METI-FS: a multi-evidence temporal integration framework for biomarker discovery in small-sample time-series transcriptomics

Selecting a small set of experimentally verifiable candidate biomarkers from small-sample, time-series RNA-seq data ( n < 50, p > 10,000) remains a critical bottleneck for downstream functional studies. Current approaches suffer from three unresolved limitations: differential expression-based pre-filtering lacks temporal awareness, machine learning-based feature selection lacks automatic sparsity control, and individual analytical steps are not integrated into a coherent end-to-end workflow. We developed METI-FS (Multi-Evidence Temporal Integration for Feature Selection), a multi-evidence integration framework whose principal contribution is temporally aware upstream compression: through the intersection of maSigPro temporal regression, WGCNA co-expression network analysis, and DESeq2-based minimum effect size testing, METI-FS systematically reduces the feature space from p > 10,000 to a high-quality candidate pool containing several hundred to several thousand genes, thereby improving the selection stability of a broad range of downstream machine learning methods. A data-driven gap-union threshold automatically determines the final feature set size with minimal user-specified parameters. Across 210 simulation scenarios encompassing six co-expression architectures, METI-FS achieved a relaxed precision (precision_any) of 0.904, indicating that over 90% of the selected genes carried genuine temporal signal (strict precision: 0.408 ± 0.308; F1: 0.296 ± 0.178). Ablation analysis with full paired comparisons identified maSigPro temporal filtering as the most critical upstream component (ΔF1 = −0.206 ± 0.169). On four cross-domain GEO datasets spanning immune activation, anti-tumor drug response, viral infection, and cartilage inflammation, METI-FS achieved compression ratios ranging from 602:1 to 2722:1 without any parameter adjustment. Across all four datasets, the effect size filter (which tests whether each gene's log₂ fold change statistically exceeds a minimum threshold) removed 14–77% of genes that passed temporal and co-expression filters, demonstrating that statistical significance and biological significance require independent assessment. A two-layer benchmarking framework demonstrated that maSigPro-based upstream filtering improved downstream Boruta selection stability by 65% on average across datasets (mean pairwise Jaccard index: 0.288 to 0.475). Cross-validation against published literature across all four datasets confirmed that 75% (41 of 55) of the candidate genes had independent functional evidence, demonstrating that METI-FS, without any biological prior knowledge, independently converges on genes with established functional roles. METI-FS provides a largely automated end-to-end framework for biomarker discovery in small-sample, time-series transcriptomics. Its core contribution is temporally aware upstream compression that systematically enhances the selection stability of a broad range of downstream methods, coupled with data-driven sparsity control that eliminates manual parameter tuning. The R implementation and all analysis scripts are publicly available at GitHub.