# Revealing intra-group immunotherapy response heterogeneity in metastatic urothelial carcinoma through interpretable feature extraction and spectral clustering

## TL;DR

This study identifies distinct patient subgroups in metastatic urothelial carcinoma with different responses to immunotherapy, using interpretable machine learning and clustering.

## Contribution

A novel framework combining SHAP-based feature extraction and spectral clustering to reveal immune response heterogeneity in cancer patients.

## Key findings

- Four patient clusters with distinct immune phenotypes and ICI response patterns were identified.
- Cluster 3 showed high response rates with inflamed tumor features, while Cluster 1 had no response and immune-desert features.
- The framework was validated in an independent cohort, showing consistent immune subclass patterns.

## Abstract

Immune checkpoint inhibitors (ICIs) have improved outcomes in metastatic urothelial carcinoma (mUC) but clinical responses remain highly heterogenous. Traditional binary classification of response overlooks clinically relevant variability within each group but a more detailed understanding of intra-group heterogeneity may support subclass-specific therapeutic strategies.

We developed a novel analysis framework that integrates interpretable feature extraction and spectral clustering to identify patient subclasses associated with heterogeneous responses to ICIs. This method was applied to tumor transcriptomic data from the IMvigor210 cohort (n = 298), comprising mUC patients treated with atezolizumab. Interpretable features based on SHapley Additive exPlanations (SHAP) were computed from a response classification model to quantify patient-level gene contributions, which were then used for spectral clustering. An independent cohort (GSE176307, n = 88) was used for external validation.

This approach identified four patient clusters with distinct immune phenotypes and response patterns. Cluster 3 (92.3% responders) showed an inflamed phenotype with high PD-L1 expression, T cell activation, and TP53 mutations. Cluster 1 (100% non-responders) displayed an immune-desert phenotype with FGFR3 mutations and elevated TGF-β signaling. Cluster 2 was more heterogeneous, containing two subgroups (Sub 1 and Sub 2) with differing immune activity and immunosuppressive gene expression, corresponding to response rates of 23.2% and 77.3%, respectively. Similar patterns were observed in the validation cohort.

Our framework, which combines SHAP-based interpretable feature extraction with spectral clustering, revealed subclass-level heterogeneity in ICI response, highlighting biologically distinct immune subclasses. This approach may facilitate the development of subclass-specific therapeutic strategies.