Statistical analyses methods for integrating multi-omics data with health outcomes for biomarker discovery and personalized medication

Radiogenomics, an emerging field that integrates radiological imaging, genomics, and clinical data, holds the potential to enhance the accuracy of models for predicting patient outcomes through a multi-modal approach. However, the challenge lies in selecting a manageable number of informative features from the vast array of available features, especially given the complex intrinsic group structures, e.g. biological pathways, and limited availability of datasets that contain both genomic and imaging data. To address these challenges, we propose a joint modeling approach that integrates imaging and genomic data to improve the prediction of clinical outcomes. Specifically, we jointly consider two models, where Model 1 regresses imaging features on genomic features, and Model 2 regresses patient's clinical outcome (either continuous or time-to-event) on genomic features. A sparse group lasso method is used to select informative features while accounting for intrinsic group structures. To enhance the likelihood of selecting shared features, for each penalty term of one model, we introduce a weight based on the model coefficients of the other model to increase the selection chance of features selected by the other model. This weighting mechanism enables the integration of information between the two models to strengthen feature selection. An accelerated generalized coordinate descent algorithm is proposed to obtain model parameter estimates. Our joint model allows the use of two separate datasets to fit the two models, where the dataset for Model 2 does not necessarily contain imaging data. This flexibility enables the use of large-scale genomic datasets, even when corresponding imaging data is unavailable, thereby increasing statistical power. Simulation studies indicate that our method outperforms existing methods in the literature. The application of our method is demonstrated through real data analysis. 

Personalized medicine has become a cornerstone of modern healthcare, as one treatment does not fit all patients. With advancements in technology, omics data and high-dimensional datasets have become increasingly available, providing new opportunities for discovering biomarkers in personalized medicine. Integrating treatment information with multi-omics data can help identify signature molecules that serve as both prognostic and predictive variables. In this talk, we will explore how incorporating various multivariate analysis techniques—such as factor analysis, principal component analysis, and regularized variable selection methods—can effectively identify prognostic biomarkers and treatment effect modifiers. 

Mediation analysis has been instrumental in uncovering the pathways through which exposure variables impact health outcomes via intermediate variables, or mediators, in environmental studies. When dealing with numerous environmental exposures, such as chemical mixtures or pollutants, alongside multiple potential mediators like metabolites, advanced methodologies are required to effectively disentangle direct and indirect effects. In this paper, we propose a novel method designed for this purpose and evaluate its performance against existing approaches through simulation studies and an application to real-world data. 

Multi-omics analysis offers unparalleled insights into the interlinked molecular interactions that govern the underlying biological processes.
In the era of big data, driven by the emergence of high-throughput technologies, we are well-positioned to gain a more comprehensive and detailed understanding of complex systems.
Nevertheless, the challenges lie in developing methods to effectively integrate and analyze this wealth of data.
This challenge is even more apparent when the type of -omics data (e.g., pathomics) lack pixel-to-pixel or region-to-region correspondence across the population.
In this paper, we introduce a novel sample-specific cooperative learning framework designed to adaptively manage diverse multi-omics data types, even when there is no direct correspondence between regions.
We outline this framework for both continuous and categorical outcomes with theoretical guarantees based on finite samples.
We demonstrate the model performance and compare with existing methods in two real world datasets with 1) proteomics and metabolomics; and 2) radiomics and pathomics.