Small-sample RNA-sequencing studies often promise a window into the molecular changes that accompany disease, treatment response, or infection. Yet when fewer than 50 biological samples must be used to interrogate more than 10,000 genes, separating genuine biology from statistical noise becomes exceptionally difficult. A new computational framework called METI-FS, developed by researchers in China and reported in BMC Bioinformatics, is designed to address this problem in time-series transcriptomics, where the order and timing of molecular changes may be as important as the differences between experimental groups. The method combines temporal regression, gene co-expression networks, and statistical effect-size testing before applying machine-learning feature selection. Its aim is not simply to identify genes that change, but to produce a compact and more stable list of candidates that can be tested in the laboratory.
The challenge is particularly acute in studies of infection and immune activation. A viral infection can trigger waves of gene activity rather than a single, static response: innate immune genes may rise rapidly, inflammatory pathways can peak later, and repair or immune-regulatory programs may dominate at subsequent time points. Conventional differential-expression analysis often compares selected time points or averages across a study, potentially overlooking these dynamic patterns. At the same time, machine-learning methods can struggle when the number of variables greatly exceeds the number of samples. With thousands of possible predictors and only a small cohort, models may select different genes each time the data are resampled. Such instability makes it difficult to distinguish robust biomarkers from features that happen to fit one dataset.
METI-FS, or Multi-Evidence Temporal Integration for Feature Selection, tackles the problem through what its developers describe as temporally aware upstream compression. Instead of sending the full transcriptome directly into a feature-selection algorithm, the framework first filters the data using three complementary forms of evidence. The first is maSigPro, a temporal regression approach that identifies genes whose expression trajectories differ across time or experimental conditions. This step is intended to retain genes with meaningful temporal behavior rather than merely large changes at one isolated point. The second is WGCNA, or weighted gene co-expression network analysis, which groups genes into modules based on correlated expression patterns. The third uses DESeq2-based minimum effect-size testing to determine whether a gene’s change is not only statistically significant but also large enough to be biologically meaningful.
This sequence is important because statistical significance and biological relevance are not interchangeable. In large datasets, even very small expression changes can achieve low p-values, while in small datasets, substantial effects may fail to reach conventional significance thresholds. METI-FS treats the minimum effect size as an independent criterion, asking whether a gene’s log2 fold change exceeds a predefined biological threshold with statistical support. According to the study, this filter eliminated between 14% and 77% of genes that had already passed temporal and co-expression screening in four publicly available GEO datasets. The result was a substantial reduction in dimensionality, with reported compression ratios ranging from 602 to one to 2,722 to one, without changing parameters for each dataset.
After this upstream reduction, METI-FS can be paired with downstream machine-learning methods such as Boruta, LASSO, support-vector-machine recursive feature elimination, or random forests. These algorithms approach feature selection differently: some emphasize predictive coefficients, others compare variables against randomized controls, and others rank genes according to their contribution to classification. Yet all face a common problem when thousands of noisy variables compete for attention. By presenting them with a smaller, evidence-supported candidate pool, METI-FS is intended to improve the consistency of their results. The framework also incorporates a data-driven gap-union threshold to determine the final number of selected genes. Rather than requiring users to choose the feature-set size manually, the method looks for a natural separation in the ranking of candidate genes and uses that gap to define the cutoff.
The researchers evaluated the framework across 210 simulated scenarios representing six different co-expression architectures. These simulations were designed to test whether the method could recover genes carrying genuine temporal signals under varied relationships among genes. METI-FS achieved a relaxed precision, called precision_any, of 0.904, meaning that more than 90% of selected genes carried some genuine temporal signal under that definition. Its strict precision was lower, at 0.408 on average with a standard deviation of 0.308, while the mean F1 score was 0.296 with a standard deviation of 0.178. The distinction between these measures is consequential: a method may select many genes with some relevant signal while still including false positives or missing part of the complete true set. The results therefore point to useful enrichment, but not perfect recovery.
An ablation analysis provided additional insight into which elements of the workflow mattered most. In full paired comparisons, removing the maSigPro temporal filter produced the largest decline in performance, with a mean change in F1 of minus 0.206 and a standard deviation of 0.169. This finding supports the central premise that time-aware filtering is more than an optional preprocessing step for longitudinal transcriptomic data. If temporal structure is ignored, downstream algorithms may prioritize genes associated with sample-specific variation, batch effects, or isolated contrasts rather than coordinated trajectories. The study also used a two-layer benchmarking design to distinguish the quality of the upstream candidate pool from the stability of the final machine-learning selection. Across datasets, maSigPro-based filtering improved the stability of Boruta selections by 65% on average, increasing the mean pairwise Jaccard index from 0.288 to 0.475.
The framework was then tested on four cross-domain GEO datasets involving immune activation, anti-tumor drug response, viral infection, and cartilage inflammation. This range was intended to examine whether the method could operate beyond a single disease or experimental system. In the viral-infection setting, the biological value of temporal modeling is especially clear because host responses can change rapidly as infection progresses, viral replication rises, and immune signaling becomes established or resolves. The study reports that, across all four datasets, 41 of 55 candidate genes—75%—had independent functional evidence in the published literature. This cross-validation did not use biological prior knowledge to guide the initial selection, but rather assessed whether the resulting candidates had previously documented roles. Such agreement strengthens the plausibility of the candidates, although literature support cannot substitute for prospective experimental validation.
The authors present METI-FS as a largely automated end-to-end workflow for small-sample, time-series biomarker discovery, rather than as a replacement for laboratory testing or independent clinical validation. Its strongest contribution is the integration of temporal evidence before machine learning, combined with an explicit test for minimum biological effect and an automated approach to sparsity control. The reported results suggest that these steps can make feature selection more reproducible when data are high-dimensional and sample numbers are limited. At the same time, the simulation results show that precision and F1 performance remain imperfect, and the framework depends on choices such as the effect-size threshold, the quality of time-point sampling, and the assumptions built into co-expression analysis. Candidate genes emerging from METI-FS should therefore be treated as prioritized hypotheses. The researchers have made the R implementation and analysis scripts publicly available, allowing other groups to test the approach in infection biology, drug development, and other forms of longitudinal transcriptomics.
Subject of Research: A computational framework for biomarker discovery and feature selection in small-sample, time-series RNA-sequencing studies.
Article Title: METI-FS: a multi-evidence temporal integration framework for biomarker discovery in small-sample time-series transcriptomics
Article References: Zhang, Z., Ma, T., Xu, Y. et al. “METI-FS: a multi-evidence temporal integration framework for biomarker discovery in small-sample time-series transcriptomics.” BMC Bioinformatics (2026).
Image Credits: AI Generated
DOI: 10.1186/s12859-026-06588-9
Keywords: Feature selection, time-series transcriptomics, biomarker discovery, multi-evidence temporal integration, effect-size testing, maSigPro, WGCNA, RNA-seq, viral infection, machine learning
