A stability framework for variable selection, graphical modelling and clustering: applications in lung cancer research
File(s)
Author(s)
Bodinier, Barbara
Type
Thesis
Abstract
Decades of research have shown that various biological mechanisms are deregulated during carcinogenesis. By offering an agnostic view of individual molecular profiles, high-throughput technologies have enabled new avenues to investigate these molecular changes. These technological developments have also raised novel statistical challenges for the generation of interpretable and robust findings from high dimensional and heterogenous datasets. Regularisation has been instrumental in the analysis of such data by enabling estimations in high dimension and inducing sparsity in the results. Reproducibility issues can be circumvented by combining regularised models with resampling techniques in stability selection or consensus clustering. However, these stability approaches are still under-used, possibly due to the sensitivity of the results to the choice of hyper-parameters and the complexity of their calibration. In the present thesis, I propose to calibrate hyper-parameters of stability selection (for regression or graphical modelling) or consensus clustering by maximising novel scores measuring results stability. To evaluate the performance of this procedure, and compare it with state-of-the-art calibration techniques, I develop flexible simulation models based on the multivariate Normal distribution. Extensive simulation studies suggest that calibration using the novel scores outperforms existing approaches for no increase in computational cost. The proposed models are then used on real omics datasets in a lung cancer case-control study to identify molecular markers of lung carcinogenesis. The use of stability selection in a structural causal modelling framework enables the identification of both (i) potential biological mediators of the effect of tobacco smoking on lung carcinogenesis, and (ii) smoking-independent markers of lung cancer. To facilitate the use of these novel techniques, I have created two R packages, freely available on CRAN. Stability selection and consensus clustering, along with the novel calibration procedures, have been implemented in the R package sharp. Simulation functions are available in the R package fake.
Version
Open Access
Date Issued
2022-10
Date Awarded
2023-04
Copyright Statement
Creative Commons Attribution NonCommercial Licence
License URL
Advisor
Chadeau-Hyam, Marc
Elliott, Paul
Vermeulen, Roel
Sponsor
Medical Research Council (Great Britain)
Publisher Department
School of Public Health
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)