Unsupervised classification of multi-omics data during cardiac remodeling using deep learning

Abstract

Integration of multi-omics in cardiovascular diseases (CVDs) presents high potentials for translational discoveries. By analyzing abundance levels of heterogeneous molecules over time, we may uncover biological interactions and networks that were previously unidentifiable. However, to effectively perform integrative analysis of temporal multi-omics, computational methods must account for the heterogeneity and complexity in the data. To this end, we performed unsupervised classification of proteins and metabolites in mice during cardiac remodeling using two innovative deep learning (DL) approaches. First, long short-term memory (LSTM)-based variational autoencoder (LSTM-VAE) was trained on time-series numeric data. The low-dimensional embeddings extracted from LSTM-VAE were then used for clustering. Second, deep convolutional embedded clustering (DCEC) was applied on images of temporal trends. Instead of a two-step procedure, DCEC performes a joint optimization for image reconstruction and cluster assignment. Additionally, we performed K-means clustering, partitioning around medoids (PAM), and hierarchical clustering. Pathway enrichment analysis using the Reactome knowledgebase demonstrated that DL methods yielded higher numbers of significant biological pathways than conventional clustering algorithms. In particular, DCEC resulted in the highest number of enriched pathways, suggesting the strength of its unified framework based on visual similarities. Overall, unsupervised DL is shown to be a promising analytical approach for integrative analysis of temporal multi-omics.

Keywords: Cardiovascular; Clustering; Integrative analysis; Multi-omics; Time-series; Unsupervised deep learning.

Figure 1.. Integrative Clustering Workflow for Temporal Multi-Omics Data.

Missing values in proteomics and metabolomics datasets are completed by fully conditional specification (FCS)-based multiple imputations. Cubic splines were employed to smooth the temporal trends. Images were also generated based on the temporal trends over 14 days. As a baseline, K-means clustering, hierarchical clustering (HC), partitioning around medoids (PAM)) were employed. In contrast, two deep neural network architectures for unsupervised learning were also implemented. First, long short-term memory (LSTM)-based variational autoencoder (LSTM-VAE) was used to extract low-dimensional embeddings from the time-series numeric data and K-means clustering was employed on the embeddings. Second, deep convolutional embedded clustering (DCEC) was employed to perform clustering of molecules based on image data. The results obtained from the clustering methods were fed to Reactome knowledgebase for pathway enrichments analyses.

Figure 2:. 3D Visualization of 6 clusters in the Embedded Feature Space of LSTM-VAE.

Three axes represent the three dimensions of the latent space obtained from LSTM-VAE. Different colors denote six clusters identified by K-means clustering. The ellipsoids represent 80% of the concentration for each cluster.

Figure 3:. t-SNE Visualization of DCEC Embeddings.

The t-SNE visualization of the embeddings, optimized for both reconstruction and cluster losses, clearly shows the separation of samples into six clusters. Each cluster obtained by DCEC is represented by a different color.

Figure 4.. Cluster Centers for Different Clustering Methods.

Cluster centers for the each of the 6 clusters obtained by K-means, HC, PAM, LSTM-VAE and DCECE methods are plotted. Based on the temporal similarities of cluster centers across methods, we labeled the clusters as increase, decrease, increase-decrease, decrease-increase, late increase and late decrease. However, there are some obvious differences, which can be attributed to the differences in learning algorithms and input data types (e.g., time-series numeric or image).

Figure 5.. Shared and Unique Molecules Across Different Clustering Methods in Merged Clusters.

In each merged cluster, majority of the molecules are shared by at least three clustering methods. All the methods have some unique molecules in each cluster due to differences in learning algorithm and/or input data type. Hierarchal clustering generally has a high concentration of unique molecules while LSTM-VAE has the least.