Yi Yang, PhD

While deep learning has recently been used for identifying risk variants in genome-wide association studies, its stochastic nature and the complex correlation within genetic data have posed significant challenges for deep-learning-based methods to pinpoint risk variants. We introduce KNOT, a knockoff-augmented neural network for stabilized variable selection with false discovery rate control. KNOT employs contrastive learning and attention mechanisms to model sample relatedness and linkage disequilibrium, alongside a permutation test to detect epistatic interactions. In applications to two autism spectrum disorder family cohorts, KNOT outperforms conventional methods by identifying more known and putative risk loci and epistatic interactions for autism.

Effective risk stratification is essential in clinical practice, enabling better resource allocation and improved patient outcomes. Although machine learning models have been widely used for risk prediction and stratification in electronic health record (EHR) data, conventional interpretability metrics for machine learning models typically lack decision rules for clinicians to determine which patient features significantly influence outcomes. We proposed Knockoff-ML, a model-free machine learning framework that simultaneously accomplishes outcome prediction and identification of risk features through integrating a knockoff framework with various predictive machine learning algorithms. Specifically, Knockoff-ML augments traditional machine learning models with the knockoff framework that enables machine learning models to perform variable selection with false discovery rate (FDR) control in the presence of complex, nonlinear associations between features and outcomes in EHR data. We extensively evaluated Knockoff-ML in both simulations and real-data applications. Our simulation results demonstrated that Knockoff-ML consistently achieved high statistical power to identify risk features while rigorously controlling the FDR, whereas conventional feature selection methods exhibited inflated FDR in most scenarios. In applications to a cohort of 50,591 intensive care unit (ICU) patients from the Medical Information Mart for Intensive Care (MIMIC)-IV database, Knockoff-ML identified risk features significantly associated with short- and long-term mortality. Prediction models with identified risk features in Knockoff-ML also achieved comparable prediction accuracy with full models using all available features. Furthermore, Knockoff-ML exhibited substantially higher predictive power and clinical utility compared to conventional ICU scoring systems such as SOFA and SAPS II. The robust performance and interpretability of Knockoff-ML make it a useful tool for enhancing clinical decision-making, with the potential to significantly improve patient outcomes and optimize healthcare delivery.

Both trio and population designs are popular study designs for identifying risk genetic variants in genome-wide association studies (GWASs). The trio design, as a family-based design, is robust to confounding due to population structure, whereas the population design is often more powerful due to larger sample sizes. Here, we propose KnockoffHybrid, a knockoff-based statistical method for hybrid analysis of both the trio and population designs. KnockoffHybrid provides a unified framework that brings together the advantages of both designs and produces powerful hybrid analysis while controlling the false discovery rate (FDR) in the presence of linkage disequilibrium and population structure. Furthermore, KnockoffHybrid has the flexibility to leverage different types of summary statistics for hybrid analyses, including expression quantitative trait loci (eQTL) and GWAS summary statistics. We demonstrate in simulations that KnockoffHybrid offers power gains over non-hybrid methods for the trio and population designs with the same number of cases while controlling the FDR with complex correlation among variants and population structure among subjects. In hybrid analyses of three trio cohorts for autism spectrum disorders (ASDs) from the Autism Speaks MSSNG, Autism Sequencing Consortium, and Autism Genome Project with GWAS summary statistics from the iPSYCH project and eQTL summary statistics from the MetaBrain project, KnockoffHybrid outperforms conventional methods by replicating several known risk genes for ASDs and identifying additional associations with variants in other genes, including the PRAME family genes involved in axon guidance and which may act as common targets for human speech/language evolution and related disorders.

Family-based designs can eliminate confounding due to population substructure and can distinguish direct from indirect genetic effects, but these designs are underpowered due to limited sample sizes. Here, we propose KnockoffTrio, a statistical method to identify putative causal genetic variants for father-mother-child trio design built upon a recently developed knockoff framework in statistics. KnockoffTrio controls the false discovery rate (FDR) in the presence of arbitrary correlations among tests and is less conservative and thus more powerful than the conventional methods that control the family-wise error rate via Bonferroni correction. Furthermore, KnockoffTrio is not restricted to family-based association tests and can be used in conjunction with more powerful, potentially nonlinear models to improve the power of standard family-based tests. We show, using empirical simulations, that KnockoffTrio can prioritize causal variants over associations due to linkage disequilibrium and can provide protection against confounding due to population stratification. In applications to 14,200 trios from three study cohorts for autism spectrum disorders (ASDs), including AGP, SPARK, and SSC, we show that KnockoffTrio can identify multiple significant associations that are missed by conventional tests applied to the same data. In particular, we replicate known ASD association signals with variants in several genes such as MACROD2, NRXN1, PRKAR1B, CADM2, PCDH9, and DOCK4 and identify additional associations with variants in other genes including ARHGEF10, SLC28A1, ZNF589, and HINT1 at FDR 10%.

Although genome-wide association studies (GWAS) often collect data on multiple correlated traits for complex diseases, conventional gene-based analysis is usually univariate, and therefore, treating traits as uncorrelated. Multivariate analysis of multiple correlated traits can potentially increase the power to detect genes that affect some or all of these traits. In this study, we propose the multivariate hierarchically structured variable selection (HSVS-M) model, a flexible Bayesian model that tests the association of a gene with multiple correlated traits. With only summary statistics, HSVS-M can account for the correlations among genetic variants and among traits simultaneously and can also estimate the various directions and magnitudes of associations between a gene and multiple traits. Simulation studies show that HSVS-M substantially outperforms competing methods in various scenarios, particularly when variants in a gene are associated with a trait in similar directions and magnitudes. We applied HSVS-M to the summary statistics of a meta-analysis GWAS on four lipid traits from the Global Lipids Genetics Consortium and identified 15 genes that have also been confirmed as risk factors in previous studies.

Although genome-wide association studies have been widely used to identify associations between complex diseases and genetic variants, standard single-variant analyses often have limited power when applied to rare variants. To overcome this problem, set-based methods have been developed with the aim of boosting power by borrowing strength from multiple rare variants. We propose the adaptive hierarchically structured variable selection (HSVS-A) before test for association of rare variants in a set with continuous or dichotomous phenotypes and to estimate the effect of individual rare variants simultaneously. HSVS-A has the flexibility to integrate a pairwise weighting scheme, which adaptively induces desirable correlations among variants of similar significance such that we can borrow information from potentially causal and noncausal rare variants to boost power. Simulation studies show that for both continuous and dichotomous phenotypes, HSVS-A is powerful when there are multiple causal rare variants, either in the same or opposite direction of effect, with the presence of a large number of noncausal variants. We also apply HSVS-A to the Wellcome Trust Case Control Consortium Crohn’s disease data for testing the association of Crohn’s disease with rare variants in pathways. HSVS-A identifies two pathways harboring novel protective rare variants for Crohn’s disease.