Loading…
Histopathology model predicting TP53 mutation status, TP53 RNA expression, and tumour taxonomy from H&E whole-slide images across 32 solid cancers.
No providers recorded yet. Browse all providers
A pathologist reading an H&E slide can name the tumour but not its genotype. Establishing whether TP53 carries a pathogenic driver mutation means sending tissue out for sequencing, and prognosis means a third analysis again — three assays, three turnaround times, and in low-resource settings often only the first is reachable. This model collapses those readouts into a single pass over the slide already on the microscope: from one whole-slide image it emits tumour type among 32 solid cancers, TP53 mutation status, TP53 RNA expression level, and four survival-related outputs.
The mechanism is weakly supervised multiple-instance learning. A gigapixel slide is too large for a transformer to read whole, and region-level annotation of TP53-associated morphology does not exist at pan-cancer scale, so each slide is tiled into patches, every patch is embedded by a frozen self-supervised DINOv2 vision transformer into 768 dimensions, and that token sequence is read by a learned six-layer transformer aggregator with a prepended CLS token. Only the slide's recorded label supervises training; the aggregator works out which patches matter. Because all seven heads read the same CLS vector, one representation must be informative about morphology, genotype and outcome at once — and those vectors cluster by tumour type under t-SNE, with TP53 status stratifying within each cluster.
It was built at the University of Tasmania's Menzies Institute for Medical Research with the pathology department at Royal Hobart Hospital and Pandani Solutions, posted to medRxiv in November 2025 and published in The American Journal of Pathology. It sits between single-biomarker slide models such as HRDPath and general slide-level encoders such as Threads, which fit a fresh probe for each new task; here one checkpoint answers all seven questions as trained.
11,060 primary tumours were assembled from the TCGA Pan-Cancer Atlas with matched somatic mutations, batch-corrected RNA-seq and clinical endpoints; TP53 variants were graded as pathogenic driver, passenger, benign or wild-type using the COSMIC Mutant Census and TP53 hotspot annotations. 7,784 whole-slide images were pulled from the GDC portal, and the first training stage additionally used TCGA-UT, 1.6 million tumour patches from 8,736 slides across 7,175 patients. Patches are 518×518 pixels at 6× downsampling, filtered for tissue fraction, blur and ink artefacts, and each slide is presented as a bag of 300 tokens. The aggregator runs 12 attention heads per layer, and training combines focal cross-entropy, Cox partial likelihood and Smooth L1 losses across the heads under AdamW on a single 80 GB GPU.
On a held-out partition of 1,729 TCGA slides, argmax tumour-type accuracy is 0.659, and per-class Youden thresholds put one-versus-rest AUROC above 0.88 for every tumour type except ovarian. TP53 mutation detection reaches AUROC 0.766 (95% CI 0.743–0.789) at 70.0% sensitivity and 72.9% specificity, and TP53 expression regression gives a mean squared error of 1.213. Risk stratification separates Kaplan–Meier curves (hazard ratio 1.73 for overall survival, 1.56 for progression-free interval), but concordance falls by roughly 48% and 41% against the internal split, and the authors conclude that prognostic utility remains limited.
The realistic use is triage, not replacement: flagging which slides likely carry a pathogenic TP53 alteration so confirmatory sequencing is spent where it will pay, and supplying a tumour-type and molecular read where genomic assays are slow, costly or unavailable. Because the input is the H&E slide already scanned during routine diagnosis, no extra staining or tissue is consumed — which suits remote and resource-limited pathology services.
The claim that matters is multi-task pan-cancer modelling at the slide level: most computational pathology models answer one question for one cancer, while one supervised aggregator over a frozen general-purpose encoder holds taxonomy and a molecular biomarker together. Its limits are equally on the record — heavy class imbalance across tumour types, downsampled inputs that discard fine nuclear detail, and prognostic heads that do not survive the held-out partition, which is itself drawn from TCGA rather than an outside institution. Nothing has been released: no code, no checkpoint, no hosted inference and no stated model licence, with four of the five authors declaring a founding interest in a company commercialising computational pathology. The architecture and the training recipe are public; the model is not.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.