This study developed ASTRA, an author-defined, literature-informed Astragalus mechanistic prior, to derive mechanism-guided whole-blood transcriptomic scores and test their association with 28‑day mortality in ICU sepsis. The objective was to evaluate whether integrated and node-level transcriptomic scores anchored to hypothesized Astragalus-related mechanisms, and specifically to inflammasome biology, related to mortality risk.
The primary analysis used publicly available transcriptomic and clinical data from GSE65682 comprising 479 adults with sepsis. An exploratory directional cross-cohort assessment evaluated Day‑1 samples from 51 patients with septic shock in GSE95233. Analyses used mean‑Z scores as the primary scoring metric and applied age‑adjusted logistic regression with false‑discovery‑rate (FDR) correction to test associations with 28‑day mortality.
Sensitivity and stability checks included alternate scoring (singscore), restricted cubic splines to assess nonlinearity, label permutation tests, leave‑one‑gene‑out analyses, correlations with IL1B and IL6 mRNA, and adjustment for transcriptome-derived estimates of myeloid cell composition from MCP-counter and xCell.
ASTRA produced integrated scores reflecting the assembled mechanistic prior and node-level scores for component mechanisms. The authors highlighted a three‑gene node linked to the NLRP3 inflammasome as of particular interest based on literature-informed selection and observed signal strength. Mean‑Z scoring was the main metric reported, with supplementary sensitivity scoring approaches conducted to evaluate robustness.
The integrated ASTRA score showed a nominal inverse association with 28‑day mortality but did not survive FDR correction. In contrast, a node-level, NLRP3-related three‑gene score demonstrated the strongest and statistically robust association in GSE65682: odds ratio (OR) per 1‑SD increase 0.70 (95% CI 0.57–0.86; q = 0.0072). Leave‑one‑gene‑out analyses preserved the directionality of this association, indicating no single gene fully drove the result.
The NLRP3 score correlated positively with IL1B and IL6 mRNA levels and with MCP-counter estimates of monocytic-lineage and neutrophil abundance, suggesting that this transcriptomic signal partly reflects concurrent inflammatory transcription and estimated myeloid-cell composition.
In the smaller GSE95233 cohort of 51 Day‑1 septic shock patients, the age‑adjusted NLRP3 estimate was directionally concordant with GSE65682 but imprecise: OR 0.68 (95% CI 0.37–1.23). The authors present this as an exploratory directional assessment rather than a definitive replication given limited precision.
Several sensitivity approaches were performed to evaluate robustness. Correlations with IL1B and IL6 supported a link with inflammatory transcription. Adjustment for estimated cell composition had method-sensitive effects: joint adjustment using MCP-counter attenuated the NLRP3 association to OR 0.79 (95% CI 0.58–1.06), while adjustment for xCell neutrophil estimates strengthened the association to OR 0.64 (95% CI 0.50–0.83). These divergent effects indicate that the observed NLRP3-linked transcriptomic state partly tracks estimated myeloid-cell composition and that results depend on the cell‑deconvolution method used.
Other sensitivity checks included singscore recalculation, restricted cubic spline evaluation for nonlinearity, label permutation testing to assess chance findings, and leave‑one‑gene‑out analyses to test influence of individual genes; the NLRP3 signal remained directionally stable across these checks.
The authors interpret ASTRA as a reproducible framework for mechanism-guided transcriptomic scoring in sepsis. The NLRP3-linked signal represents an observational whole-blood transcriptomic state associated with lower observed 28‑day mortality in the primary cohort and directionally concordant in an exploratory cohort. Importantly, authors caution that these observational transcriptomic associations do not establish functional or causal conclusions: they do not demonstrate protective inflammasome activity, efficacy of Astragalus as a therapy, or engagement of a pharmacologic target.
Limitations explicitly reported include reliance on secondary analysis of publicly available datasets, dependence of findings on cell‑composition estimation methods, limited precision in the smaller exploratory cohort, and the observational nature of transcriptomic associations that cannot infer causality or therapeutic effect.
ASTRA offers a mechanism‑guided approach for generating transcriptomic scores tied to hypothesized biological priors. The NLRP3 three‑gene node produced the most consistent mortality-associated signal in GSE65682, with evidence that the signal partly tracks myeloid-cell estimates and inflammatory gene expression. The authors emphasize that further experimental and clinical work would be required to establish any causal role for inflammasome biology in outcomes, or to evaluate Astragalus-related interventions. The study was a secondary analysis of de‑identified public data and the authors declared no competing interests.