Figure 2.

Machine learning with heterogenous data. ML algorithms can handle heterogenous data in different ways. For naive feature concatenation-based data integration, tree-based methods (e.g., decision trees and random forest), and penalized linear models (e.g., elastic net and LASSO) can be employed. A two-stage elastic net-based approach like TANDEM is useful if data sources with continuous features (e.g., gene expression) dominate the data sources with binary features (e.g., mutation). Multiple kernel learning (MKL), a robust integrative analysis approach with heterogenous data, employs different kernels or similarity functions for data from different sources and fuses them into a global matrix. Bayesian multitask MKL and simpleMKL are notable examples in this category. Network fusion methods such as similarity network fusion (SNF) employ similarity network for each data type and fuse heterogenous networks. PAthway Recognition Algorithm using Data Integration on Genomic Models (PARADIGM) can incorporate different heterogenous data including gene expression, copy number and curated pathways. Network-based Integration of Multi-omics Data (NetICS) integrates multi-omics data on a directed functional interaction network. Heterogenous information networks like HetroMed can handle raw text, numeric, and categorical data in electronic health records (EHRs) for medical diagnosis. Integrative methods including iCluster+, Scluster and mixed variable restricted Boltzmann machine (MV-RBM) first transform data from heterogenous sources into latent sub-space, and then perform clustering on the latent profiles. Deep learning models such as improved Clustering and PageRank (CPR), Deep Neural Network Synergy model with Autoencoders (AuDNNsynergy), multimodal deep belief networks (DBN) and deep neural networks (DNN) have been employed to perform integrative analysis of heterogenous data by learning complex features through data transformations at multiple layers.