TL;DR
This study leverages deep learning to predict genetic mutations directly from standard H&E stained histopathology slides across 23 cancer types. The findings demonstrate that individual gene mutations leave stronger morphological signatures than entire pathways, suggesting a more granular approach to predicting genetic alterations from histology.
What Problem Does This Paper Address?
Histopathology is the cornerstone of cancer diagnosis, but it primarily provides morphological information. Meanwhile, genetic mutations drive cancer development and progression, and many are clinically actionable. While recent advances in deep learning have shown promise in linking histological features to specific genetic alterations, most prior work focused on a limited number of genes or tumor types. This paper addresses the gap of systematically evaluating the predictability of a broader range of genetic alterations, including both individual genes and functional pathways, across a diverse set of solid tumors. Understanding how genetic alterations manifest morphologically is crucial for developing AI tools that can assist pathologists and potentially guide treatment decisions.
Methods
The researchers employed a deep learning approach using convolutional neural networks (CNNs). They trained these networks in an end-to-end manner to predict the mutational status of specific genes and signaling pathways directly from H&E stained whole slide images. The core methodology involved:
- Image Preprocessing: Manually annotating tumor regions in whole slide images and tessellating them into smaller tiles.
- Gene and Pathway Selection: Identifying 69 clinically relevant genes based on mutation prevalence and then grouping them into 12 functional pathways.
- Model Training: Training CNNs (specifically a modified ShuffleNet architecture) using a 3-fold cross-validation approach at the patient level to predict binary targets (mutated vs. wild-type for genes, or pathway-altered vs. not-altered for pathways).
- Performance Evaluation: Assessing model performance using Area Under the Receiver Operating Curve (AUROC), F-Score, and Matthew's Correlation Coefficient (MCC).
The study also included specific experiments to investigate the impact of "dominant" genes on pathway prediction and to explore the correlation between prediction scores and allele frequency.
Datasets
The study utilized data from The Cancer Genome Atlas (TCGA) project. This included:
- Histopathology Images: Anonymized, digitized H&E stained whole slide images from 7,829 patients across 23 different solid tumor types.
- Mutation Data: Molecular data (mutation status) for the selected genes, obtained from publicly available sources like cBioPortal, MSK-IMPACT, and OncoKB.
The study did not introduce a new benchmark dataset but rather leveraged existing, large-scale public resources.
Key Insights and Ideas
- Individual Genes Shine Brighter: The most striking insight is that individual gene mutations are generally more predictable from histopathology than entire pathway alterations. This suggests that specific genes have more distinct morphological "signatures" that CNNs can readily detect.
- Pathways are Complex: The predictability of pathway alterations appears to be largely driven by the mutations in one or two key "dominant" genes within that pathway. When these dominant genes were excluded, pathway prediction performance significantly dropped.
- Morphological Footprint of Genetics: The study reinforces the idea that genetic changes are not just molecular events but have tangible effects on tissue morphology, which can be captured by AI.
- Broad Applicability: The systematic approach across 23 tumor types demonstrates the wide-ranging potential of deep learning in inferring genetic alterations from routine pathology.
- Adenocarcinomas are More Predictable: Genetic alterations in adenocarcinomas were found to be more predictable than in other tumor types, potentially due to their more frequent and distinct morphological changes associated with genetic alterations.
Key Results
- Gene Prediction: Mutations in 44 out of 69 analyzed genes were predictable in at least one tumor type. Genes like TP53, PTEN, SETD2, and KRAS showed high predictability across multiple cohorts. For example, TP53 mutations were predictable in 11 out of 23 cohorts with an average AUROC of 0.68.
- Pathway Prediction: Alterations in 12 out of 14 clinically relevant pathways were detectable. However, pathway prediction AUROCs were generally lower than for individual genes.
- High Performance Genes: The neural network achieved AUROCs higher than 0.75 for twenty genes. Notably, FGFR3 mutations in bladder cancer were predicted with an AUROC of 0.78, IDH1 in low-grade glioma with 0.764, and NOTCH2 in colorectal cancer with 0.934.
- Tumor Type Differences: Adenocarcinomas (e.g., STAD, UCEC, CRC) showed the highest number of predictable genes and pathways. Squamous cell carcinomas generally showed lower predictability.
Limitations and Future Work
- TCGA Reliance: The study relied solely on TCGA data. Validation on independent, external cohorts is crucial to confirm these findings.
- Dataset Size Imbalance: The unequal dataset sizes across tumor types might have led to underestimation of predictability in smaller cohorts. Future studies should re-analyze genes in larger cohorts.
- Potential Biases: TCGA data may carry biases due to patient selection and the computational pathology process.
- Limited Mutation Types: The study focused on single nucleotide variants and small deletions/insertions, excluding fusion genes, copy number changes, or expression data.
- Intratumor Heterogeneity: The mismatch between the tissue slide region and the DNA source for genetic characterization could dilute results due to intratumor heterogeneity. Future work should quantify this impact.
- Future Directions: The authors suggest exploring prediction of other genetic alteration types, incorporating clinical variant classification data, and evaluating promising candidates in multicentric image collections. They also highlight the potential for AI to identify patient subgroups that could benefit from targeted therapies.
Summary updated