Predicting atezolizumab response in metastatic urothelial carcinoma patients using machine learning on integrated tumour gene expression and clinical data

TL;DR

Researchers developed machine learning models to predict which metastatic urothelial carcinoma patients will respond to atezolizumab treatment based on gene expression and clinical data.

Contribution

The study introduces a machine learning model (CART-OMC) that outperforms existing biomarkers in predicting atezolizumab response using gene expression data.

Findings

The CART-OMC model achieved a validation MCC of 0.437 using only 29 genes, including CXCL9 and IFNG.

Common biomarkers like TMB and PD-L1 had lower predictive accuracy compared to the ML models.

The LGBM-OMC model performed best on merged datasets with an MCC of 0.252, surpassing other approaches like EaSIeR and JADBio.

Abstract

Atezolizumab is a treatment for metastatic urothelial carcinoma (mUC), yet only 23% of mUC patients benefit from it. Worse yet, accurately predicting such responders remains challenging, despite existing biomarkers. Here we employed eight machine learning (ML) algorithms to predict mUC patient response to atezolizumab using tumours’ gene expression profiling and clinical data from two independent cohorts. The CART-OMC model developed on the discovery dataset achieved the highest performance, with a validation set Matthews correlation coefficient (MCC) of 0.437, using the expressions of just 29 ML-selected genes, including CXCL9 and IFNG. Univariate biomarkers like TMB, TNB, and PD-L1 were less predictive with MCCs of 0, 0.316, and 0, respectively. Upon merging these datasets, the best-performing model (LGBM-OMC; MCC of 0.252) also outperformed top modelling approaches such as EaSIeR (MCC ~ 0) and JADBio (MCC of 0.179). We make these promising ML models freely available to predict atezolizumab response in other mUC patients.

Figures