The analytical strategy behind the study was deliberately layered. The researchers began with GSE134347, a bulk microarray dataset generated on the Affymetrix Human Transcriptome Array 2.0 platform, which contains transcriptomic profiles from 156 patients with sepsis and 83 healthy controls. To balance group sizes and reduce analytical heterogeneity, they randomly selected 50 sepsis samples and 50 healthy controls, performed principal component analysis for quality control, excluded outlier samples, and normalized the remaining data before downstream analysis. Using an absolute log2 fold-change cutoff of 1 and a false discovery rate threshold of 0.05, the differential-expression analysis identified 718 differentially expressed messenger RNAs in sepsis patients relative to controls, comprising 299 upregulated and 419 downregulated genes.

In parallel, the team integrated two single-cell RNA-sequencing datasets, GSE167363 and GSE220189, drawing on five sepsis patients and two healthy individuals from the former and five healthy individuals from the latter. After stringent quality control, which retained cells with at least 200 detected genes, at least 1,000 unique molecular identifiers, a log10 genes-per-UMI ratio of at least 0.7, mitochondrial gene proportions no higher than 30 percent, and hemoglobin gene proportions no higher than 5 percent, putative doublets were removed with DoubletFinder. Harmony was applied to correct batch effects, and cell identities were annotated using SingleR against the Human Primary Cell Atlas reference. This single-cell analysis identified 243 differentially expressed mRNAs, including 153 upregulated and 90 downregulated genes, revealing substantial transcriptomic alterations at the individual-cell level that mirrored the bulk-tissue signal.

Functional enrichment analysis of the bulk-derived differentially expressed genes showed that they were mainly concentrated in immune-related cellular components and molecular functions, while Kyoto Encyclopedia of Genes and Genomes pathway analysis highlighted inflammation- and immunity-related pathways. In the single-cell datasets, the enriched genes were predominantly involved in antiviral responses, immune responses, and immune system processes. Venn analysis then identified 57 overlapping differentially expressed mRNAs between the bulk microarray dataset and the single-cell data. Protein-protein interaction analysis of these overlapping genes, performed with the STRING database at a minimum interaction score of 0.4, revealed two major hub modules: one centered on CD3D, CD2, GZMK, and CCL5, and another centered on THBS1, MMP9, and PPBP. Given its high network connectivity, consistent differential expression in public datasets, and potential prognostic relevance, THBS1 was selected as the key candidate target for all subsequent analyses.

The clinical component of the study reused an exploratory peripheral-blood RNA-sequencing cohort of 23 patients with sepsis, diagnosed according to the Sepsis-3 criteria, and 10 healthy controls recruited from the emergency intensive care unit of the Affiliated Hospital of Southwest Medical University. Peripheral blood was collected into EDTA tubes, total RNA was extracted with TRIzol reagent, and paired-end sequencing was performed on the BGISEQ-500/MGISEQ-2000 platform at BGI-Shenzhen. In this cohort, the tabulated THBS1 expression values had means of 12.510 in sepsis patients and 10.881 in controls, yielding a descriptive log2 ratio of +0.201 and a standardized mean difference (Hedges’ g) of +0.171. An exploratory two-sided exact Mann-Whitney test did not detect a statistically significant between-group difference (P = 0.893). The authors explicitly state that, because the cohort was small and the distributions variable, these results do not support clinical validation, and the cohort was retained only as an exploratory description rather than presented as an independent validation set.

Prognostic relevance was explored in the public sepsis dataset GSE65682, which contains peripheral blood transcriptomic data and clinical follow-up information. Patients were divided into high- and low-expression groups according to the median expression level of each candidate gene, and Kaplan-Meier curves were compared with the log-rank test. The patterns diverged sharply by gene: high expression of CD3D and CD2 was associated with better survival, whereas high THBS1 expression was associated with significantly shorter survival, with a log-rank P value of 0.0029. The association for PPBP did not reach statistical significance (P = 0.36). The authors caution that this THBS1 survival association is exploratory, because SOFA and APACHE II severity scores were not uniformly available in the dataset, so a combined prognostic model adjusting for baseline disease severity could not be constructed, and residual confounding by severity cannot be excluded.

To test whether the differential expression of THBS1 held up across independent cohorts, the researchers conducted a meta-analysis of ten public GEO datasets, encompassing 867 patients with sepsis and 307 healthy controls. Because expression values came from different platforms and preprocessing pipelines, standardized mean difference was used as the pooled effect size, with a random-effects model applied given substantial heterogeneity. The pooled random-effects estimate was −0.7847 (95 percent confidence interval, −1.2030 to −0.3664), with heterogeneity of I² = 86 percent. Because the effect was calculated as healthy control minus sepsis, the negative sign denotes higher THBS1 transcript levels in sepsis. A leave-one-out sensitivity analysis showed that the pooled estimate remained negative and statistically significant after the sequential exclusion of each dataset, ranging from −0.897 to −0.674, with I² values between 74.1 and 84.3 percent, indicating that no single dataset drove the direction or significance of the pooled association.

An exploratory platform-stratified analysis added nuance to this picture. Across five microarray datasets the pooled SMD was −1.0804 (95 percent CI, −1.5780 to −0.5829; I² = 82.33 percent), while across five RNA-sequencing datasets it was −0.4475 (95 percent CI, −0.8149 to −0.0801; I² = 50.76 percent). Both estimates were negative, but the nominal between-platform P value of 0.0449 was not robust to exclusion of the discovery dataset GSE134347; when that dataset was removed, the microarray subgroup estimate shifted to −0.9680 and the between-group P value rose to 0.1888. The authors emphasize that the pooled SMD quantifies a standardized difference in transcript abundance, is neither a fold change nor a measure of diagnostic performance, and should be read as exploratory evidence of a cross-cohort transcript-level association rather than a precise universal effect estimate. They also stress that case-control expression differences and within-disease prognostic stratification address different analytical questions and must be interpreted separately.

Perhaps the most visually striking result came from the single-cell localization analysis. After quality control and doublet removal across 12 samples, a total of 60,480 cells were retained, with per-sample cell counts ranging from 2,197 to 8,676, mean detected genes per cell of 1,173 to 1,909, and mean UMI counts of 4,194 to 7,859. The cells were annotated into populations including T cells, monocytes, B cells, natural killer cells, neutrophils, macrophages, dendritic cells, and progenitor-related populations. THBS1 was detectable across multiple sepsis-related immune-cell populations, but it appeared visually more abundant in subsets of the myeloid-related cell populations, with lower expression observed in B cells, natural killer cells, and neutrophils. The authors note that these observations are descriptive only and were not subjected to formal statistical testing between cell types or disease states, so the myeloid localization remains a visual and hypothesis-level finding.

The study closed with a computational therapeutic exploration. Candidate compounds associated with THBS1 were screened using the HERB 2.0 database, and two drug-like molecules, deoxycholic acid and resveratrol, were selected for exploratory molecular docking against the human THBS1 N-terminal domain crystal structure (PDB ID 2ERF, resolved at 1.45 angstroms). AutoDock Vina predicted binding energies of −6.6 kcal/mol for deoxycholic acid and −6.4 kcal/mol for resveratrol, both below the −5.0 kcal/mol threshold the authors used to indicate favorable predicted binding. The team is careful to stress that these are computational estimates only: without experimental binding controls or functional assays, the docking scores cannot establish direct THBS1 binding or therapeutic activity in sepsis. Taken together, the authors conclude that THBS1 may be associated with myeloid-related immune activity and adverse prognosis in sepsis, but that protein-level validation, functional experiments in macrophage models, larger prospective multicenter cohorts, and attention to platelet-related confounding are all required before THBS1 can be considered a clinically applicable biomarker or therapeutic target.

Subject of Research: Integrated bulk and single-cell transcriptomic analysis of THBS1 as a candidate myeloid-associated biomarker in sepsis

Article Title: Integrated bulk and single-cell transcriptomic analyses suggest a myeloid-associated candidate role of THBS1 in sepsis

Article References: Chen, C., Jin, Y., Chen, W., Yuan, H., Hu, Y., & Hu, J. (2026). Integrated bulk and single-cell transcriptomic analyses suggest a myeloid-associated candidate role of THBS1 in sepsis. Frontiers in Cellular and Infection Microbiology, 16, Article 1891550. https://doi.org/10.3389/fcimb.2026.1891550

Image Credits: AI Generated

DOI: 10.3389/fcimb.2026.1891550

Keywords: sepsis, THBS1, transcriptomics, single-cell RNA sequencing, myeloid cells, biomarkers, meta-analysis, molecular docking, prognosis, inflammation, macrophages, GEO datasets