TL;DR
This paper introduces HistCode, a novel computational histopathology framework that uses self-supervised contrastive learning on whole slide images (WSIs) to predict differential gene expression in cancer cells. By learning robust slide-level features, HistCode not only achieves state-of-the-art performance in tumor diagnosis but also accurately predicts gene expression changes, offering a powerful new tool for understanding cancer biology from pathology images.
What Problem Does This Paper Address?
Digital pathology has revolutionized cancer diagnosis, enabling automated analysis of histopathology images. While deep learning excels at extracting features for tasks like tumor classification and prognosis prediction, a significant gap remains in leveraging these visual features to understand the underlying molecular mechanisms, specifically differential gene expression in tumor cells. Traditional methods for gene expression analysis, like bulk RNA-seq, lose crucial spatial information and require separate molecular profiling. This paper aims to bridge this gap by developing a computational approach that can infer gene expression patterns directly from standard H&E stained pathology slides, offering a more integrated and potentially cost-effective way to study cancer at both the morphological and molecular levels.
Methods
The core of HistCode lies in its self-supervised contrastive learning framework. Given the high dimensionality of WSIs, the images are first tessellated into smaller tiles. These tiles are then processed by a convolutional neural network (CNN) encoder (ResNet50) to extract tile-level features. The key innovation is the use of adversarial contrastive learning (AdCo). This method trains the encoder to learn discriminative representations by maximizing the similarity between augmented views of the same tile (positive samples) while minimizing similarity with a large set of negative samples. AdCo specifically generates challenging negative samples to push the model to capture more informative features.
Following feature extraction, a gated-attention pooling mechanism is employed to aggregate tile-level features into a single slide-level representation. This attention mechanism allows the model to weigh the importance of different tiles, effectively focusing on the most informative regions within a WSI.
These learned slide-level features are then transferred to downstream tasks:
- Tumor Diagnosis: A simple fully connected layer and softmax classifier are used to predict tumor presence.
- Differential Gene Expression Prediction: For this task, the model leverages attention weights to select tiles most likely representing tumor and normal tissue. The features from these selected tiles are averaged and concatenated, then fed into a linear regression model for each gene to predict its differential expression level (log10 fold change).
Datasets
The study utilized several large-scale datasets:
- TCGA-BRCA: 1,979 WSIs from 1,094 breast cancer patients (1,580 tumor, 399 normal).
- CPTAC-BRCA: 642 WSIs from 134 breast cancer patients, used as an independent test set to evaluate model generalizability.
- TCGA-LUSC and TCGA-LUAD: 2,168 WSIs from 1,010 lung cancer patients (1,577 tumor, 591 normal).
- RNA-seq datasets from TCGA: Used for differential expression analysis, with 153 matched breast cancer samples.
The authors also mention using a pathologist to manually annotate a few slides for validation of spatial localization.
Key Insights and Ideas
- Bridging the Morphology-Molecular Gap: The paper demonstrates that robust molecular information, like differential gene expression, can be inferred directly from H&E stained pathology images, a significant step towards integrated multi-modal cancer analysis.
- Adversarial Contrastive Learning for Histopathology: AdCo proves to be a powerful self-supervised learning technique for extracting meaningful tile-level features from unannotated WSIs, outperforming other methods.
- Attention is Key for Aggregation: Gated-attention pooling significantly improves performance in both tumor diagnosis and differential gene expression prediction compared to simpler aggregation methods like max or mean pooling.
- Fold Change Matters: The accuracy of differential gene expression prediction is positively correlated with the magnitude of the gene's fold change, suggesting that more dramatic molecular changes are more readily reflected in observable phenotypic features.
- Spatial Interpretability: The attention scores not only help in feature aggregation but also provide a spatial map of informative regions within the WSI, aligning well with pathologist annotations of tumor and necrosis areas. This interpretability extends to visualizing the spatial distribution of immune cells based on marker gene expression.
Key Results
- Tumor Diagnosis: HistCode achieved state-of-the-art performance, with an accuracy of 0.963 and ROC-AUC of 0.976 on the TCGA-LUNG cohort, outperforming competitive methods by at least 4%. Similar high performance was observed on the TCGA-BRCA cohort (0.960 accuracy, 0.965 ROC-AUC) and validated on the independent CPTAC-BRCA cohort (0.931 accuracy).
- Differential Gene Expression Prediction:
- For the TCGA-BRCA cohort, the average Pearson correlation coefficient for predicting differential expression of 200 cancer driver genes was 0.185 (p<0.01).
- A significant portion of genes (88%) showed statistically significant prediction results (p<0.05) compared to a random baseline.
- In a binary classification task (log2(fc) > 1.5), 33% of genes achieved >0.9 accuracy, and 60% achieved >0.8 accuracy.
- The paper found a positive correlation between gene fold-change level and prediction accuracy.
- Frequently mutated and immune-related genes were predicted with higher accuracy, with immune cell marker genes achieving an average Pearson correlation of 0.624.
Limitations and Future Work
- Limitations:
- The study focused on a limited number of cancer driver genes (top 200 most mutated).
- While the model shows strong performance, the Pearson correlation coefficients for differential gene expression prediction are moderate (average 0.185), indicating room for improvement.
- The interpretation of spatial localization of immune cells is based on marker genes, which is an indirect inference.
- Future Work:
- The authors suggest that computational pathology can promote the spatial visualization of the tumor immune microenvironment.
- The framework could be extended to infer other molecular patterns beyond differential gene expression.
- Further exploration of the relationship between phenotypic features and molecular variations is warranted.
- The authors believe their work will yield inspiring insights into computational pathology and greatly extend its clinical application.
Summary updated