Meta-analysis of tomato RNA-seq data pinpoints drought-responsive genes

Drought is one of the most damaging stresses that tomato plants can face, and it disrupts growth, yield, and fruit quality in ways that matter to growers worldwide. Yet when individual laboratories use RNA sequencing to find the genes that respond to water deficit, their candidate lists often overlap only partially. Differences in cultivars, drought-induction methods, treatment duration, and analytical pipelines mean that a gene highlighted in one experiment may barely register in another. A new study in Applied Biological Chemistry tackles this reproducibility problem head-on by combining four independent public RNA-seq datasets into a single meta-analysis that explicitly measures how consistent each gene’s response is across studies.

The research, led by Hongman Moon, Hyejin Cho, and Chanseok Shin, analyzed four BioProjects from the NCBI Sequence Read Archive—PRJEB83804, PRJNA624892, PRJNA947261, and PRJNA1298269—comprising 51 tomato leaf samples. The datasets represented different water-deficit induction systems, including soil drying, dehydration, and polyethylene glycol-induced osmotic stress. Rather than simply intersecting lists of differentially expressed genes, the team processed all raw reads through a unified pipeline, aligning them to the tomato reference genome SL4.0 with HISAT2 and counting gene-level reads with featureCounts using the ITAG4.0 annotation. Differential expression was then computed independently for each BioProject with DESeq2, defining genes with an adjusted P-value below 0.05 and an absolute log2 fold-change of at least 1 as differentially expressed.

Before any pooling, the four datasets showed striking variability. Principal component analysis of normalized expression profiles separated samples mainly by BioProject rather than by drought treatment alone, although control and treated samples showed partial separation within some datasets. The number of differentially expressed genes differed markedly among studies, with PRJNA947261 showing the largest response and PRJNA1298269 and PRJNA624892 showing smaller sets. Pairwise comparisons revealed that the proportion of shared genes regulated in the same direction ranged from 37 to 89 percent, meaning that genes flagged in two datasets were not always regulated the same way. This pre-analysis variability, the authors argue, is precisely why simple overlap-based selection can mislead.

The core of the method lies in how the study integrates evidence. Two complementary approaches were applied: Fisher’s combined probability test, which integrates nominal P-values from the individual analyses, and an inverse-variance weighted fixed-effect model that pools log2 fold-change estimates and their standard errors to produce summary effect sizes, Z statistics, and P-values. The fixed-effect model served as a precision-weighted screening estimator of expression-change direction rather than an assumption of a single universal biological effect. A Paule-Mandel random-effects model was added as a sensitivity check on the final filtered gene set. Multiple testing was controlled with the Benjamini-Hochberg false discovery rate method.

What distinguishes this work is its treatment of between-study heterogeneity. Using Cochran’s Q test and the I-squared statistic, the researchers quantified how much of the variation in each gene’s pooled effect size was attributable to differences among studies rather than random sampling error. Genes were classified as having low, moderate, high, or very high heterogeneity, and those with I-squared values of 75 percent or more were excluded. The sequential workflow evaluated 23,853 genes; 11,917 met the fixed-effect FDR criterion, 11,351 had log2 fold-change estimates in at least three datasets, and removing 5,670 very-high-heterogeneity genes left 5,681 heterogeneity-filtered drought-responsive meta-DEGs. Fixed- and random-effects pooled directions agreed for 5,647 genes, or 99.4 percent, and 3,294 genes remained significant at a random-effects FDR below 0.05. Forty genes met stricter high-confidence-core criteria requiring representation in all four datasets, concordant direction, I-squared below 50 percent, an absolute pooled log2 fold-change of at least 1.5, and random-effects FDR below 0.05.

The analysis showed that a large pooled effect size alone is not a reliable signal of consistent regulation. Some genes with strong positive or negative log2 fold-changes also carried high heterogeneity, indicating that they responded strongly in one water-deficit condition but weakly or differently in another. Even after filtering, principal component analysis based on the retained meta-DEGs still showed dataset-specific structure, meaning that heterogeneity filtering reduced extreme inconsistency but did not eliminate all biological and experimental differences among the public datasets. The authors caution that because developmental stage, induction method, and treatment duration co-varied among datasets, meta-regression was not performed, and I-squared values should be read as overall between-study heterogeneity.

Functional context came from Gene Ontology enrichment analysis of effect-size-significant meta-DEGs represented in at least three datasets, with up- and downregulated genes analyzed separately. Upregulated genes showed nominal enrichment for proteolysis involved in protein catabolic process, vacuolar transport, protein folding, the electron transport chain, and protein retention in the endoplasmic reticulum lumen. Downregulated genes showed nominal enrichment for photosynthesis, response to auxin, cell redox homeostasis, transmembrane transport, and amino acid transmembrane transport. The authors note that these nominally enriched terms provide functional context for the analyzed gene set rather than direct evidence of individual gene function, and that several broadly stress-associated processes should not be interpreted as drought-specific mechanisms.

To test candidates in a living plant, the team grew Micro-Tom seedlings under controlled conditions and imposed drought by withholding irrigation, collecting leaf tissue at 0, 5, 7, and 10 days after irrigation withdrawal. No time-matched well-watered controls were included at the later time points, so the experiment was interpreted as a time course relative to the 0-day baseline rather than a drought-versus-control comparison, and drought severity was not quantitatively measured. Wilting and leaf curling increased over time, and the drought-responsive marker genes SlPP2C3, SlAREB1, and SlNCED1 were induced during progressive soil drying, providing molecular evidence of a response to irrigation withdrawal. Ten representative candidate genes were then examined by RT-qPCR with three biological and three technical replicates, normalized to SlActin and expressed relative to the 0-day baseline.

The RT-qPCR results revealed distinct temporal patterns rather than a uniform response. Among candidates with positive pooled effects, SlMAPKKK55 showed increased expression at 5 days, strong repression at 7 days, and recovery toward baseline at 10 days, while SlMAPKKK51 was significantly induced at 10 days after repression at 7 days. Among the seven candidates with negative pooled effects—including the SAUR-like genes SlSAUR6 and SlSAUR9, the cuticle-associated genes Solyc07g049430 and GDSL2, the glutaredoxin gene Solyc10g008930, the subtilisin-like protease Solyc02g071580, and Solyc11g011730—expression was reduced at one or more time points. Notably, Solyc06g035960, annotated as a 4-coumarate-CoA ligase-like gene, was repressed throughout despite a positive pooled prediction, a discrepancy the authors attribute to its high heterogeneity of 52.4 percent across datasets. Overall, five candidates were directionally consistent with the pooled direction, four were time-dependent or partial, and one was discordant.

The authors emphasize that the identified genes should be regarded as candidates supported across the analyzed water-deficit-related conditions, not as universal drought markers or evidence of a conserved drought-specific mechanism. Transcript-level responsiveness, they stress, does not demonstrate a causal contribution to drought tolerance, and the absence of time-matched controls means irrigation-withdrawal effects cannot be fully separated from time- or developmental-stage-dependent changes. Still, the study demonstrates that RNA-seq meta-analysis combined with formal heterogeneity assessment can prioritize drought-responsive candidates in tomato and delivers a curated set of 5,681 genes—along with a 20-gene candidate pool and ten experimentally profiled representatives—for future functional studies under diverse water-deficit conditions.

Subject of Research: Heterogeneity-guided RNA-seq meta-analysis of drought-responsive genes in tomato

Article Title: Heterogeneity-guided RNA-seq meta-analysis reveals reproducible drought-responsive genes in tomato

Article References: Moon, H., Cho, H., & Shin, C. (2026). Heterogeneity-guided RNA-seq meta-analysis reveals reproducible drought-responsive genes in tomato. Applied Biological Chemistry, 69(1), Article 75. https://doi.org/10.1186/s13765-026-01137-7

Image Credits: AI Generated

DOI: 10.1186/s13765-026-01137-7

Keywords: tomato, drought stress, RNA-seq, meta-analysis, heterogeneity, gene expression, RT-qPCR, Micro-Tom, water deficit, transcriptomics, candidate genes, abiotic stress