Impact of joint Dimension Reduction methods for survival prediction - extension of a multi-omics benchmark study
Résumé
The complexity of certain diseases requires multiple measurements on the human genome to fully understand the underlying dysregulations, which led to the generation of multi-omics datasets. The high-dimensionality and heterogeneity inherent to multi-omics datasets appear to be challenging to analyse. Nevertheless, in the context of survival analysis and exploiting 18 distinct cancer datasets from TCGA (Herrmann et al., 2021) showed that models considering the inherent group structure of multi-omics datasets can help leverage their full potential. Yet, from their conclusions, it remains unclear whether the joint analysis of molecular and clinical data increases predictive power in comparison to the analysis of clinical data only. Extending this benchmark, we aim to tackle this limitation by exploring methods that extract links between omics data blocks by employing 4 joint Dimension Reduction (jDR) techniques: RGCCA, JIVE, IntNMF and MOFA. For each method, several combinations of hyperparameters are tested. Our approach initially estimates a reduced space from molecular data alone via unsupervised or supervised techniques, followed by survival prediction using a Cox model on this joint reduced space, with and without clinical data. We demonstrate that jDR methods, when combined with a Cox model, can significantly outperform traditional ones. When jointly analyzing clinical and omics data, some of the jDR methods perform significantly better than the baseline model (Cox model on clinical data only). Furthermore, when prediction is done only from omics data, most of the compared methods perform significantly lower than this baseline, with a notable exception for methods that can be supervised.
Domaines
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |