A new study published in npj Digital Medicine highlights the potential of an interpretable deep learning framework to improve the prediction of responses to neoadjuvant chemotherapy (NAC) in patients with muscle-invasive bladder cancer (MIBC). The Graph-based Multimodal Late Fusion (GMLF) model, integrates histopathological images and gene expression data to predict treatment outcomes and identify key biomarkers.
MIBC is a high-risk disease where the standard treatment combines NAC with radical cystectomy. However, only about 35% of patients achieve a complete pathological response (pCR), and the current "one-size-fits-all" approach poses risks of toxicity and delayed surgery. Tumor heterogeneity remains a significant challenge in predicting treatment responses, and robust predictive models are lacking.
To address this gap, researchers developed the GMLF model, which leverages both histopathological whole-slide images (WSI) and RNA sequencing data. The goal was to enhance the accuracy of predicting NAC responses and identify relevant biomarkers to optimize treatment strategies.
The GMLF model integrates H&E-stained histopathological images and RNA sequencing data. Graph neural networks were used to extract spatial and cellular information from the images, while multilayer perceptrons analyzed gene expression data. The model employs a late fusion strategy to combine predictions from both modalities.
Data for the study were derived from the SWOG S1314 (COXEN) clinical trial, which included 180 patients with MIBC. The dataset comprised 182 gigapixels WSI and corresponding gene expression data, with 56 patients (30.8%) achieving pCR and 126 patients (69.2%) showing partial or no response. The data were split into a discovery set (80% of patients) and a hold-out test set (20% of patients), with performance evaluated using a 5-fold cross-validation and an 80/20 train-test split.
The GMLF model demonstrated superior performance compared to single-modal models, achieving an average AUC of 0.74 in 5-fold cross-validation and 0.72 in the 80/20 test set split.
Through SHAP analysis, the study identified genes associated with NAC response, including TP63, CCL5, and DCN. Gene set enrichment analysis revealed that basal differentiation and myofibroblast pathways were critical predictors of NAC response.
Histopathological analysis showed that in pCR-associated regions, there was an increase in cancer and stromal cell counts, a decrease in necrotic cells, and a significant rise in the tumor-stroma ratio.
The study innovatively combined histopathological images and gene expression data, capturing complementary information that single-modal data cannot provide. The use of a GNN-based Slide Graph+ architecture allowed for the analysis of spatial and cellular relationships in histopathological images.
However, limitations included limited sample size, lack of external validation, and model complexity. Future research will focus on validating the model with larger external datasets and optimizing feature selection to enhance clinical applicability.
Link to Original Study: Predicting response to neoadjuvant chemotherapy in muscle-invasive bladder cancer via interpretable multimodal deep learning
