SLIDE: Significant Latent Factor Interaction Discovery and Exploration across biological domains

Introduction:: Modern multi-omic technologies can generate deep multi-scale profiles. However, differences in data modalities, multicollinearity of the data, and large numbers of irrelevant features make the analyses and integration of high-dimensional omic datasets challenging. For example, multicollinearity can increase the variance of regression coefficients and lead to deflation of corresponding P values. This is a significant barrier to meaningful inference in a regression setting for high-dimensional multi-collinear data. Further, human biological systems are complex, multi-factorial and organized hierarchically, with complex interaction rules at each hierarchy. A linear model is often inadequate at capturing relevant higher-order relationships in such a system. Finally, while recent methods developed by us and others have harnessed these high-dimensional multi-scale multi-modal datasets to predict different outcomes/groups of interest, they do not provide meaningful inference beyond prediction. In fact, approaches that do provide insights into the underlying mechanistic bases of outcome are tailored primarily for low-dimensional datasets, and often trade predictive power for inference.

Materials and Methods:: To address these, we present SLIDE, a novel data-distribution-free approach to analyze high-dimensional multi-omic datasets and uncover latent factors that drive the outcome of interest. SLIDE makes no assumptions regarding the distribution of the underlying data as it builds on a unique latent-factor regression framework developed by us. It takes into account an extremely large search space of relationships to converge on a very small subset of biologically relevant and actionable latent factors. Critically, SLIDE incorporates both linear and non-linear relationships, including complex hierarchical structures. It uncovers significant interacting latent factors in diverse contexts that span scales of organization from cellular/molecular phenotypes (e.g., extent of clonal expansion of CD4 T cells) to organismal phenotypes (e.g., disease severity of patients with diffuse systemic sclerosis). The discovery of these relationships is also coupled to rigorous FDR control, something that is extremely difficult to do in a large search space of non-linear/hierarchical relationships. Here, this is made possible using our unique analytical framework that creatively adapts ultra-modern methods for FDR control14. SLIDE comes with provable statistical guarantees regarding identifiability of the latent factors, corresponding inference of significant interacting latent factors. This is fundamentally different from several methods that have been published recently that rely on clever heuristics but do not have formal statistical guarantees or work only when strong biological priors are available. SLIDE has rigorous statistical guarantees, recapitulates known biological mechanisms and helps uncover novel biological mechanisms.

Results, Conclusions, and Discussions:: We tested the predictive performance of SLIDE on a range of datasets, and it outperformed several state-of-the-art approaches. Further, it provided novel inference not afforded by any existing approaches, thus being one of the only methods that simultaneously provides meaningful inference for high-dimensional data without compromising on predictive power. When analyzing datasets from SSc patients to elucidate the basis of SSc pathogenesis, SLIDE recovered altered transcriptomic states in myeloid cells and fibroblasts, a well elucidated basis of SSc disease severity. But it also identified an unexplored keratinocyte-centric signature (validated by protein staining), and a novel mechanism involving an interaction between the altered transcriptomic states in myeloid cells and fibroblasts with HLA signaling in macrophages. This mechanism has strong support in recent genetic association analyses. SLIDE also worked well on spatial transcriptomic data and was able to accurately identify significant interacting latent factors underlying immune cell partitioning by 3D location within lymph nodes. In the characterization of latent factors underlying clonal expansion of CD4 T cells, SLIDE recapitulated well known inhibitory receptors and markers of activation/exhaustion including Lag3, Pdcd1 (Pd1), Tigit and Cd200 that standard differential expression (DE) analyses would have picked up. However, it also identified several novel markers that standard DE analyses would have missed. This includes novel gene sets and functional modules that have both significant linear and non-linear relationships to clonal expansion. Overall, SLIDE is a first-in-class approach for the discovery of significant interacting latent factors from high-dimensional multi-modal datasets that help infer novel biological mechanisms. It is an engine for biological discovery from modern multi-omic datasets.


Acknowledgements (Optional): :

References (Optional): :