Explainable machine learning identifies small high-performance gene signatures from full transcriptomes in human liver preservation cohorts

Transcriptomic studies of liver preservation and ischemia-reperfusion injury (IRI) often report large gene lists that are difficult to translate into deployable diagnostic tools. Genome-wide profiling remains impractical within organ procurement timelines, where decisions must be made rapidly and assay infrastructure is limited. A compact, interpretable gene panel deployable via targeted platforms such as NanoString or RT-qPCR could bridge this gap between transcriptomic discovery and clinical utility. We asked whether full transcriptomes from human liver cohorts can be compressed into such panels without loss of classification performance. Seven GEO datasets covering static cold storage, hypothermic/normothermic machine perfusion, ischemic preconditioning, and peri-transplant biopsies (GSE112713, GSE276531, GSE12720, GSE14951, GSE15480, GSE151648_TIME, GSE151648_IRI) were re-processed into ML-ready expression and phenotype matrices. All gene identifiers were harmonized to HGNC symbols before any modelling step. For each dataset, elastic net (EN) logistic regression and XGBoost were trained and evaluated within a repeated nested cross-validation framework (fivefold × 3-repeat; 15 outer folds), with feature selection performed exclusively on the outer training partition. The union of the top-30 EN coefficient and top-30 XGBoost gain importance genes (EN ∪ XGB) was evaluated on the held-out outer test fold. Nested CV yielded AUCs of 0.842-0.965 across four datasets with reliable sample sizes (n ≥ 30). The two largest cohorts, GSE151648_TIME (n = 80) and GSE12720 (n = 42) achieved AUC = 0.959 ± 0.052 and 0.965 ± 0.059 respectively, with specificity and PPV > 0.90. GSE276531 (n = 32) did not yield reliable transcriptomic signal in either label orientation and is reported as a negative result. Cross-dataset transferability assessed by Leave-One-Dataset-Out validation was consistently poor, with balanced accuracy of 0.500 and macro F1 ≤ 0.333 in five of seven test sets, indicating majority-class collapse rather than genuine cross-dataset signal, suggesting that panels are largely context-specific. An exception was observed among the three reperfusion-focused cohorts (GSE12720, GSE15480, GSE151648_TIME), where high bidirectional panel transfer within this subset (AUC = 0.99-1.00) coincided with convergent selection of inflammatory, proteotoxic stress, and AP-1 axis genes across independently derived panels. Explainable ML models can compress liver transplant transcriptomes into compact, context-specific gene panels with strong within-cohort diagnostic accuracy. The convergence of biologically coherent gene signatures across reperfusion-focused cohorts lends biological support to these panels, while consistently poor LODO transfer underscores the need for cohort-tailored assay development and prospective validation.