Tumor subtyping moves beyond broad categories: HHRM classifies eight hierarchy levels and reaches 0.7924 accuracy on pan-cancer test data

Tumor subtype prediction moves from emitting a flat label at once to deciding level by level from primary cancer type down to molecular subtype, with each level's output feeding the next level's decision. Earlier flat classifiers treated subtype prediction as an ordinary multi-label problem and lost performance as the label space grew deeper and wider; existing hierarchical methods either modeled each cancer type separately or used only two levels. On a pan-cancer cohort of 11,476 samples, 26 primary cancers, 405 subtypes and up to eight levels, HHRM exceeded the compared hierarchical models on all four metrics in local, global and mixed output modes; in local mode its accuracy was 0.7924, against 0.7896 for the closest model, HMCN_R.

Each level's module first embeds gene expression through cancer hallmark gene networks, then a gating unit filters the previous level's predictions into the current level. The hallmark networks come from ten cancer hallmark gene sets built from GO terms, totaling 6,448 genes, plus 75,869 molecular interactions from Reactome. The model has eight recursive modules, each with a hallmark graph convolution, a highway gate and a two-layer fully connected output unit; training used 10,589 TRUE samples expanded to 11,518 with synthetic samples, and 887 FALSE samples served as validation.

Removing either the hallmark graph convolution or the highway gate lowers performance, indicating both components contribute. Replacements included self-attention, a graph attention network, a convolutional network, and a three-layer MLP in place of the highway gate. Against the variant using only the immediately preceding level's output, the replacement variants lost 0.0029 to 0.078 in accuracy and 0.0179 to 0.0532 in macro-F1; the graph attention variant had the highest accuracy at 0.7892, and the convolutional variant was lowest on all four metrics.

The demonstrated scope is a Treehouse-derived pan-cancer cohort with 26 primary cancers, 405 subtypes and up to eight levels, plus 1,608 breast cancer samples from METABRIC with four subtypes and two levels; the latter reached 0.9105 test accuracy and 0.8659 validation accuracy only after retraining.

On 887 difficult validation samples whose labels were entirely discordant with the original diagnosis, HHRM reached 0.3450 accuracy against full eight-level labels and 0.3788 against original levels, ranging from 0.0 to 1.0 across cancer types; applied to METABRIC without retraining, accuracy was 0.3495, which the authors attribute to microarray versus RNA sequencing platform differences and imperfect label correspondence.