New AI Model Predicts TP53 Mutations Across 32 Cancers; EGFR Models Show Ancestry Variability
New AI model predicts TP53 mutations across 32 cancers. EGFR prediction models vary by ancestry. Breast pathology review finds AI improves accuracy but notes barriers.
Researchers have developed a novel AI model that can analyze a routine whole histopathology image and simultaneously predict cancer subtype, specific genetic mutations, and survival outcomes across 32 different solid cancers. A separate study evaluating two open-source AI models for EGFR prediction found that model performance varies with patient ancestry, with accuracy dropping in Asian patients.
The AI-based Vision Transformer model was trained on a dataset that included more than 11,000 primary tumor cases retrieved from the Pan-Cancer Atlas, with corresponding somatic mutation, RNA-sequencing, and clinical outcome data. It achieved a strong predictive accuracy score (AUROC of 0.766) for TP53 mutation detection across 32 solid tumor types in an independent validation set of 1,729 slides. The model also demonstrated the ability to infer TP53 RNA expression levels and tumor taxonomy directly from whole slide images. Because whole slide images are extremely large and complex, the researchers deployed a weakly supervised learning strategy, enabling the model to learn from slide-level labels without requiring exhaustive pixel- or region-level annotations. The findings from the study in The American Journal of Pathology, published by Elsevier, highlight the potential of computational pathology to connect routine diagnostic imaging with molecular oncology.
The co-lead investigator from the Menzies Institute for Medical Research and School of Medicine, University of Tasmania, explained: “Standard molecular profiling for TP53 mutations is often costly and inaccessible in underprivileged or remote clinical settings. We developed a single model that can generate seven outputs simultaneously from the whole histopathology image, including TP53 mutation status, TP53 RNA expression, tumor type, and survival-related outcomes at the slide level.” Another co-lead investigator noted that the approach could help identify patients who may benefit from confirmatory molecular testing, support triage in settings with limited genomic testing, and provide additional decision support to clinicians, while emphasizing that it should be complementary to molecular testing, not a replacement.
In a study of patients with lung adenocarcinoma (LUAD), researchers validated the reproducibility of two open-source AI models, EAGLE and DeepGEM, for EGFR prediction from hematoxylin-eosin (H&E) whole-slide images. The reproduced AUCs on the TCGA LUAD dataset (n = 463) matched published results: 0.89 (95% CI, 0.85-0.93) for EAGLE and 0.88 (95% CI, 0.82-0.93) for DeepGEM. Overall, 2,098 patients with LUAD were included. In the Dana-Farber Cancer Institute (DFCI) cohort (n = 1,759), EAGLE showed an AUC of 0.83 (95% CI, 0.81-0.85) vs 0.68 (95% CI, 0.65-0.70) for DeepGEM. In the EU-based TNM-I cohort (n = 339), EAGLE again outperformed DeepGEM with an AUC of 0.81 (95% CI, 0.74-0.88) vs 0.75 (95% CI, 0.68-0.83).
EGFR mutations were most frequent among Asian patients (60 [63%]), followed by African patients (17 [31%]) in the DFCI cohort. EAGLE maintained robust performance (AUC >0.80) across most ancestry groups, but its AUC dropped to 0.68 (95% CI, 0.55-0.78) in Asian patients. The study also noted that in a recent global survey across 90 countries, more than 40% of clinicians reported treating patients before biomarker results were available, with cost, turnaround time, sample quality and access identified as leading barriers.
A separate review of AI in breast pathology found that artificial intelligence is reshaping diagnostic breast pathology by improving diagnostic accuracy, efficiency, and reproducibility. The review documents clinical applications including detection of lymph node metastases, Nottingham grading, benign versus malignant classification, automated biomarker quantification, prognostic prediction, risk stratification, and analysis of the tumor microenvironment. Common barriers to real-world implementation include data quality and bias, regulatory considerations, cost and infrastructure, and the challenge of integrating tools into existing laboratory workflows.