An integrative multi-project transcriptomic and structural prediction framework identifies candidate cold-responsive...

Cold stress limits alfalfa (Medicago sativa) growth and persistence, but public transcriptomic datasets differ widely in genotype, tissue, treatment duration, and experimental design. We integrated RNA-seq data from ten independent BioProjects using a common processing workflow while retaining project-specific structures. A recurrent contrast-level DEG-derived pool of 4,354 genes was ranked by random forest using expression profiles from 240 samples. The original model showed strong internal discrimination (OOB ROC-AUC = 0.937), whereas fully nested leave-one-BioProject-out validation yielded an accuracy of 0.729, balanced accuracy of 0.676, and ROC-AUC of 0.727. PlantTFDB annotation identified MsG0680033896.01, MsG0680033848.01, and MsG0480021906.01 as the three highest-ranked transcription factors. The first two candidates showed greater stability in project-held-out and alternative machine-learning analyses. In project-aware multilevel meta-analysis, neither the primary 50-contrast analysis nor the 54-contrast sensitivity analysis identified genome-wide significant transcripts after Benjamini-Hochberg correction. However, MsG0680033896.01 and MsG0680033848.01 showed predominantly positive effects, positive pooled estimates, and confidence intervals excluding zero in both analyses, whereas MsG0480021906.01 showed weaker directional consistency. Co-expression, promoter prediction, chromosomal localization, and AlphaFold3 modeling provided additional computational context, including localization of the two leading candidates within a Chr6 CBF/DREB1-like-enriched region. These results prioritize MsG0680033896.01 and MsG0680033848.01 as high-confidence computational candidates and retain MsG0480021906.01 as an additional project-sensitive candidate for future functional testing.