Hybrid quantum-classical attention yields selective gains in small-cohort gene prediction from histopathology

The study tested the QDSM attention layer across 29 TCGA cohorts plus an independent CPTAC pancreatic-cancer set. Leave-one-cancer-out mixed-effects modeling showed that baseline molecular features explained part of the gene-level benefit while residuals flagged cancer-specific programs that improved or declined beyond expectation. In adrenocortical carcinoma, the genes that gained most were enriched for adverse-survival associations, linking the method to prognostically relevant biology rather than uniform transcriptome-wide uplift.

QDSM did not deliver blanket gains; it redistributed predictive accuracy across pathways, improving selected metabolic and lineage genes in pancreatic transfer experiments while worsening others. Hardware experiments on IBM processors recovered the underlying doubly stochastic matrix primitive, confirming the quantum component was executable but not yet demonstrating end-to-end clinical superiority.

The approach targets precisely the settings where sequencing fails—limited tissue, absent molecular data, or cohorts too small for robust training—yet the preprint provides no prospective validation or head-to-head comparison against existing image-based predictors such as those trained on larger pan-cancer sets. External replication in independent, multi-institutional cohorts with locked preprocessing pipelines is required before any claim of clinical triage utility can be sustained.