25.2 Transformation-based integration
In transformation-based integration, omic datasets are first transformed into an intermediate representation, typically a graph or a kernel matrix, and they are then merged before building the final model. This approach preserves the specific properties of each omic layer if they are transformed into appropriate intermediate representations, and a wide range of omic data can be combined as long as they share a unique identifier (i.e. a sample ID). Graph-based analyses have the advantage of easier interpretability and lower computational requirements whereas, overall, kernel-based methods provide higher predictive performance [70].
There are several methods available for transformation-based unsupervised analysis. Regularised Multiple Kernel Learning for Locality Preserving Projections (rMKL-LPP) [71] and PAMOGK [72] are examples of kernel- and graph-based methods that can be used for clustering. Meta-analytic SVM (Meta-SVM) [73] and NEighborhood based Multi-Omics clustering (NEMO) [74] are other methods available for transformation-based unsupervised analysis. Most of the methods for transformation-based supervised analysis are kernel- or graph-based algorithms [70]. The kernel-based integration approaches include Semi-Definite Programming SVM (SDP-SVM) [75], Multiple Kernel Learning with Feature Selection (FSMKL) [76], Relevance Vector Machine (RVM) [77] and Ada-boost RVM [78]. The graph-based integration approaches include graph-based semi-supervised learning (included in supervised analyses following Reel et al. 2021 [58]) [79], graph sharpening [80] and composite network [81]. Graph-based analyses have the advantage of easier interpretability and lower computational requirements whereas, overall, kernel-based methods provide higher predictive performance [70]. However, see Multi-Omics Graph Convolutional Networks (MOGONET) [82] for a high performing graph-based classification method.