Weighted dimensionality reduction and robust Gaussian mixture model based cancer patient subtyping from gene expression data.

BACKGROUND

The heterogeneous nature of cancer necessitates subtyping of cancer patients into distinct and well separated subgroups. However, computational issues arise because gene expression data is noisy and contains outliers apart from being high dimensional. As such, an attempt to subtype cancer patients from gene expression data leads to highly overlapping Kaplan-Meier (KM) survival plots and thus clear distinction among the discovered subtypes becomes difficult. Here we attempt to achieve a greater separation among the subtypes through a robust clustering pipeline.

METHODS

We propose a robust framework to achieve a better separation among the discovered subtypes. Our framework is based on dimensionality reduction of a weighted gene expression matrix using t-distributed Stochastic Neighbor Embedding (t-SNE) and a robust Gaussian mixture model based clustering approach. Every gene is weighted according to the median absolute deviation (MAD) of the gene before dimensionality reduction. The results are quantified by measuring the minimum pairwise separation among the KM plots and minimum hazard ratio among the subtypes. We also introduce a novel method, called cumulative survival separation, to quantify the separation among the discovered subtypes.

RESULTS

To validate the proposed methodology we obtained five cancer gene expression datasets from The Cancer Genome Atlas (TCGA) and comparisons with Consensus Clustering (CC), Consensus non-negative matrix factorization (CNMF), fast density-aware spectral clustering (Spectrum) and Neighborhood based Multi-Omics clustering (NEMO) methodologies show that the proposed method is able to achieve a greater separation compared to the aforementioned methods in literature. For instance, the minimum pairwise life expectancy difference (in days) between the discovered subtypes for GBM is 61 days for the proposed methodology with MAD scores, whereas it is approximately 33, 19, 49 and 33 days only for CC, Spectrum, Nemo and CNMF respectively. Comparisons are also shown for the proposed framework with and without using the MAD scores and it is observed that MAD score significantly improves the subtype separation. Hazard ratio analysis also shows that the proposed methodology performs better. Furthermore, pathway over-representation analyses were carried to identify relevant genetic pathways which can be possible targets for treatment.

CONCLUSION

The results suggest that the use of median absolute deviation and a robust clustering methodology are helpful in achieving greater separation among the subtypes with better statistical and clinical significance.