Explainable Deep Learning of Transcriptomes Prioritizes Candidate Biomarkers With Preferential Performance in Gastric...

Gastric cardia adenocarcinoma is biologically distinct from distal disease, but deployable molecular markers remain scarce. We investigated whether explainable deep learning applied to public transcriptomes could nominate diagnostic and survival-associated candidates. We developed and externally evaluated an explainable transcriptome-based classifier. The Asian Cancer Research Group SuperSeries GSE66229 (300 tumors and 100 patient-matched non-tumor tissues) underwent robust multi-array average preprocessing and quality control. An attention-based deep neural network was evaluated by stratified fivefold cross-validation for tumor-versus-non-tumor classification. Inputs were restricted before training to genes available in all cohorts and ordered identically; no missing model inputs were imputed. Shapley additive explanations nominated 20 genes, and univariable Cox models evaluated overall survival associations. External testing without refitting used GSE29272 and The Cancer Genome Atlas stomach adenocarcinoma cohort. Pathway analyses provided biological context. Cross-validated receiver operating characteristic and precision-recall areas under the curve were both 1.00. External values were 0.85/0.93 for cardia and 0.50/0.67 for non-cardia in GSE29272, and 0.78/0.80 and 0.71/0.50, respectively, in The Cancer Genome Atlas cohort. Six genes showed nominal survival associations in GSE66229. None replicated statistically in the strict external cardia subset; LVRN and WISP2 were nominally concordant in the full stomach adenocarcinoma cohort, but neither survived six-test correction. Enrichment implicated immune and lipid-related processes. Explainable deep learning prioritized candidates with stronger external performance in cardia-versus-non-tumor contrasts. These exploratory diagnostic and survival findings require clinically adjusted, prospective validation before clinical use.