Machine learning prediction of atezolizumab treatment response in metastatic urothelial carcinoma using gene expression and clinical data
File(s)
Author(s)
Piyawajanusorn, Chayanit
Type
Thesis
Abstract
Cancer heterogeneity leads to variability in treatment responses among patients. Precision medicine addresses this by tailoring treatments based on individual tumour molecular profiles. ML offers a promising approach for predicting drug responses, but is challenged by high-dimensional, low-sample datasets, requiring careful data preparation and analysis. This PhD project demonstrates successful ML application in predicting drug responses across four drug–cancer-type binomials. Since the predictive performance depends on the specific problem, it is important to incorporate diverse molecular profiles tailored to various ML algorithms to enhance the accuracy of drug response predictions. In atezolizumab-treated mUC, CART-OMC trained on the discovery dataset achieved the highest performance, with an MCC of 0.437 in the validation set using just 29 genes. Univariate biomarkers such as TMB, TNB, and PD-L1 were less predictive (MCCs of 0, 0.316, and 0, respectively). When datasets were merged, LGBM-OMC outperformed top modelling approaches like EaSIeR (MCC ~ 0) and JADBio (MCC 0.179), achieving an MCC of 0.252. For taxane-anthracycline-treated breast cancer, LR using mRNAs achieved an MCC of 0.396, with comparable performance on an independent dataset (MCC of 0.351). In doxorubicin-treated breast cancer, CART achieved MCCs of 0.56 and 0.32 using 4 isomiRs and 3 miRNAs, respectively. In gemcitabine-treated pancreatic cancer, RF-OMC with 4 mRNAs achieved an MCC of 0.44, while XGBoost with 12 CpG probes reached an MCC of 0.32. These findings demonstrate the robustness and generality of our proposed methodology, with strong predictive performance across four case studies. The ML models identified predictive, reproducible, and robust drug response signatures, aiding clinicians in identifying patients most likely to benefit from treatment to propose suitable drugs without delay while minimising the risk of side effects for patients unlike to respond. We have made our ML models publicly available for prospective validation on patients with the same drug-cancer types.
Version
Open Access
Date Issued
2024-12-02
Date Awarded
01/05/2025
License URL
Advisor
Ballester, Pedro
Publisher Department
Department of Bioengineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)