Biomedical data are becoming increasingly multimodal and thereby capture the underlying complex relationships among biological processes. Deep learning (DL)-based data fusion strategies are a popular approach for modeling these nonlinear relationships. Therefore, we review the current state-of-the-art of such methods and propose a detailed taxonomy that facilitates more informed choices of fusion strategies for biomedical applications, as well as research on novel methods. By doing so, we find that deep fusion strategies often outperform unimodal and shallow approaches. Additionally, the proposed subcategories of fusion strategies show different advantages and drawbacks. The review of current methods has shown that, especially for intermediate fusion strategies, joint representation learning is the preferred approach as it effectively models the complex interactions of different levels of biological organization. Finally, we note that gradual fusion, based on prior biological knowledge or on search strategies, is a promising future research path. Similarly, utilizing transfer learning might overcome sample size limitations of multimodal data sets. As these data sets become increasingly available, multimodal DL approaches present the opportunity to train holistic models that can learn the complex regulatory dynamics behind health and disease.

Making breast cancer prognosis from gene expression profiles of the primary tumor has become a promising application of deep learning. Yet, to be relevant to real world applications in the clinic and for knowledge discovery, these models must be robust to common distribution shifts. In this study, we evaluate recently proposed methods for improving domain and subgroup shifts. We test the in-distribution and out-of-distribution generalization of multiple episode learning, stochastic weight averaging, group distributionally robust optimization, and a subsampling scheme on one training and four external breast cancer prognosis datasets. The evaluation found that the methods can, to various degrees, improve generalization across domains, although there remain, partially high, generalization gaps. Additionally, in-distribution and out-of-distribution generalization differs between clinical subtypes of breast cancer. Thus, we conclude that further research into methods specifically addressing challenges in breast cancer prognosis from gene expression data are warranted. 

In recent years, serum-based tests for early detection and detection of tissue of origin are being developed. Circulating microRNA has been shown to be a potential source of diagnostic information that can be collected non-invasively. In this study, we investigate circulating microRNAs as predictors of cancer stage. Specifically, we predict whether a sample stems from a patient with early stage (0-II) or late stage cancer (III-IV). We trained five machine learning algorithms on a data set of cancers from twelve different primary sites. The results showed that cancer stage can be predicted from circulating microRNA with a sensitivity of 71.73%, specificity of 79.97%, as well as positive and negative predictive value of 54.81% and 89.29%, respectively. Furthermore, we compared the best pan-cancer model with models specialized on individual cancers and found no statistically significant difference. Finally, in the best performing pan-cancer model 185 microRNAs were significant. Comparing the five most relevant circulating microRNAs in the best performing model with the current literature showed some known associations to various cancers. In conclusion, the study showed the potential of circulating microRNA and machine learning algorithms to predict cancer stage and thus suggests that further research into its potential as a non-invasive clinical test is warranted.