Nikhil Milind

Transcriptome-wide association studies (TWAS) are widely used to identify genes involved in complex traits and to infer the direction of gene effects on traits. However, despite their popularity, it remains unclear how accurately TWAS recover the true direction of a gene’s effect on a trait. Here we estimate the false sign rate of TWAS for plasma proteins, leveraging the expectation that increased gene expression should generally increase protein expression. We then extend this framework to complex traits, where loss-of-function burden tests provide the expected direction of effect. In both analyses, we observe high discordance with expectations, with TWAS showing a false sign rate of 23% for plasma proteins and 33% for complex traits. We propose that much of the sign discordance arises when expression quantitative trait loci in non-trait-relevant tissues tag haplotypes associated with traits in genome-wide association studies via distinct, tightly linked regulatory variants, yielding spurious TWAS associations with the correct genes but with unreliable direction of effect. These findings show that TWAS-based direction-of-effect estimates should be interpreted with caution and raise concerns about the reliability of TWAS more broadly.

The classic theory of mutation-selection balance predicts the equilibrium frequency of genetic variation under negative selection. The model predicts a simple relationship between the total frequency of deleterious variants, mutation rate, and strength of selection, with different functions for recessive and (co-)dominant genes. In this study, we investigate whether genes associated with human recessive disorders fit the predictions of this classic model. By comparing observed frequencies of loss of function variants (LoFs) to those expected under mutation-selection balance we find that, for nearly all recessive genes, the observed frequencies are too low to be explained by purely recessive selection. Analyzing the effects of heterozygous LoFs on quantitative traits from the UK Biobank, we find that recessive disease genes have widespread quantitative effects in heterozygotes. Together, these results suggest that for a large proportion of recessive disease genes, the majority of selection acting on pathogenic mutations appears to operate through heterozygotes. We conclude that very few human genes follow the classic model of recessive mutation-selection balance.

Transcriptome-wide association studies (TWAS) are widely used to identify genes involved in complex traits and to infer the direction of gene effects on traits. However, despite their popularity, it remains unclear how accurately TWAS recover the true direction of a gene's effect on a trait. Here, we estimate the false sign rate (FSR) of TWAS for plasma proteins, leveraging the expectation that increased gene expression should generally increase protein expression. We then extend this framework to complex traits, where loss-of-function burden tests provide the expected direction-of-effect. In both analyses, we observe high discordance with expectations, with TWAS showing an FSR of 23% for plasma proteins and 33% for complex traits. While colocalization-based filtering reduced the FSR, substantial discordance remained, and with substantial loss of recall. However, when we restricted gene-direction assignments for plasma proteins to using only relevant tissues in combination with colocalization-based filtering, the FSR dropped to 11%, and to just 5% if we excluded brain-specific proteins. We propose that much of the sign discordance arises when eQTLs in non-trait-relevant tissues tag GWAS-associated haplotypes via distinct, tightly-linked regulatory variants, yielding spurious TWAS associations with the correct genes but with unreliable direction-of-effect. These findings show that TWAS-based direction-of-effect estimates should be interpreted with caution and raise concerns about the reliability of TWAS more broadly.

The classic theory of mutation-selection balance predicts the equilibrium frequency of genetic variation under negative selection. The model predicts a simple relationship between the total frequency of deleterious variants, mutation rate, and strength of selection, with different functions for recessive and (co-)dominant genes. In this study, we investigate whether genes associated with human recessive disorders fit the predictions of this classic model. By comparing observed frequencies of loss of function variants (LoFs) to those expected under mutation-selection balance we find that, for nearly all recessive genes, the observed frequencies are too low to be explained by purely recessive selection. Analyzing the effects of heterozygous LoFs on quantitative traits from the UK Biobank, we find that recessive disease genes have widespread quantitative effects in heterozygotes. Together, these results suggest that most selection experienced by pathogenic mutations in recessive disease genes may be due to stabilizing selection in heterozygotes. We conclude that very few human genes follow the classic model of recessive mutation-selection balance.

Standard genome-wide association studies (GWAS) and rare variant burden tests are essential tools for identifying trait-relevant genes. Although these methods are conceptually similar, we show by analyzing association studies of 209 quantitative traits in the UK Biobank that they systematically prioritize different genes. This raises the question of how genes should ideally be prioritized. We propose two prioritization criteria: 1) trait importance — how much a gene quantitatively affects a trait; and 2) trait specificity — a gene’s importance for the trait under study relative to its importance across all traits. We find that GWAS prioritize genes near trait-specific variants , while burden tests prioritize trait-specific genes . Because non-coding variants can be context specific, GWAS can prioritize highly pleiotropic genes, while burden tests generally cannot. Both study designs are also affected by distinct trait-irrelevant factors, complicating their interpretation. Our results illustrate that burden tests and GWAS reveal different aspects of trait biology and suggest ways to improve their interpretation and usage.

The genome-wide burdens of deletions, loss-of-function mutations, and duplications correlate with many traits. Curiously, for most of these traits, variants that decrease expression have the same genome-wide average direction of effect as variants that increase expression. This seemingly contradicts the intuition that for individual genes reducing expression should have the opposite effect on a phenotype as increasing expression. To understand this paradox, we use the gene dosage response curve (GDRC), which relates changes in gene expression to expected changes in phenotype. We show that, for many traits, GDRCs are systematically biased in one trait direction relative to the other, and we develop a simple theoretical model that explains this bias in trait direction. Our results have broad implications for complex traits, drug discovery, and statistical genetics.

Genetics play an important role in the development of metabolic diseases. However, the relative influence of genetic variation on metabolism is not well defined, particularly in tissues, where metabolic dysfunction that leads to disease occurs. We used inbred strains of laboratory mice to evaluate the impact of genetic variation on the metabolomes of tissues that play central roles in metabolic diseases. We chose a set of four common inbred strains that have different levels of susceptibility to obesity, insulin resistance, and other common metabolic disorders. At the ages used, and under standard husbandry conditions, these lines are not overtly diseased. Using global metabolomics profiling, we evaluated water-soluble metabolites in liver, skeletal muscle, and adipose from A/J, C57BL/6J, FVB/NJ, and NOD/ShiLtJ mice fed a standard mouse chow diet. We included both males and females to assess the relative influence of strain, sex, and strain-by-sex interactions on metabolomes. The mice were also phenotyped for systems level traits related to metabolism and energy expenditure. Strain explained more variation in the metabolite profile than did sex or its interaction with strain across each of the tissues, especially in liver. Purine and pyrimidine metabolism and pathways related to amino acid metabolism were identified as pathways that discriminated strains across all three tissues. Based on the results from ANOVA, sex and sex-by-strain interaction had modest influence on metabolomes relative to strain, suggesting that the tissue metabolome remains largely stable across sexes consuming the same diet. Our data indicate that genetic variation exerts a fundamental influence on tissue metabolism.

Late-Onset Alzheimer’s disease (LOAD) is a common, complex genetic disorder well-known for its heterogeneous pathology. The genetic heterogeneity underlying common, complex diseases poses a major challenge for targeted therapies and the identification of novel disease-associated variants. Case-control approaches are often limited to examining a specific outcome in a group of heterogenous patients with different clinical characteristics. Here, we developed a novel approach to define relevant transcriptomic endophenotypes and stratify decedents based on molecular profiles in three independent human LOAD cohorts. By integrating post-mortem brain gene co-expression data from 2114 human samples with LOAD, we developed a novel quantitative, composite phenotype that can better account for the heterogeneity in genetic architecture underlying the disease. We used iterative weighted gene co-expression network analysis (WGCNA) to reduce data dimensionality and to isolate gene sets that are highly co-expressed within disease subtypes and represent specific molecular pathways. We then performed single variant association testing using whole genome-sequencing data for the novel composite phenotype in order to identify genetic loci that contribute to disease heterogeneity. Distinct LOAD subtypes were identified for all three study cohorts (two in ROSMAP, three in Mayo Clinic, and two in Mount Sinai Brain Bank). Single variant association analysis identified a genome-wide significant variant in TMEM106B (p-value < 5×10⁻⁸, rs1990620^G) in the ROSMAP cohort that confers protection from the inflammatory LOAD subtype. Taken together, our novel approach can be used to stratify LOAD into distinct molecular subtypes based on affected disease pathways. One of the challenges of identifying risk variants for late-onset Alzheimer’s disease based on transcript abundance from post-mortem brain tissue is the inherent heterogeneity of human cohorts. This is evident from studies that have employed increasingly larger cohorts without detecting any new risk variants. In this study, we have developed a novel methodology to decompose transcriptomic data into submodules to reduce experimental noise and increase power by accounting for patient heterogeneity. We demonstrate the effectiveness of our submodules by identifying associated variants that are either novel or have been implicated previously in late-onset Alzheimer’s disease. A systems-level understanding of the transcriptome is especially important for translating human disease-associated variants into models that can accelerate the development of therapies.

Late-Onset Alzheimer’s disease (LOAD) is a common, complex genetic disorder well-known for its heterogeneous pathology. The genetic heterogeneity underlying common complex diseases poses a major challenge for targeted therapies and the identification of novel disease-associated variants. Case-control approaches are often limited to examining a specific outcome in a group of heterogenous patients with different clinical characteristics. Here, we developed a novel approach to define relevant transcriptomic endophenotypes and stratify decedents based on molecular profiles in three independent human LOAD cohorts. By integrating post-mortem brain gene co-expression data from 2114 human samples with LOAD, we developed a novel quantitative, composite phenotype that can better account for the heterogeneity in genetic architecture underlying the disease. We used iterative weighted gene co-expression network analysis (WGCNA) analysis to reduce data dimensionality and to isolate gene sets that are highly co-expressed within disease subtypes and represent specific molecular pathways. We then performed single variant association testing using whole genome-sequencing data for the novel composite phenotype in order to identify genetic loci that contribute to disease heterogeneity. Distinct LOAD subtypes were identified for all three study cohorts (two in ROSMAP, three in Mayo Clinic, two in Mount Sinai Brain Bank). Single variant association analysis identified a genome-wide significant variant in TMEM106B (p-value < 5×10 ⁻⁸ , rs1990620 ^G ) in the ROSMAP cohort that confers protection from the inflammatory LOAD subtype. Taken together, our novel approach can be used to stratify LOAD into distinct molecular subtypes based on affected disease pathways.