Vision Transformer Model Predicts Gene Mutations and Biomarkers Across 32 Cancer Types

Novel AI model accurately detects key gene mutations and predicts biomarkers across 32 cancer types

According to Medical Xpress, citing a study published in The American Journal of Pathology, researchers from the Menzies Institute for Medical Research at the University of Tasmania have trained a single Vision Transformer that ingests a routine hematoxylin and eosin (H&E)-stained whole-slide histopathology image and emits seven simultaneous outputs — including TP53 mutation status, TP53 RNA expression, tumor taxonomy, and slide-level survival — across 32 solid tumor types. The headline metric is an AUROC of 0.766 for TP53 mutation detection on an independent validation cohort of 1,729 slides, establishing a multi-cancer baseline for weakly supervised computational pathology rather than another single-cancer classifier.

Architecture and training setup

The network operates on whole-slide images of human solid tumors, the standard input format in clinical histopathology. Training was conducted on more than 11,000 primary tumor cases drawn from the Pan-Cancer Atlas, each paired with somatic mutation calls, RNA-sequencing counts, and clinical outcome labels. Because WSIs are extremely large and pixel-level expert annotations are costly and subjective, the authors adopted a weakly supervised learning regime: molecular labels such as TP53 mutation status were available at the patch level, but slides containing TP53-associated morphology were not manually delineated. The design replaces the prevailing one-task-per-network convention with a single forward pass producing seven outputs.

Validation results

On an independent validation set of 1,729 WSIs spanning 32 solid tumor types, the model recorded an AUROC of 0.766 for TP53 mutation detection. The same network further inferred TP53 RNA expression levels and tumor taxonomy directly from slide pixels, consolidating what are normally separate single-task pipelines. Cross-cancer generalization is the operative variable; per-tumor-type AUROC figures and ablation breakdowns were not disclosed in the public reporting.

Position in clinical workflows

The investigators frame the system as complementary to molecular testing rather than a replacement. Co-lead investigator Abadh K. Chaurasia, affiliated with the Menzies Institute and Pandani Solutions Pty Ltd, identifies intended uses in confirmatory triage, prioritization where genomic assays are inaccessible, and decision support for pathologists. Co-lead investigator Alex W. Hewitt frames the tool as a more practical alternative for underprivileged and remote settings where standard molecular profiling for TP53 remains costly.

What to verify

Code, pretrained weights, and the full training recipe were not specified in the cited coverage. Readers planning to reproduce, fine-tune, or benchmark the architecture should confirm the availability of an open repository against the journal version before citing the reported AUROC, and should request per-cancer-type performance breakdowns and ablation results directly from the authors to assess parameter efficiency versus established single-cancer baselines.