485; Table S4).

The NLRP3 estimate was therefore directionally concordant with the primary GSE65682 result, but statistically imprecise. Of the 20 prespecified nodes, 15 had at least two detected genes and were eligible for exploratory node-level analysis; four single-gene nodes were retained in Table S5 only as coverage-audit entries, and HCAR2 was non-estimable. Among the eligible nodes, TLR4 and Mitochondria showed nominal associations, but none survived BH correction. Given the small cohort and 17 deaths, these findings provide limited exploratory directional support rather than formal external validation and were not used to redefine the primary NLRP3 finding.

Discussion

Principal findings

Functional NLRP3 activation in experimental sepsis involves inflammasome assembly, caspase-1 activation, cytokine maturation, and pyroptosis and is generally considered detrimental. The present score, however, measured only whole-blood mRNA abundance of NLRP3, PYCARD, and CASP1. Its positive correlations with IL1B and IL6 indicate that the inverse mortality association did not reflect a transcriptionally quiescent inflammatory state and do not support a simple separation between the score and concurrent inflammatory transcription. Because neither mature cytokines nor functional inflammasome readouts were available, these transcript-level correlations do not resolve the directional paradox.

Deconvolution provided a second interpretive boundary. The score correlated with MCP-counter monocytic-lineage and neutrophil estimates; separate adjustment modestly attenuated the association, and joint adjustment shifted the OR to 0.79, although bootstrap uncertainty was wide. xCell neutrophil adjustment instead strengthened the inverse estimate, consistent with a method-specific statistical suppression pattern. Estimated myeloid composition may therefore contribute to the whole-blood signal, but the data do not establish that the association is either entirely attributable to or independent of cell composition. In the absence of verified Astragalus exposure and signed pharmacodynamic encoding, the score should be interpreted as an observational, composition-sensitive transcriptomic state rather than functional NLRP3 activity or Astragalus target engagement. The age-adjusted GSE95233 estimate remained directionally concordant but imprecise and therefore provides only limited cross-cohort support.

Directional and cross-modal interpretation of the NLRP3-linked transcriptomic signal

NLRP3 remains mechanistically relevant to sepsis because TLR4–NF-κB-dependent priming increases the expression of inflammasome components, whereas microbial and damage-associated stress signals promote inflammasome assembly and downstream caspase-1 activation. Activated inflammasome signaling drives IL-1β and IL-18 maturation and GSDMD-dependent pyroptosis; excessive functional activation has generally been linked to tissue injury and adverse experimental sepsis phenotypes [12, 13, 31–33].

The estimated cell-composition analyses further narrow the interpretation of the whole-blood NLRP3 signal. Its positive correlations with MCP-counter monocytic-lineage and neutrophil scores indicate that the three-gene module partly tracks variation in the myeloid composition or myeloid transcriptional state of whole blood. Adjustment for either MCP-counter score separately produced modest attenuation, while joint adjustment reduced the log-OR magnitude by approximately one third and widened the confidence interval across the null. However, the bootstrap interval for attenuation was broad, precluding a firm conclusion that cellular composition explained the association.

The xCell neutrophil sensitivity analysis produced the opposite coefficient change, with a stronger inverse NLRP3 estimate after adjustment. Because the xCell neutrophil covariate was itself statistically imprecise and changed coefficient direction after inclusion of the NLRP3 score, this result is best regarded as a method-specific statistical suppression pattern rather than a biological interaction. The selected xCell signatures did not directly contain NLRP3, PYCARD, or CASP1, reducing concern about direct gene-set overlap, but both the predictor and the composition estimates remained derived from the same bulk expression matrix. Accordingly, the current data cannot distinguish differences in leukocyte abundance from cell-intrinsic expression, immune activation state, or other correlated sepsis endotypes.

In the clinical cohort, however, higher whole-blood expression scores derived from NLRP3, PYCARD, and CASP1 were associated with lower 28-day mortality in GSE65682. This directional discordance does not indicate that functional NLRP3 inflammasome activation is protective, nor does it demonstrate that Astragalus suppresses NLRP3 activity in patients. The score quantified coordinated transcript abundance rather than inflammasome assembly, caspase-1 cleavage, cytokine maturation, GSDMD pore formation, or ASC-speck formation. Transcriptome-derived cell-composition analyses indicated that estimated leukocyte composition may contribute to the observed association, although the direction and magnitude of adjustment were method-dependent and these analyses could not distinguish cellular abundance from within-cell transcriptional regulation. Potential contributions from mRNA–protein discordance, sampling time, and sepsis endotype remain unresolved and should be regarded as hypothesis-generating. Whole-blood transcriptomes in sepsis capture heterogeneous and time-dependent host-response states and cannot be assumed to represent pathway activity linearly [3–5]. The age-adjusted estimate in GSE95233 remained directionally concordant with the primary GSE65682 result and differed little from the corresponding unadjusted estimate. However, harmonizing the statistical model did not strengthen the underlying evidence: the external cohort contained only 17 deaths, and the resulting wide confidence interval precludes claims of robust replication, independent prognostic performance, or clinical transportability. The GSE95233 result therefore provides only limited exploratory cross-cohort directional support and does not constitute formal external or functional validation. Accordingly, the finding is best interpreted as an inverse mortality-associated NLRP3-linked transcriptomic pattern whose biological basis requires cell-resolved and protein- or function-level investigation.

A related cross-modal limitation concerns the apparent contrast between experimental Astragalus-associated NLRP3 inhibition and the inverse mortality association of the clinical mRNA score. Experimental studies of astragaloside IV in sepsis models have reported attenuation of NLRP3-related inflammatory activity and intestinal injury [34, 35]. However, GSE65682 contained no verified information on Astragalus exposure, and the present score reflected baseline whole-blood expression rather than treatment-induced molecular change. Moreover, experimental directionality annotations were not encoded as signed weights in the transcriptomic score. A higher score therefore indicates greater relative expression of NLRP3, PYCARD, and CASP1, not stronger Astragalus-like inhibition, functional inflammasome activation, or pharmacological target engagement. In this application, ASTRA tests whether an literature-informed node can be represented and evaluated in clinical transcriptomes; it does not establish directional pharmacodynamic concordance. Such concordance would require intervention datasets with verified Astragalus exposure and matched transcriptomic, protein-level, and functional inflammasome measurements.

Relation to sepsis transcriptomics and mechanism-guided scoring

Previous sepsis transcriptomic studies have identified reproducible diagnostic signatures, host-response endotypes, and outcome-associated molecular patterns using whole-blood expression data [4–6, 36]. Many of these analyses rely on data-driven signatures or general-purpose enrichment frameworks. Such approaches are powerful, but their biological interpretation can be limited by pathway redundancy, overlapping gene membership, and the lack of a direct connection between the statistical signal and a predefined mechanistic hypothesis [7–9].

The present work takes a complementary approach. Instead of beginning with differentially expressed genes and then searching post hoc for enriched pathways, we first organized literature-informed Astragalus-related mechanistic concepts into a predefined, author-defined node–gene framework and then tested whether these nodes generated interpretable patient-level transcriptomic scores in sepsis. This prior-driven design does not eliminate confounding or prove mechanism, but it improves traceability: each score can be mapped back to a defined mechanistic node and a small set of representative genes. In this sense, ASTRA provides a structured way to ask whether a botanical mechanistic hypothesis leaves a measurable transcriptomic footprint in human sepsis data.

The results also suggest that not all prespecified ASTRA nodes are equally informative in the available whole-blood transcriptomic setting. The integrated score showed only nominal evidence, and most node-level scores did not survive FDR correction. This pattern argues against overinterpreting ASTRA as a broad protective signature. Instead, the current evidence points more narrowly to the NLRP3-related transcriptomic module as the most stable mortality-associated component in the primary dataset, with limited and statistically imprecise directional support from GSE95233. This narrower conclusion is scientifically stronger and more defensible than claiming a generalized Astragalus protective effect across all inflammatory and repair pathways.

Methodological implications

Several methodological features strengthen the internal consistency of the analysis. First, the ASTRA prior was constructed before outcome modeling and then converted into integrated and node-level gene sets, reducing post hoc selection of genes from the clinical dataset. Second, the primary analysis used a uniform age-adjusted logistic framework across eligible gene sets, and multiplicity was handled using Benjamini–Hochberg correction across the primary mean-Z comparison family. Third, the NLRP3 finding was supported by sensitivity analyses, including rank-based singscore comparison, restricted cubic spline assessment, and label-permutation testing.

At the same time, these robustness checks should not be overinterpreted. Agreement between mean-Z and singscore supports scoring-method stability, but it is not external validation. Restricted cubic spline analysis supports an approximately linear association over the observed score range, but it does not establish a biological dose–response relationship. Permutation testing indicates that the observed NLRP3 association is unlikely under random outcome allocation within this dataset, but it does not address unmeasured confounding, immune-cell composition, treatment effects, or cohort-specific sampling structure. Therefore, the evidence supports hypothesis generation and mechanistic prioritization rather than clinical deployment.

Strengths & limitations

This study has several strengths. It integrates a literature-informed, author-defined mechanistic prior with patient-level transcriptomic analysis, providing a reproducible route from predefined biological hypotheses to clinical-cohort assessment. The node-based architecture improves interpretability by linking statistical estimates to specified immune-inflammatory and homeostatic-repair modules rather than to broad, overlapping pathway labels. Analytical robustness was examined through two single-sample scoring methods, explicit gene-set coverage checks, false-discovery-rate correction, functional-form assessment, permutation testing, leave-one-gene-out analysis, inflammatory-transcript correlations, and transcriptome-derived cell-composition sensitivity analyses. The locked ASTRA resources and analysis files were organized for reuse, and the principal NLRP3 estimate was additionally examined in GSE95233 using the same age-adjusted logistic-regression framework without outcome-driven gene reselection or model tuning. Although imprecise, the directionally concordant GSE95233 estimate provided limited cross-cohort support.

The limitations define the boundaries of interpretation. The primary analysis used an observational public transcriptomic cohort and therefore cannot establish causality, clinical utility, treatment responsiveness, or therapeutic efficacy. Models were adjusted only for age because detailed comorbidities, treatment exposures, immunomodulatory medications, sampling times, organ-support variables, infection characteristics, and longitudinal trajectories were unavailable or incompletely harmonized. Residual confounding is therefore likely. Direct leukocyte counts and protein- or function-level inflammasome measurements were also unavailable. Transcriptome-derived deconvolution and IL1B/IL6 correlations provided exploratory assessments of cellular composition and concurrent inflammatory transcription, but all variables originated from the same bulk whole-blood expression matrix. These analyses cannot distinguish cellular abundance from within-cell transcriptional regulation, establish causal mediation, or identify the leukocyte population responsible for the mortality-associated signal. MCP-counter and xCell scores are not interchangeable with measured absolute cell counts or proportions, and the xCell monocyte score showed marked practical zero inflation and was unsuitable for inferential adjustment. Accordingly, the NLRP3 score cannot be interpreted as a direct measure of cell-intrinsic inflammasome activation, protein-level pathway activity, pharmacological target engagement, or a protective NLRP3 state.

The construction of the mechanistic prior also has important limitations. The broad literature pool was generated through rule-assisted bibliographic and title/abstract-level prioritization rather than complete full-text evidence extraction for every candidate record. Residual duplicate identifiers, abbreviation-related false-positive matches, archive inconsistencies, publication-selection bias, and incomplete directionality annotation may remain. The 2,222-record archive should therefore be interpreted as a source-discovery resource rather than as a quantitatively synthesized evidence base. Moreover, the original workflow did not preserve a complete publication-by-gene audit trail or a formally discriminative evidence-grading system for all 37 representative genes. The final gene sets should consequently be regarded as literature-informed, author-defined mechanistic priors rather than products of systematic quantitative evidence synthesis. These provenance limitations affect evidentiary strength and generalizability but do not alter the reproducibility of the locked scoring analyses.

Finally, the exploratory cross-cohort assessment was constrained by the small GSE95233 sample and only 17 deaths. Although sample-level age was complete and the modeling framework was harmonized with GSE65682, confidence intervals remained wide. The GSE95233 cohort therefore does not establish formal validation, robust replication, independent prognostic performance, clinical transportability, or biological confirmation. Generalizability across sepsis cohorts, platforms, disease stages, infection sources, treatment settings, and sampling windows remains uncertain.

Future directions

Future validation should prioritize larger independent sepsis cohorts with measured differential leukocyte counts or cell-resolved transcriptomics, matched protein- or function-level inflammasome readouts, standardized sampling times, broader clinical covariates, and verified treatment exposure. Longitudinal sampling linked to organ-dysfunction and recovery trajectories would help determine whether ASTRA-derived scores reflect dynamic host-response states rather than static baseline risk. Further refinement of the mechanistic prior should include a publication-by-gene audit trail, explicit directionality annotation, and prospectively locked node definitions before clinical or therapeutic interpretation.

Conclusion

ASTRA provides a reproducible, literature-informed framework for mechanism-guided transcriptomic scoring in sepsis. The NLRP3 module showed the most stable mortality-associated signal, but it co-varied with inflammatory transcription and estimated myeloid composition and showed method-dependent changes after composition adjustment. The finding should remain an observational, hypothesis-generating blood transcriptomic association rather than evidence of protective inflammasome activity, Astragalus efficacy, or pharmacological target engagement.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

X.Y., T.H., J.W., Z.G., W.Y., and B.G. contributed to this work as follows: X.Y. and B.G. conceptualized the study. X.Y., T.H., Z.G., and B.G. developed the methodology. X.Y. and J.W. curated the data. X.Y., T.H., and Z.G. performed the formal analyses, with T.H. leading the software implementation and validation. X.Y., T.H., and J.W. created visualizations. X.Y. drafted the original manuscript. X.Y., T.H., J.W., Z.G., W.Y., and B.G. reviewed and edited the manuscript. W.Y. and B.G. provided supervision and project administration and acquired funding. All authors reviewed and approved the final manuscript.

Funding

This work was supported by the Military Key Discipline Construction Projects of China (HL21JD1206).

Data availability

The clinical transcriptomic and phenotype data analyzed in this study are publicly available from the NCBI Gene Expression Omnibus under accession numbers GSE65682 and GSE95233. The ASTRA mechanistic prior, harmonized analytical inputs, derived result tables, figure source data, and R scripts required to reproduce the reported analyses are available through Mendeley Data at DOI: 10.17632/82nyt2mxft.2.

Declarations

Conflict of interest

The authors declare no competing interests.

Ethical approval and consent to participate

This study involved a secondary analysis of publicly available, de-identified transcriptomic and clinical data from the NCBI Gene Expression Omnibus datasets GSE65682 and GSE95233. No new participants were recruited, and no identifiable personal information was accessed. Therefore, additional ethics approval and consent to participate were not required for the present study. The original studies were conducted in accordance with the ethical approvals and informed-consent procedures reported by the respective investigators.

Consent for publication

Not applicable. The manuscript contains no identifiable individual-level information, images, or clinical case details.