With the increasing amount of high-throughput sequencing data becoming available, the proper integration of differently sized and heterogeneous molecular and clinical groups of variables has become crucial in cancer survival models. Due to the difficulty of multi-omics integration, the Cox Proportional-Hazards (Cox PH) model using clinical data has remained one of the best-performing methods [Herrmann et al., 2021]. This motivates the need for new models which can successfully perform

more »

... s integration in survival models and outperform the Cox PH model. Furthermore, there is a strong need to make multi-omics models more sparse and interpretable to encourage their usage in clinical settings. We developed a neural architecture, termed Supervised Hierarchical Autoencoder (SHAE), based on supervised autoencoders and Sparse-Group-Lasso regularization. Our new method performed competitively with the best performing statistical models used for multi-omics survival analysis. Moreover, it outperformed the Cox PH model using clinical data across all 17 cancers from The Cancer Genome Atlas (TCGA) considered in our work. We further showed that surrogate linear models for SHAE trained on a subset of multi-omics groups achieved competitive performance at consistently high sparsity levels, enabling usage within clinics. Alternatively, surrogate models can act as a feature selection step, permitting improved performance in arbitrary downstream survival models. Code for the reproduction of our results is available on Github.