Genotype stratified adjunctive dexamethasone for tuberculous meningitis in HIV-negative Adults: the LAST ACT trial

Tuberculous meningitis (TBM) is the most severe form of tuberculosis. Adjunctive corticosteroids are recommended for HIV-negative adults, although their benefit appears modest and may depend on host leukotriene A4 hydrolase (LTA4H) genotype. The LAST ACT trial (NCT03100786) was a genotype-stratified, randomised, double-blind, placebo-controlled Phase III trial that evaluated dexamethasone in HIV-negative Vietnamese adults with TBM. A total of 613 adults with LTA4H CC or CT genotypes were randomised to receive dexamethasone or placebo; 89 TT-genotype participants received open-label dexamethasone. The trial found no benefit from adjunctive dexamethasone in CC/CT-genotype participants.

This dataset supports the analysis of a secondary outcome: changes in blood and CSF inflammatory responses. Whole-blood RNA sequencing was performed for 202 participants after quality control (day 0: n=202; day 14: n=188; day 60: n=153). CSF inflammatory proteins were measured in 646 participants using the Olink Explore 384 Inflammation panel (day 0: n=638; day 30: n=391), resulting in 1029 CSF samples with high-quality data after quality control. Ten pre-specified cytokines were targeted, but four (IL-2, IL-4, IL-5, IL-13) were excluded due to poor detection.

Analyses focused on five inflammation-related pathways: TNF signalling, interferons, cytokines, neutrophils, and eicosanoids. Pathway activity (enrichment scores) was calculated using single-sample gene set enrichment analysis (ssGSEA) and z-score metrics. Longitudinal changes (rate of reduction) in pathway activity by treatment and LTA4H genotype were assessed using Bayesian joint models (JMbayes2), which combined linear mixed effects and survival components, accounting for early mortality bias. This dataset enabled a detailed investigation of corticosteroid effects on host inflammatory responses in TBM.

Methods
CSF inflammatory proteins were measured using the Olink Explore 384 Inflammation Panel (Olink Proteomics, Uppsala, Sweden). Olink measurements were conducted for 675 participants on day 0 (n=675) and day 30 (n=397) at the Human Genomics Facility of the Genetic Laboratory, Department of Internal Medicine, Erasmus MC (Rotterdam, Netherlands). Raw protein expression data were reported as Normalized Protein eXpression (NPX) units, which were log₂-transformed and normalized using the plate control method to minimize technical variation. To correct for batch effects, the NPX values then were further adjusted using the ComBat function from the sva R package.31 Quality control (QC) procedures were conducted at both the sample and protein levels. At the sample level, poor-quality samples were excluded if they exhibited a failure rate of ≥50% across protein assays, as determined by Olink’s internal QC criteria. Outliers were identified via principal component analysis (PCA) and excluded if they deviated by more than three standard deviations from the first principal component. At the protein level, proteins with a limit of detection exceeding 75% of samples were filtered out. Finally, 17 participants who died before randomization were excluded, resulting in a final dataset of 275 proteins in 1029 CSF samples from 646 participants available for analysis (day 0: *n=638; day 30: n=*391). Among the 10 planned CSF cytokines - TNF-α, interleukin (IL)-1β, IL-2, IL-6, IL-12b, interferon (IFN)-γ, IL-4, IL-5, IL-10, and IL-13 - four (IL-2, IL-4, IL-5, and IL-13) did not pass quality control due to their limit of detection (LOD) exceeding 75% of samples.

Whole blood RNA sequencing was performed for the first 207 consecutively enrolled participants on day 0 (n=207), day 14 (n=191), and day 60 (n=156). Whole blood samples were preserved in PAXgene Blood RNA collection tubes at -80°C. Total RNA was subsequently extracted using the PAXgene Blood RNA Kit (Qiagen, Valencia, CA, USA), following the manufacturer's protocol. Extracted RNA was shipped to the Ramaciotti Centre for Genomics (University of New South Wales, Sydney, Australia) for high-throughput sequencing. Library preparation was performed using the TruSeq Stranded Total RNA with Ribo-Zero Globin kit (Illumina, San Diego, CA, USA) to deplete globin transcripts and ribosomal RNA. Sequencing was conducted on the Illumina NovaSeq 6000 platform, generating approximately 30 million 100 bp paired-end reads per sample. Raw sequencing data was subjected to quality control and aligned to the human reference genome (GRCh38 build 99) with the STAR aligner (v2.5.2a).32 Gene-level quantification from aligned reads was performed using FeatureCounts (v2.0.0), generating raw counts for 60,067 genes.33 Prior to analysis, five participants were excluded: two who died before randomization and three whose RNA sequencing data were poor quality (RNA integrity number < 4 and uniquely mapped reads < 10 million). This resulted in a final sequencing dataset of 202 participants (day 0: n=202, day 14: n=188, day 60: n=153). To further clean the dataset, hemoglobin genes, ribosomal RNA genes, and genes with low expression (median count < 10) were filtered out, reducing the gene set to 20,533. Gene expression values were then normalized and log₂-transformed using the variance stabilizing transformation algorithm implemented in the DESeq2 package in R (v1.34.0) to enable downstream statistical analyses.34 For each targeted pathway, a single sample enrichment score was calculated using the z-score method to evaluate the activity of the pathway at each time point for both whole blood transcriptomics and CSF proteomics, for each patient.

Statistical analysis

We followed the published statistical analysis plan of the LAST ACT clinical trial, a leukotriene A4 hydrolase–stratified non-inferiority trial of adjunctive corticosteroids in HIV-negative adults with tuberculous meningitis (DOI: 10.12688/wellcomeopenres.22498.2). The present dataset was used to analyse secondary endpoints described in Section 6, Measurements of blood and cerebrospinal fluid inflammation.

We employed a joint modelling framework that combined a survival model with a linear mixed-effects model for longitudinal blood and CSF markers. The survival sub-model consisted of time to all-cause mortality within three months as the outcome, with covariates including treatment allocation and the subject-specific fitted values of longitudinal marker measurements. Models were estimated in a Bayesian framework using the R package JMbayes2. This approach accounts for informative dropout due to early death, occurring within the first 60 days for transcriptomic analyses and the first 30 days for proteomic analyses after randomisation.