Integrative algorithms for multi-modal biological data for classification and interpretation

Abstract

Accurate diagnostic classification and biological interpretation are both very important in biology and medicine, which are data-rich sciences. The integration of different data types is challenging for high predictive accuracy of clinical phenotypes and discovery of complex interactions. In this dissertation, we propose an unsupervised model, a supervised model, and a loss function algorithm to integrate multi-modal biological data. In the first part of this dissertation, we propose a joint deep semi-non-negative matrix factorization (JDSNMF) model, which uses a hierarchical non-linear feature extraction approach that can capture shared latent features from the complex multi-omics data. The extracted latent features obtained from JDSNMF enabled a variety of downstream tasks, including disease prediction and module analysis. The proposed model is also applicable not only to sample-matched data but also to feature-matched data, so it can be flexibly applied in various cases. We demonstrate the capabilities of JDSNMF using simulated data and multi-omics datasets from Alzheimer’s disease (AD) cohorts, and evaluate the feature extraction performance in the context of classification. We identify AD- and age-related modules from the latent matrices using an explainable artificial intelligence and regression model. In the second part of this dissertation, we propose a novel multi-task attention learning algorithm for multi-omics data termed MOMA, which captures important biological processes for high diagnostic performance and interpretability. MOMA vectorizes features and modules using a novel geometric approach and focuses on important modules in multi-omics data via an attention mechanism. Experiments on public data on AD and cancer datasets with various classification tasks demonstrate the superior performance of this approach. The utility of MOMA is also verified using a comparison experiment with an attention mechanism that is turned on or off and biological analysis. Multi-modal learning often outperforms its unimodal counterparts by using unimodal contributions and cross-modal interactions. However, focusing solely on integrating multi-modally learned instances into a unified comprehensive representation often overlooks the unimodal characteristics. In real data, the contributions of modalities can vary from instance to instance, and they often reinforce or conflict with each other. In the third part of this dissertation, we introduce a novel MultiModal loss for multi-modal learning by subgrouping instances according to their unimodal contributions. This loss is empirically shown to perform better on one synthetic and four real datasets, and we validate that it accelerates convergence. Furthermore, we show that it generates a reliable prediction score for each modality, which is essential for subgrouping.