Project title: Biomarker identification for cancer patients’ survival prognosis using multi-omics network information
Supervisors: Zhi Zhao, Manuela Zucknick
Background
Accurate disease risk and treatment outcomes prediction based on multi-omics (e.g., DNA methylation, mRNA expression, protein expression) and clinical data leads to more effective personalized treatment strategies. Although thousands of disease-associated genomic and molecular features have been uncovered, the prognostic prediction of cancer patients remains moderate or poor, which is largely due to the challenges related to identifying the functionally predictive features. Tumour samples are characterized by widely heterogeneous and partly unknown cancer tissues and subtypes, high-dimensional genomic features with gene-gene correlations, and heterogeneity between various molecular data sources, which makes it even more challenging to develop accurate and personalized prediction models for personalized patient management.
Outline of the Work
Initially, the thesis work will exploit existing data sets, for example the Cancer Genome Atlas (TCGA). You may reuse the models that we have developed or extend them for survival prediction. Based on these models, we will incorporate prior information about functional annotations of multi-omics data, i.e., correlations between genomic features within molecular pathways or between multi-omics layers (Figure 1a), to identify multi-omics features and to better estimate their effect sizes accurately for prognostic prediction of cancer patients.
Depending on interest and progress, we may develop novel and timely statistical machine learning methods that integrate multi-omics datasets in a state-of-the-art manner with coherent biological justifications motivated by their hierarchical dependencies (as illustrated in Figure 1a) and efficient computational speed (i.e., scalable Bayesian variable selection as our previous work in Madjar et al., 2021 and Zhao et al., 2021). Ultimately, such improved ML models can lead to improved prognostic models, which in turn can improve personalized cancer therapies.
Figure 1. Pan-cancer survival data and multi-omics signatures. (a) The multi-omics cascade. (b) TCGA pan-cancer patient cohorts. (c) Overall survival times of cancer patients from TCGA. Novel methods are needed to model the high-dimensional multi-omics data and leverage information from heterogeneous cancer cohorts for improved prognostic prediction and biomarker identification.
Learning Outcome
You will become familiar with (i) clinical and molecular data of cancer patients, (ii) Bayesian machine learning for survival analysis, and (iii) interpretation of biological findings based on advanced statistical analysis.
Prerequisites
To start the project, you should be familiar with (i) statistical learning methods, (ii) survival analysis, and (iii) Bayesian statistics. Prior exposure to sparse Bayesian models may be helpful, but is not required.
References
Madjar K, Zucknick M, Ickstadt K, Rahnenführer (2021). Combining heterogeneous subgroups with graph-structured variable selection priors for Cox regression. BMC Bioinformatics, 22, 258.
Zhao Z, Banterle M, Lewin A, Zucknick M (2021). Structured Bayesian variable selection for multiple correlated response variables and high-dimensional predictors. arXiv: 2101.05899.