Fig. 1: Machine learning workflow for predicting response to atezolizumab in mUC patients.

A Schematic illustrations of developing and validating supervised ML models using gene expression profiles of mUC patient tumours treated with atezolizumab, along with clinical data (Table 1). The IMVigor210 phase II clinical trial cohort (298 patients) served as the discovery dataset for model development, while the Synder et al. cohort (22 patients) was used as a validation dataset for model validation. Both datasets were subsequently pre-processed for ML. Each of the eight regression algorithms was used to develop an ML model with optimal model complexity (OMC) feature selection, which was applied to reduce the model features from the initially considered 20,000 features. The model performance was evaluated by 10-fold cross-validation (CV) with five repetitions, each with a different random seed. The predicted responses were classified into responders or non-responders using a threshold of 2.0. In this way, the model accessed all the information about patients’ response labels (C.R., P.R., S.D. and P.D.), thus capturing the full granularity of data, rather than grouping them (e.g., C.R. and P.R. as responders, and S.D. and P.D. as non-responders) prior to modelling. The best OMC model, the one with the highest MCC, was selected and applied to the validation dataset. MCC, ROC-AUC and PR-AUC were reported. B The performance of the best-performing model was compared against clinical biomarkers (e.g. TMB per megabase, TNB per megabase and PD-L1 expression), as well as state-of-the-art ML models (e.g. EaSIeR and JADBio) for atezolizumab response prediction.