Integrating transcriptomics, metabolomics, and GWAS helps reveal molecular mechanisms for metabolite levels and disease risk.

Transcriptomics data have been integrated with genome-wide association studies (GWASs) to help understand disease/trait molecular mechanisms. The utility of metabolomics, integrated with transcriptomics and disease GWASs, to understand molecular mechanisms for metabolite levels or diseases has not been thoroughly evaluated. We performed probabilistic transcriptome-wide association and locus-level colocalization analyses to integrate transcriptomics results for 49 tissues in 706 individuals from the GTEx project, metabolomics results for 1,391 plasma metabolites in 6,136 Finnish men from the METSIM study, and GWAS results for 2,861 disease traits in 260,405 Finnish individuals from the FinnGen study. We found that genetic variants that regulate metabolite levels were more likely to influence gene expression and disease risk compared to the ones that do not. Integrating transcriptomics with metabolomics results prioritized 397 genes for 521 metabolites, including 496 previously identified gene-metabolite pairs with strong functional connections and suggested 33.3% of such gene-metabolite pairs shared the same causal variants with genetic associations of gene expression. Integrating transcriptomics and metabolomics individually with FinnGen GWAS results identified 1,597 genes for 790 disease traits. Integrating transcriptomics and metabolomics jointly with FinnGen GWAS results helped pinpoint metabolic pathways from genes to diseases. We identified putative causal effects of UGT1A1/UGT1A4 expression on gallbladder disorders through regulating plasma (E,E)-bilirubin levels, of SLC22A5 expression on nasal polyps and plasma carnitine levels through distinct pathways, and of LIPC expression on age-related macular degeneration through glycerophospholipid metabolic pathways. Our study highlights the power of integrating multiple sets of molecular traits and GWAS results to deepen understanding of disease pathophysiology.

Author Correction: Exome sequencing of Finnish isolates enhances rare-variant association power.

An Amendment to this paper has been published and can be accessed via a link at the top of the paper.

Exome sequencing of Finnish isolates enhances rare-variant association power.

Exome-sequencing studies have generally been underpowered to identify deleterious alleles with a large effect on complex traits as such alleles are mostly rare. Because the population of northern and eastern Finland has expanded considerably and in isolation following a series of bottlenecks, individuals of these populations have numerous deleterious alleles at a relatively high frequency. Here, using exome sequencing of nearly 20,000 individuals from these regions, we investigate the role of rare coding variants in clinically relevant quantitative cardiometabolic traits. Exome-wide association studies for 64 quantitative traits identified 26 newly associated deleterious alleles. Of these 26 alleles, 19 are either unique to or more than 20 times more frequent in Finnish individuals than in other Europeans and show geographical clustering comparable to Mendelian disease mutations that are characteristic of the Finnish population. We estimate that sequencing studies of populations without this unique history would require hundreds of thousands to millions of participants to achieve comparable association power.

Association of structural variation with cardiometabolic traits in Finns.

The contribution of genome structural variation (SV) to quantitative traits associated with cardiometabolic diseases remains largely unknown. Here, we present the results of a study examining genetic association between SVs and cardiometabolic traits in the Finnish population. We used sensitive methods to identify and genotype 129,166 high-confidence SVs from deep whole-genome sequencing (WGS) data of 4,848 individuals. We tested the 64,572 common and low-frequency SVs for association with 116 quantitative traits and tested candidate associations using exome sequencing and array genotype data from an additional 15,205 individuals. We discovered 31 genome-wide significant associations at 15 loci, including 2 loci at which SVs have strong phenotypic effects: (1) a deletion of the ALB promoter that is greatly enriched in the Finnish population and causes decreased serum albumin level in carriers (p = 1.47 × 10-54) and is also associated with increased levels of total cholesterol (p = 1.22 × 10-28) and 14 additional cholesterol-related traits, and (2) a multi-allelic copy number variant (CNV) at PDPR that is strongly associated with pyruvate (p = 4.81 × 10-21) and alanine (p = 6.14 × 10-12) levels and resides within a structurally complex genomic region that has accumulated many rearrangements over evolutionary time. We also confirmed six previously reported associations, including five led by stronger signals in single nucleotide variants (SNVs) and one linking recurrent HP gene deletion and cholesterol levels (p = 6.24 × 10-10), which was also found to be strongly associated with increased glycoprotein level (p = 3.53 × 10-35). Our study confirms that integrating SVs in trait-mapping studies will expand our knowledge of genetic factors underlying disease risk.

Mitochondrial genome copy number measured by DNA sequencing in human blood is strongly associated with metabolic traits via cell-type composition differences.

BACKGROUND: Mitochondrial genome copy number (MT-CN) varies among humans and across tissues and is highly heritable, but its causes and consequences are not well understood. When measured by bulk DNA sequencing in blood, MT-CN may reflect a combination of the number of mitochondria per cell and cell-type composition. Here, we studied MT-CN variation in blood-derived DNA from 19184 Finnish individuals using a combination of genome (N = 4163) and exome sequencing (N = 19034) data as well as imputed genotypes (N = 17718). RESULTS: We identified two loci significantly associated with MT-CN variation: a common variant at the MYB-HBS1L locus (P = 1.6 × 10-8), which has previously been associated with numerous hematological parameters; and a burden of rare variants in the TMBIM1 gene (P = 3.0 × 10-8), which has been reported to protect against non-alcoholic fatty liver disease. We also found that MT-CN is strongly associated with insulin levels (P = 2.0 × 10-21) and other metabolic syndrome (metS)-related traits. Using a Mendelian randomization framework, we show evidence that MT-CN measured in blood is causally related to insulin levels. We then applied an MT-CN polygenic risk score (PRS) derived from Finnish data to the UK Biobank, where the association between the PRS and metS traits was replicated. Adjusting for cell counts largely eliminated these signals, suggesting that MT-CN affects metS via cell-type composition. CONCLUSION: These results suggest that measurements of MT-CN in blood-derived DNA partially reflect differences in cell-type composition and that these differences are causally linked to insulin and related traits.

Genome-wide mapping of brain phenotypes in extended pedigrees with strong genetic loading for bipolar disorder.

Bipolar disorder is a highly heritable illness, associated with alterations of brain structure. As such, identification of genes influencing inter-individual differences in brain morphology may help elucidate the underlying pathophysiology of bipolar disorder (BP). To identify quantitative trait loci (QTL) that contribute to phenotypic variance of brain structure, structural neuroimages were acquired from family members (n = 527) of extended pedigrees heavily loaded for bipolar disorder ascertained from genetically isolated populations in Latin America. Genome-wide linkage and association analysis were conducted on the subset of heritable brain traits that showed significant evidence of association with bipolar disorder (n = 24) to map QTL influencing regional measures of brain volume and cortical thickness. Two chromosomal regions showed significant evidence of linkage; a QTL on chromosome 1p influencing corpus callosum volume and a region on chromosome 7p linked to cortical volume. Association analysis within the two QTLs identified three SNPs correlated with the brain measures.

Understanding the Hidden Complexity of Latin American Population Isolates.

Most population isolates examined to date were founded from a single ancestral population. Consequently, there is limited knowledge about the demographic history of admixed population isolates. Here we investigate genomic diversity of recently admixed population isolates from Costa Rica and Colombia and compare their diversity to a benchmark population isolate, the Finnish. These Latin American isolates originated during the 16th century from admixture between a few hundred European males and Amerindian females, with a limited contribution from African founders. We examine whole-genome sequence data from 449 individuals, ascertained as families to build mutigenerational pedigrees, with a mean sequencing depth of coverage of approximately 36×. We find that Latin American isolates have increased genetic diversity relative to the Finnish. However, there is an increase in the amount of identity by descent (IBD) segments in the Latin American isolates relative to the Finnish. The increase in IBD segments is likely a consequence of a very recent and severe population bottleneck during the founding of the admixed population isolates. Furthermore, the proportion of the genome that falls within a long run of homozygosity (ROH) in Costa Rican and Colombian individuals is significantly greater than that in the Finnish, suggesting more recent consanguinity in the Latin American isolates relative to that seen in the Finnish. Lastly, we find that recent consanguinity increased the number of deleterious variants found in the homozygous state, which is relevant if deleterious variants are recessive. Our study suggests that there is no single genetic signature of a population isolate.

Genetic analysis of activity, brain and behavioral associations in extended families with heavy genetic loading for bipolar disorder.

BACKGROUND: Disturbed sleep and activity are prominent features of bipolar disorder type I (BP-I). However, the relationship of sleep and activity characteristics to brain structure and behavior in euthymic BP-I patients and their non-BP-I relatives is unknown. Additionally, underlying genetic relationships between these traits have not been investigated. METHODS: Relationships between sleep and activity phenotypes, assessed using actigraphy, with structural neuroimaging (brain) and cognitive and temperament (behavior) phenotypes were investigated in 558 euthymic individuals from multi-generational pedigrees including at least one member with BP-I. Genetic correlations between actigraphy-brain and actigraphy-behavior associations were assessed, and bivariate linkage analysis was conducted for trait pairs with evidence of shared genetic influences. RESULTS: More physical activity and longer awake time were significantly associated with increased brain volumes and cortical thickness, better performance on neurocognitive measures of long-term memory and executive function, and less extreme scores on measures of temperament (impulsivity, cyclothymia). These associations did not differ between BP-I patients and their non-BP-I relatives. For nine activity-brain or activity-behavior pairs there was evidence for shared genetic influence (genetic correlations); of these pairs, a suggestive bivariate quantitative trait locus on chromosome 7 for wake duration and verbal working memory was identified. CONCLUSIONS: Our findings indicate that increased physical activity and more adequate sleep are associated with increased brain size, better cognitive function and more stable temperament in BP-I patients and their non-BP-I relatives. Additionally, we found evidence for pleiotropy of several actigraphy-behavior and actigraphy-brain phenotypes, suggesting a shared genetic basis for these traits.

Characterization of Expression Quantitative Trait Loci in Pedigrees from Colombia and Costa Rica Ascertained for Bipolar Disorder.

The observation that variants regulating gene expression (expression quantitative trait loci, eQTL) are at a high frequency among SNPs associated with complex traits has made the genome-wide characterization of gene expression an important tool in genetic mapping studies of such traits. As part of a study to identify genetic loci contributing to bipolar disorder and other quantitative traits in members of 26 pedigrees from Costa Rica and Colombia, we measured gene expression in lymphoblastoid cell lines derived from 786 pedigree members. The study design enabled us to comprehensively reconstruct the genetic regulatory network in these families, provide estimates of heritability, identify eQTL, evaluate missing heritability for the eQTL, and quantify the number of different alleles contributing to any given locus. In the eQTL analysis, we utilize a recently proposed hierarchical multiple testing strategy which controls error rates regarding the discovery of functional variants. Our results elucidate the heritability and regulation of gene expression in this unique Latin American study population and identify a set of regulatory SNPs which may be relevant in future investigations of complex disease in this population. Since our subjects belong to extended families, we are able to compare traditional kinship-based estimates with those from more recent methods that depend only on genotype information.

The Contribution of GWAS Loci in Familial Dyslipidemias.

Familial combined hyperlipidemia (FCH) is a complex and common familial dyslipidemia characterized by elevated total cholesterol and/or triglyceride levels with over five-fold risk of coronary heart disease. The genetic architecture and contribution of rare Mendelian and common variants to FCH susceptibility is unknown. In 53 Finnish FCH families, we genotyped and imputed nine million variants in 715 family members with DNA available. We studied the enrichment of variants previously implicated with monogenic dyslipidemias and/or lipid levels in the general population by comparing allele frequencies between the FCH families and population samples. We also constructed weighted polygenic scores using 212 lipid-associated SNPs and estimated the relative contributions of Mendelian variants and polygenic scores to the risk of FCH in the families. We identified, across the whole allele frequency spectrum, an enrichment of variants known to elevate, and a deficiency of variants known to lower LDL-C and/or TG levels among both probands and affected FCH individuals. The score based on TG associated SNPs was particularly high among affected individuals compared to non-affected family members. Out of 234 affected FCH individuals across the families, seven (3%) carried Mendelian variants and 83 (35%) showed high accumulation of either known LDL-C or TG elevating variants by having either polygenic score over the 90th percentile in the population. The positive predictive value of high score was much higher for affected FCH individuals than for similar sporadic cases in the population. FCH is highly polygenic, supporting the hypothesis that variants across the whole allele frequency spectrum contribute to this complex familial trait. Polygenic SNP panels improve identification of individuals affected with FCH, but their clinical utility remains to be defined.

Genetic contributions to circadian activity rhythm and sleep pattern phenotypes in pedigrees segregating for severe bipolar disorder.

Abnormalities in sleep and circadian rhythms are central features of bipolar disorder (BP), often persisting between episodes. We report here, to our knowledge, the first systematic analysis of circadian rhythm activity in pedigrees segregating severe BP (BP-I). By analyzing actigraphy data obtained from members of 26 Costa Rican and Colombian pedigrees [136 euthymic (i.e., interepisode) BP-I individuals and 422 non-BP-I relatives], we delineated 73 phenotypes, of which 49 demonstrated significant heritability and 13 showed significant trait-like association with BP-I. All BP-I-associated traits related to activity level, with BP-I individuals consistently demonstrating lower activity levels than their non-BP-I relatives. We analyzed all 49 heritable phenotypes using genetic linkage analysis, with special emphasis on phenotypes judged to have the strongest impact on the biology underlying BP. We identified a locus for interdaily stability of activity, at a threshold exceeding genome-wide significance, on chromosome 12pter, a region that also showed pleiotropic linkage to two additional activity phenotypes.

Re-sequencing expands our understanding of the phenotypic impact of variants at GWAS loci.

Genome-wide association studies (GWAS) have identified >500 common variants associated with quantitative metabolic traits, but in aggregate such variants explain at most 20-30% of the heritable component of population variation in these traits. To further investigate the impact of genotypic variation on metabolic traits, we conducted re-sequencing studies in >6,000 members of a Finnish population cohort (The Northern Finland Birth Cohort of 1966 [NFBC]) and a type 2 diabetes case-control sample (The Finland-United States Investigation of NIDDM Genetics [FUSION] study). By sequencing the coding sequence and 5' and 3' untranslated regions of 78 genes at 17 GWAS loci associated with one or more of six metabolic traits (serum levels of fasting HDL-C, LDL-C, total cholesterol, triglycerides, plasma glucose, and insulin), and conducting both single-variant and gene-level association tests, we obtained a more complete understanding of phenotype-genotype associations at eight of these loci. At all eight of these loci, the identification of new associations provides significant evidence for multiple genetic signals to one or more phenotypes, and at two loci, in the genes ABCA1 and CETP, we found significant gene-level evidence of association to non-synonymous variants with MAF<1%. Additionally, two potentially deleterious variants that demonstrated significant associations (rs138726309, a missense variant in G6PC2, and rs28933094, a missense variant in LIPC) were considerably more common in these Finnish samples than in European reference populations, supporting our prior hypothesis that deleterious variants could attain high frequencies in this isolated population, likely due to the effects of population bottlenecks. Our results highlight the value of large, well-phenotyped samples for rare-variant association analysis, and the challenge of evaluating the phenotypic impact of such variants.

Brain structure-function associations in multi-generational families genetically enriched for bipolar disorder.

Recent theories regarding the pathophysiology of bipolar disorder suggest contributions of both neurodevelopmental and neurodegenerative processes. While structural neuroimaging studies indicate disease-associated neuroanatomical alterations, the behavioural correlates of these alterations have not been well characterized. Here, we investigated multi-generational families genetically enriched for bipolar disorder to: (i) characterize neurobehavioural correlates of neuroanatomical measures implicated in the pathophysiology of bipolar disorder; (ii) identify brain-behaviour associations that differ between diagnostic groups; (iii) identify neurocognitive traits that show evidence of accelerated ageing specifically in subjects with bipolar disorder; and (iv) identify brain-behaviour correlations that differ across the age span. Structural neuroimages and multi-dimensional assessments of temperament and neurocognition were acquired from 527 (153 bipolar disorder and 374 non-bipolar disorder) adults aged 18-87 years in 26 families with heavy genetic loading for bipolar disorder. We used linear regression models to identify significant brain-behaviour associations and test whether brain-behaviour relationships differed: (i) between diagnostic groups; and (ii) as a function of age. We found that total cortical and ventricular volume had the greatest number of significant behavioural associations, and included correlations with measures from multiple cognitive domains, particularly declarative and working memory and executive function. Cortical thickness measures, in contrast, showed more specific associations with declarative memory, letter fluency and processing speed tasks. While the majority of brain-behaviour relationships were similar across diagnostic groups, increased cortical thickness in ventrolateral prefrontal and parietal cortical regions was associated with better declarative memory only in bipolar disorder subjects, and not in non-bipolar disorder family members. Additionally, while age had a relatively strong impact on all neurocognitive traits, the effects of age on cognition did not differ between diagnostic groups. Most brain-behaviour associations were also similar across the age range, with the exception of cortical and ventricular volume and lingual gyrus thickness, which showed weak correlations with verbal fluency and inhibitory control at younger ages that increased in magnitude in older subjects, regardless of diagnosis. Findings indicate that neuroanatomical traits potentially impacted by bipolar disorder are significantly associated with multiple neurobehavioural domains. Structure-function relationships are generally preserved across diagnostic groups, with the notable exception of ventrolateral prefrontal and parietal association cortex, volumetric increases in which may be associated with cognitive resilience specifically in individuals with bipolar disorder. Although age impacted all neurobehavioural traits, we did not find any evidence of accelerated cognitive decline specific to bipolar disorder subjects. Regardless of diagnosis, greater global brain volume may represent a protective factor for the effects of ageing on executive functioning.

Sequencing strategies and characterization of 721 vervet monkey genomes for future genetic analyses of medically relevant traits.

BACKGROUND: We report here the first genome-wide high-resolution polymorphism resource for non-human primate (NHP) association and linkage studies, constructed for the Caribbean-origin vervet monkey, or African green monkey (Chlorocebus aethiops sabaeus), one of the most widely used NHPs in biomedical research. We generated this resource by whole genome sequencing (WGS) of monkeys from the Vervet Research Colony (VRC), an NIH-supported research resource for which extensive phenotypic data are available. RESULTS: We identified genome-wide single nucleotide polymorphisms (SNPs) by WGS of 721 members of an extended pedigree from the VRC. From high-depth WGS data we identified more than 4 million polymorphic unequivocal segregating sites; by pruning these SNPs based on heterozygosity, quality control filters, and the degree of linkage disequilibrium (LD) between SNPs, we constructed genome-wide panels suitable for genetic association (about 500,000 SNPs) and linkage analysis (about 150,000 SNPs). To further enhance the utility of these resources for linkage analysis, we used a further pruned subset of the linkage panel to generate multipoint identity by descent matrices. CONCLUSIONS: The genetic and phenotypic resources now available for the VRC and other Caribbean-origin vervets enable their use for genetic investigation of traits relevant to human diseases.

Magnitude and distribution of linkage disequilibrium in population isolates and implications for genome-wide association studies.

The genome-wide distribution of linkage disequilibrium (LD) determines the strategy for selecting markers for association studies, but it varies between populations. We assayed LD in large samples (200 individuals) from each of 11 well-described population isolates and an outbred European-derived sample, using SNP markers spaced across chromosome 22. Most isolates show substantially higher levels of LD than the outbred sample and many fewer regions of very low LD (termed 'holes'). Young isolates known to have had relatively few founders show particularly extensive LD with very few holes; these populations offer substantial advantages for genome-wide association mapping.

A non-human primate system for large-scale genetic studies of complex traits.

Non-human primates provide genetic model systems biologically intermediate between humans and other mammalian model organisms. Populations of Caribbean vervet monkeys (Chlorocebus aethiops sabaeus) are genetically homogeneous and large enough to permit well-powered genetic mapping studies of quantitative traits relevant to human health, including expression quantitative trait loci (eQTL). Previous transcriptome-wide investigation in an extended vervet pedigree identified 29 heritable transcripts for which levels of expression in peripheral blood correlate strongly with expression levels in the brain. Quantitative trait linkage analysis using 261 microsatellite markers identified significant (n = 8) and suggestive (n = 4) linkages for 12 of these transcripts, including both cis- and trans-eQTL. Seven transcripts, located on different chromosomes, showed maximum linkage to markers in a single region of vervet chromosome 9; this observation suggests the possibility of a master trans-regulator locus in this region. For one cis-eQTL (at B3GALTL, beta-1,3-glucosyltransferase), we conducted follow-up single nucleotide polymorphism genotyping and fine-scale association analysis in a sample of unrelated Caribbean vervets, localizing this eQTL to a region of <200 kb. These results suggest the value of pedigree and population samples of the Caribbean vervet for linkage and association mapping studies of quantitative traits. The imminent whole genome sequencing of many of these vervet samples will enhance the power of such investigations by providing a comprehensive catalog of genetic variation.

Geospatial investigations in Colombia reveal variations in the distribution of mood and psychotic disorders.

BACKGROUND: Geographical variations in mood and psychotic disorders have been found in upper-income countries. We looked for geographic variation in these disorders in Colombia, a middle-income country. We analyzed electronic health records from the Clínica San Juan de Dios Manizales (CSJDM), which provides comprehensive mental healthcare for the one million inhabitants of Caldas. METHODS: We constructed a friction surface map of Caldas and used it to calculate the travel-time to the CSJDM for 16,295 patients who had received an initial diagnosis of mood or psychotic disorder. Using a zero-inflated negative binomial regression model, we determined the relationship between travel-time and incidence, stratified by disease severity. We employed spatial scan statistics to look for patient clusters. RESULTS: We show that travel-times (for driving) to the CSJDM are less than 1 h for ~50% of the population and more than 4 h for ~10%. We find a distance-decay relationship for outpatients, but not for inpatients: for every hour increase in travel-time, the number of expected outpatient cases decreases by 20% (RR = 0.80, 95% confidence interval [0.71, 0.89], p = 5.67E-05). We find nine clusters/hotspots of inpatients. CONCLUSIONS: Our results reveal inequities in access to healthcare: many individuals requiring only outpatient treatment may live too far from the CSJDM to access healthcare. Targeting of resources to comprehensively identify severely ill individuals living in the observed hotspots could further address treatment inequities and enable investigations to determine factors generating these hotspots.

The frequencies of mental disorders vary by geographic region. Investigating such variations may lead to more equitable access to mental healthcare and to scientific discoveries that reveal specific localized factors that contribute to the causes of mental illness. This study examined the frequency of three disorders with a major impact on public health ­ schizophrenia, bipolar disorder, and major depressive disorder ­ by analyzing electronic health records from a hospital providing comprehensive mental health care for a large region in Colombia. We show that individuals receiving outpatient care mainly live relatively near the facility. Those receiving inpatient care live throughout the region, but cluster in a few scattered locations. Future research could lead to strategies for more equitable provision of mental healthcare in Colombia and identify environmental or genetic factors that affect the likelihood that someone will develop one of these disorders.

Power of IRT in GWAS: successful QTL mapping of sum score phenotypes depends on interplay between risk allele frequency, variance explained by the risk allele, and test characteristics.

As data from sequencing studies in humans accumulate, rare genetic variants influencing liability to disease and disorders are expected to be identified. Three simulation studies show that characteristics and properties of diagnostic instruments interact with risk allele frequency to affect the power to detect a quantitative trait locus (QTL) based on a test score derived from symptom counts or questionnaire items. Clinical tests, that is, tests that show a positively skewed phenotypic sum score distribution in the general population, are optimal to find rare risk alleles of large effect. Tests that show a negatively skewed sum score distribution are optimal to find rare protective alleles of large effect. For alleles of small effect, tests with normally distributed item parameters give best power for a wide range of allele frequencies. The item-response theory framework can help understand why an existing measurement instrument has more power to detect risk alleles with either low or high frequency, or both kinds.

Phenotype mining in CNV carriers from a population cohort.

Phenotype mining is a novel approach for elucidating the genetic basis of complex phenotypic variation. It involves a search of rich phenotype databases for measures correlated with genetic variation, as identified in genome-wide genotyping or sequencing studies. An initial implementation of phenotype mining in a prospective unselected population cohort, the Northern Finland 1966 Birth Cohort (NFBC1966), identifies neurodevelopment-related traits-intellectual deficits, poor school performance and hearing abnormalities-which are more frequent among individuals with large (>500 kb) deletions than among other cohort members. Observation of extensive shared single nucleotide polymorphism haplotypes around deletions suggests an opportunity to expand phenotype mining from cohort samples to the populations from which they derive.

Predicting diagnostic conversion from major depressive disorder to bipolar disorder: an EHR based study from Colombia.

Bipolar Disorder (BD) is a severe and chronic disorder characterized by recurrent episodes of depression, mania, and/or hypomania. Most BD patients initially present with depressive symptoms, resulting in a delayed diagnosis of BD and poor clinical outcomes. This study leverages electronic health record (EHR) data from the Clínica San Juan de Dios Manizales in Colombia to identify features predictive of the transition from Major Depressive Disorder (MDD) to BD. Analyzing EHR data from 13,607 patients diagnosed with MDD over 15 years, we identified 1,610 cases of conversion to BD. Using a multivariate Cox regression model, we identified severity of the initial MDD episode, the presence of psychosis and hospitalization at first episode, family history of mood or psychotic disorders, female gender to be predictive of the conversion to BD. Additionally, we observed associations with medication classes (prescriptions of mood stabilizers, antipsychotics, and antidepressants) and clinical features (delusions, suicide attempt, suicidal ideation, use of marijuana and alcohol use/abuse) derived from natural language processing (NLP) of clinical notes. Together, these risk factors predicted BD conversion within five years of the initial MDD diagnosis, with a recall of 72% and a precision of 38%. Our study confirms many previously identified risk factors identified through registry-based studies (such as female gender and psychotic depression at the index MDD episode), and identifies novel ones (specifically, suicidal ideation and suicide attempt extracted from clinical notes). These results simultaneously demonstrate the validity of using EHR data for predicting BD conversion as well as underscore its potential for the identification of novel risk factors and improving early diagnosis.