What makes a good prediction? Feature importance and beginning to open the black box of machine learning in genetics.

Genetic data have become increasingly complex within the past decade, leading researchers to pursue increasingly complex questions, such as those involving epistatic interactions and protein prediction. Traditional methods are ill-suited to answer these questions, but machine learning (ML) techniques offer an alternative solution. ML algorithms are commonly used in genetics to predict or classify subjects, but some methods evaluate which features (variables) are responsible for creating a good prediction; this is called feature importance. This is critical in genetics, as researchers are often interested in which features (e.g., SNP genotype or environmental exposure) are responsible for a good prediction. This allows for the deeper analysis beyond simple prediction, including the determination of risk factors associated with a given phenotype. Feature importance further permits the researcher to peer inside the black box of many ML algorithms to see how they work and which features are critical in informing a good prediction. This review focuses on ML methods that provide feature importance metrics for the analysis of genetic data. Five major categories of ML algorithms: k nearest neighbors, artificial neural networks, deep learning, support vector machines, and random forests are described. The review ends with a discussion of how to choose the best machine for a data set. This review will be particularly useful for genetic researchers looking to use ML methods to answer questions beyond basic prediction and classification.

Genetic Association Reveals Protection against Recurrence of Clostridium difficile Infection with Bezlotoxumab Treatment.

Bezlotoxumab is a human monoclonal antibody against Clostridium difficile toxin B, indicated to prevent recurrence of C. difficile infection (rCDI) in high-risk adults receiving antibacterial treatment for CDI. An exploratory genome-wide association study investigated whether human genetic variation influences bezlotoxumab response. DNA from 704 participants who achieved initial clinical cure in the phase 3 MODIFY I/II trials was genotyped. Single nucleotide polymorphisms (SNPs) and human leukocyte antigen (HLA) imputation were performed using IMPUTE2 and HIBAG, respectively. A joint test of genotype and genotype-by-treatment interaction in a logistic regression model was used to screen genetic variants associated with response to bezlotoxumab. The SNP rs2516513 and the HLA alleles HLA-DRB1*07:01 and HLA-DQA1*02:01, located in the extended major histocompatibility complex on chromosome 6, were associated with the reduction of rCDI in bezlotoxumab-treated participants. Carriage of a minor allele (homozygous or heterozygous) at any of the identified loci was related to a larger difference in the proportion of participants experiencing rCDI versus placebo; the effect was most prominent in the subgroup at high baseline risk for rCDI. Genotypes associated with an improved bezlotoxumab response showed no association with rCDI in the placebo cohort. These data suggest that a host-driven, immunological mechanism may impact bezlotoxumab response. Trial registration numbers are as follows: NCT01241552 (MODIFY I) and NCT01513239 (MODIFY II).IMPORTANCEClostridium difficile infection is associated with significant clinical morbidity and mortality; antibacterial treatments are effective, but recurrence of C. difficile infection is common. In this genome-wide association study, we explored whether host genetic variability affected treatment responses to bezlotoxumab, a human monoclonal antibody that binds C. difficile toxin B and is indicated for the prevention of recurrent C. difficile infection. Using data from the MODIFY I/II phase 3 clinical trials, we identified three genetic variants associated with reduced rates of C. difficile infection recurrence in bezlotoxumab-treated participants. The effects were most pronounced in participants at high risk of C. difficile infection recurrence. All three variants are located in the extended major histocompatibility complex on chromosome 6, suggesting the involvement of a host-driven immunological mechanism in the prevention of C. difficile infection recurrence.

Exome genotyping and linkage analysis identifies two novel linked regions and replicates two others for myopia in Ashkenazi Jewish families.

BACKGROUND: Myopia is one of most common eye diseases in the world and affects 1 in 4 Americans. It is a complex disease caused by both environmental and genetics effects; the genetics effects are still not well understood. In this study, we performed genetic linkage analyses on Ashkenazi Jewish families with a strong familial history of myopia to elucidate any potential causal genes. METHODS: Sixty-four extended Ashkenazi Jewish families were previously collected from New Jersey. Genotypes from the Illumina ExomePlus array were merged with prior microsatellite linkage data from these families. Additional custom markers were added for candidate regions reported in literature for myopia or refractive error. Myopia was defined as mean spherical equivalent (MSE) of -1D or worse and parametric two-point linkage analyses (using TwoPointLods) and multi-point linkage analyses (using SimWalk2) were performed as well as collapsed haplotype pattern (CHP) analysis in SEQLinkage and association analyses performed with FBAT and rv-TDT. RESULTS: Strongest evidence of linkage was on 1p36(two-point LOD = 4.47) a region previously linked to refractive error (MYP14) but not myopia. Another genome-wide significant locus was found on 8q24.22 with a maximum two-point LOD score of 3.75. CHP analysis also detected the signal on 1p36, localized to the LINC00339 gene with a maximum HLOD of 3.47, as well as genome-wide significant signals on 7q36.1 and 11p15, which overlaps with the MYP7 locus. CONCLUSIONS: We identified 2 novel linkage peaks for myopia on chromosomes 7 and 8 in these Ashkenazi Jewish families and replicated 2 more loci on chromosomes 1 and 11, one previously reported in refractive error but not myopia in these families and the other locus previously reported in the literature. Strong candidate genes have been identified within these linkage peaks in our families. Targeted sequencing in these regions will be necessary to definitively identify causal variants under these linkage peaks.

Analysis of sequence data to identify potential risk variants for oral clefts in multiplex families.

BACKGROUND: Nonsyndromic oral clefts are craniofacial malformations, which include cleft lip with or without cleft palate. The etiology for oral clefts is complex with both genetic and environmental factors contributing to risk. Previous genome-wide association (GWAS) studies have identified multiple loci with small effects; however, many causal variants remain elusive. METHODS: In this study, we address this by specifically looking for rare, potentially damaging variants in family-based data. We analyzed both whole exome sequence (WES) data and whole genome sequence (WGS) data in multiplex cleft families to identify variants shared by affected individuals. RESULTS: Here we present the results from these analyses. Our most interesting finding was from a single Syrian family, which showed enrichment of nonsynonymous and potentially damaging rare variants in two genes: CASP9 and FAT4. CONCLUSION: Neither of these candidate genes has previously been associated with oral clefts and, if confirmed as contributing to disease risk, may indicate novel biological pathways in the genetic etiology for oral clefts.

Discovery and replication of SNP-SNP interactions for quantitative lipid traits in over 60,000 individuals.

BACKGROUND: The genetic etiology of human lipid quantitative traits is not fully elucidated, and interactions between variants may play a role. We performed a gene-centric interaction study for four different lipid traits: low-density lipoprotein cholesterol (LDL-C), high-density lipoprotein cholesterol (HDL-C), total cholesterol (TC), and triglycerides (TG). RESULTS: Our analysis consisted of a discovery phase using a merged dataset of five different cohorts (n = 12,853 to n = 16,849 depending on lipid phenotype) and a replication phase with ten independent cohorts totaling up to 36,938 additional samples. Filters are often applied before interaction testing to correct for the burden of testing all pairwise interactions. We used two different filters: 1. A filter that tested only single nucleotide polymorphisms (SNPs) with a main effect of p < 0.001 in a previous association study. 2. A filter that only tested interactions identified by Biofilter 2.0. Pairwise models that reached an interaction significance level of p < 0.001 in the discovery dataset were tested for replication. We identified thirteen SNP-SNP models that were significant in more than one replication cohort after accounting for multiple testing. CONCLUSIONS: These results may reveal novel insights into the genetic etiology of lipid levels. Furthermore, we developed a pipeline to perform a computationally efficient interaction analysis with multi-cohort replication.

Comparison of parametric and machine methods for variable selection in simulated Genetic Analysis Workshop 19 data.

Current findings from genetic studies of complex human traits often do not explain a large proportion of the estimated variation of these traits due to genetic factors. This could be, in part, due to overly stringent significance thresholds in traditional statistical methods, such as linear and logistic regression. Machine learning methods, such as Random Forests (RF), are an alternative approach to identify potentially interesting variants. One major issue with these methods is that there is no clear way to distinguish between probable true hits and noise variables based on the importance metric calculated. To this end, we are developing a method called the Relative Recurrency Variable Importance Metric (r2VIM), a RF-based variable selection method. Here, we apply r2VIM to the unrelated Genetic Analysis Workshop 19 data with simulated systolic blood pressure as the phenotype. We compare the number of "true" functional variants identified by r2VIM with those identified by linear regression analyses that use a Bonferroni correction to calculate a significance threshold. Our results show that r2VIM performed comparably to linear regression. Our findings are proof-of-concept for r2VIM, as it identifies a similar number of functional and nonfunctional variants as a more commonly used technique when the optimal importance score threshold is used.

Machine learning and data mining in complex genomic data--a review on the lessons learned in Genetic Analysis Workshop 19.

In the analysis of current genomic data, application of machine learning and data mining techniques has become more attractive given the rising complexity of the projects. As part of the Genetic Analysis Workshop 19, approaches from this domain were explored, mostly motivated from two starting points. First, assuming an underlying structure in the genomic data, data mining might identify this and thus improve downstream association analyses. Second, computational methods for machine learning need to be developed further to efficiently deal with the current wealth of data.In the course of discussing results and experiences from the machine learning and data mining approaches, six common messages were extracted. These depict the current state of these approaches in the application to complex genomic data. Although some challenges remain for future studies, important forward steps were taken in the integration of different data types and the evaluation of the evidence. Mining the data for underlying genetic or phenotypic structure and using this information in subsequent analyses proved to be extremely helpful and is likely to become of even greater use with more complex data sets.

Identifying gene-gene interactions that are highly associated with Body Mass Index using Quantitative Multifactor Dimensionality Reduction (QMDR).

BACKGROUND: Despite heritability estimates of 40-70 % for obesity, less than 2 % of its variation is explained by Body Mass Index (BMI) associated loci that have been identified so far. Epistasis, or gene-gene interactions are a plausible source to explain portions of the missing heritability of BMI. METHODS: Using genotypic data from 18,686 individuals across five study cohorts - ARIC, CARDIA, FHS, CHS, MESA - we filtered SNPs (Single Nucleotide Polymorphisms) using two parallel approaches. SNPs were filtered either on the strength of their main effects of association with BMI, or on the number of knowledge sources supporting a specific SNP-SNP interaction in the context of BMI. Filtered SNPs were specifically analyzed for interactions that are highly associated with BMI using QMDR (Quantitative Multifactor Dimensionality Reduction). QMDR is a nonparametric, genetic model-free method that detects non-linear interactions associated with a quantitative trait. RESULTS: We identified seven novel, epistatic models with a Bonferroni corrected p-value of association < 0.1. Prior experimental evidence helps explain the plausible biological interactions highlighted within our results and their relationship with obesity. We identified interactions between genes involved in mitochondrial dysfunction (POLG2), cholesterol metabolism (SOAT2), lipid metabolism (CYP11B2), cell adhesion (EZR), cell proliferation (MAP2K5), and insulin resistance (IGF1R). Moreover, we found an 8.8 % increase in the variance in BMI explained by these seven SNP-SNP interactions, beyond what is explained by the main effects of an index FTO SNP and the SNPs within these interactions. We also replicated one of these interactions and 58 proxy SNP-SNP models representing it in an independent dataset from the eMERGE study. CONCLUSION: This study highlights a novel approach for discovering gene-gene interactions by combining methods such as QMDR with traditional statistics.

Next-generation analysis of cataracts: determining knowledge driven gene-gene interactions using biofilter, and gene-environment interactions using the Phenx Toolkit*.

Investigating the association between biobank derived genomic data and the information of linked electronic health records (EHRs) is an emerging area of research for dissecting the architecture of complex human traits, where cases and controls for study are defined through the use of electronic phenotyping algorithms deployed in large EHR systems. For our study, cataract cases and controls were identified within the Marshfield Personalized Medicine Research Project (PMRP) biobank and linked EHR, which is a member of the NHGRI-funded electronic Medical Records and Genomics (eMERGE) Network. Our goal was to explore potential gene-gene and gene-environment interactions within these data for 527,953 and 527,936 single nucleotide polymorphisms (SNPs) for gene-gene and gene-environment analyses, respectively, with minor allele frequency > 1%, in order to explore higher level associations with cataract risk beyond investigations of single SNP-phenotype associations. To build our SNP-SNP interaction models we utilized a prior-knowledge driven filtering method called Biofilter to minimize the multiple testing burden of exploring the vast array of interaction models possible from our extensive number of SNPs. Using Biofilter, we developed 57,376 prior-knowledge directed SNP-SNP models to test for association with cataract status. We selected models that required 6 sources of external domain knowledge. We identified 13 statistically significant SNP-SNP models with an interaction with p-value < 1 × 10(-4), as well as an overall model with p-value < 0.01 associated with cataract status. We also conducted gene-environment interaction analyses for all GWAS SNPs and a set of environmental factors from the PhenX Toolkit: smoking, UV exposure, and alcohol use;these environmental factors have been previously associated with the formation of cataracts. We found a total of 782 gene-environment models that exhibit an interaction with a p-value < 1 × 10(-4) associatedwith cataract status. Our results show these approaches enable advanced searches for epistasis and gene-environment interactions beyond GWAS, and that the EHR based approach provides an additional source of data for seeking these advanced explanatory models of the etiology of complex disease/outcome such as cataracts.

Methods of integrating data to uncover genotype-phenotype interactions.

Recent technological advances have expanded the breadth of available omic data, from whole-genome sequencing data, to extensive transcriptomic, methylomic and metabolomic data. A key goal of analyses of these data is the identification of effective models that predict phenotypic traits and outcomes, elucidating important biomarkers and generating important insights into the genetic underpinnings of the heritability of complex traits. There is still a need for powerful and advanced analysis strategies to fully harness the utility of these comprehensive high-throughput data, identifying true associations and reducing the number of false associations. In this Review, we explore the emerging approaches for data integration - including meta-dimensional and multi-staged analyses - which aim to deepen our understanding of the role of genetics and genomics in complex outcomes. With the use and further development of these approaches, an improved understanding of the relationship between genomic variation and human phenotypes may be revealed.

ATHENA: the analysis tool for heritable and environmental network associations.

MOTIVATION: Advancements in high-throughput technology have allowed researchers to examine the genetic etiology of complex human traits in a robust fashion. Although genome-wide association studies have identified many novel variants associated with hundreds of traits, a large proportion of the estimated trait heritability remains unexplained. One hypothesis is that the commonly used statistical techniques and study designs are not robust to the complex etiology that may underlie these human traits. This etiology could include non-linear gene × gene or gene × environment interactions. Additionally, other levels of biological regulation may play a large role in trait variability. RESULTS: To address the need for computational tools that can explore enormous datasets to detect complex susceptibility models, we have developed a software package called the Analysis Tool for Heritable and Environmental Network Associations (ATHENA). ATHENA combines various variable filtering methods with machine learning techniques to analyze high-throughput categorical (i.e. single nucleotide polymorphisms) and quantitative (i.e. gene expression levels) predictor variables to generate multivariable models that predict either a categorical (i.e. disease status) or quantitative (i.e. cholesterol levels) outcomes. The goal of this article is to demonstrate the utility of ATHENA using simulated and biological datasets that consist of both single nucleotide polymorphisms and gene expression variables to identify complex prediction models. Importantly, this method is flexible and can be expanded to include other types of high-throughput data (i.e. RNA-seq data and biomarker measurements). AVAILABILITY: ATHENA is freely available for download. The software, user manual and tutorial can be downloaded from http://ritchielab.psu.edu/ritchielab/software.

Next-generation analysis of cataracts: determining knowledge driven gene-gene interactions using Biofilter, and gene-environment interactions using the PhenX Toolkit.

Investigating the association between biobank derived genomic data and the information of linked electronic health records (EHRs) is an emerging area of research for dissecting the architecture of complex human traits, where cases and controls for study are defined through the use of electronic phenotyping algorithms deployed in large EHR systems. For our study, 2580 cataract cases and 1367 controls were identified within the Marshfield Personalized Medicine Research Project (PMRP) Biobank and linked EHR, which is a member of the NHGRI-funded electronic Medical Records and Genomics (eMERGE) Network. Our goal was to explore potential gene-gene and gene-environment interactions within these data for 529,431 single nucleotide polymorphisms (SNPs) with minor allele frequency > 1%, in order to explore higher level associations with cataract risk beyond investigations of single SNP-phenotype associations. To build our SNP-SNP interaction models we utilized a prior-knowledge driven filtering method called Biofilter to minimize the multiple testing burden of exploring the vast array of interaction models possible from our extensive number of SNPs. Using the Biofilter, we developed 57,376 prior-knowledge directed SNP-SNP models to test for association with cataract status. We selected models that required 6 sources of external domain knowledge. We identified 5 statistically significant models with an interaction term with p-value < 0.05, as well as an overall model with p-value < 0.05 associated with cataract status. We also conducted gene-environment interaction analyses for all GWAS SNPs and a set of environmental factors from the PhenX Toolkit: smoking, UV exposure, and alcohol use; these environmental factors have been previously associated with the formation of cataracts. We found a total of 288 models that exhibit an interaction term with a p-value ≤ 1×10(-4) associated with cataract status. Our results show these approaches enable advanced searches for epistasis and gene-environment interactions beyond GWAS, and that the EHR based approach provides an additional source of data for seeking these advanced explanatory models of the etiology of complex disease/outcome such as cataracts.

ATHENA: a tool for meta-dimensional analysis applied to genotypes and gene expression data to predict HDL cholesterol levels.

Technology is driving the field of human genetics research with advances in techniques to generate high-throughput data that interrogate various levels of biological regulation. With this massive amount of data comes the important task of using powerful bioinformatics techniques to sift through the noise to find true signals that predict various human traits. A popular analytical method thus far has been the genome-wide association study (GWAS), which assesses the association of single nucleotide polymorphisms (SNPs) with the trait of interest. Unfortunately, GWAS has not been able to explain a substantial proportion of the estimated heritability for most complex traits. Due to the inherently complex nature of biology, this phenomenon could be a factor of the simplistic study design. A more powerful analysis may be a systems biology approach that integrates different types of data, or a meta-dimensional analysis. For this study we used the Analysis Tool for Heritable and Environmental Network Associations (ATHENA) to integrate high-throughput SNPs and gene expression variables (EVs) to predict high-density lipoprotein cholesterol (HDL-C) levels. We generated multivariable models that consisted of SNPs only, EVs only, and SNPs + EVs with testing r-squared values of 0.16, 0.11, and 0.18, respectively. Additionally, using just the SNPs and EVs from the best models, we generated a model with a testing r-squared of 0.32. A linear regression model with the same variables resulted in an adjusted r-squared of 0.23. With this systems biology approach, we were able to integrate different types of high-throughput data to generate meta-dimensional models that are predictive for the HDL-C in our data set. Additionally, our modeling method was able to capture more of the HDL-C variation than a linear regression model that included the same variables.

Mitochondrial DNA variation and HIV-associated sensory neuropathy in CHARTER.

HIV-associated sensory neuropathy remains an important complication of combination antiretroviral therapy and HIV infection. Mitochondrial DNA haplogroups and single nucleotide polymorphisms (SNPs) have previously been associated with symptomatic neuropathy in clinical trial participants. We examined associations between mitochondrial DNA variation and HIV-associated sensory neuropathy in CNS HIV Antiretroviral Therapy Effects Research (CHARTER). CHARTER is a USA-based longitudinal observational study of HIV-infected adults who underwent a structured interview and standardized examination. HIV-associated sensory neuropathy was determined by trained examiners as ≥1 sign (diminished vibratory and sharp-dull discrimination or ankle reflexes) bilaterally. Mitochondrial DNA sequencing was performed and haplogroups were assigned by published algorithms. Multivariable logistic regression of associations between mitochondrial DNA SNPs, haplogroups, and HIV-associated sensory neuropathy were performed. In analyses of associations of each mitochondrial DNA SNP with HIV-associated sensory neuropathy, the two most significant SNPs were at positions A12810G [odds ratio (95 % confidence interval) = 0.27 (0.11-0.65); p = 0.004] and T489C [odds ratio (95 % confidence interval) = 0.41 (0.21-0.80); p = 0.009]. These synonymous changes are known to define African haplogroup L1c and European haplogroup J, respectively. Both haplogroups were associated with decreased prevalence of HIV-associated sensory neuropathy compared with all other haplogroups [odds ratio (95 % confidence interval) = 0.29 (0.12-0.71); p = 0.007 and odds ratio (95 % confidence interval) = 0.42 (0.18-1.0); p = 0.05, respectively]. In conclusion, in this cohort of mostly combination antiretroviral therapy-treated subjects, two common mitochondrial DNA SNPs and their corresponding haplogroups were associated with a markedly decreased prevalence of HIV-associated sensory neuropathy.

Genome-wide association study of plasma efavirenz pharmacokinetics in AIDS Clinical Trials Group protocols implicates several CYP2B6 variants.

OBJECTIVES: Prior candidate gene studies have associated CYP2B6 516G→T [rs3745274] and 983T→C [rs28399499] with increased plasma efavirenz exposure. We sought to identify novel variants associated with efavirenz pharmacokinetics. MATERIALS AND METHODS: Antiretroviral therapy-naive AIDS Clinical Trials Group studies A5202, A5095, and ACTG 384 included plasma sampling for efavirenz pharmacokinetics. Log-transformed trough efavirenz concentrations (Cmin) were previously estimated by population pharmacokinetic modeling. Stored DNA was genotyped with Illumina HumanHap 650Y or 1MDuo platforms, complemented by additional targeted genotyping of CYP2B6 and CYP2A6 with MassARRAY iPLEX Gold. Associations were identified by linear regression, which included principal component vectors to adjust for genetic ancestry. RESULTS: Among 856 individuals, CYP2B6 516G→T was associated with efavirenz estimated Cmin (P=8.5×10). After adjusting for CYP2B6 516G→T, CYP2B6 983T→C was associated (P=9.9×10). After adjusting for both CYP2B6 516G→T and 983T→C, a CYP2B6 variant (rs4803419) in intron 3 was associated (P=4.4×10). After adjusting for all the three variants, non-CYP2B6 polymorphisms were associated at P-value less than 5×10. In a separate cohort of 240 individuals, only the three CYP2B6 polymorphisms replicated. These three polymorphisms explained 34% of interindividual variability in efavirenz estimated Cmin. The extensive metabolizer phenotype was best defined by the absence of all three polymorphisms. CONCLUSION: Three CYP2B6 polymorphisms were independently associated with efavirenz estimated Cmin at genome-wide significance, and explained one-third of interindividual variability. These data will inform continued efforts to translate pharmacogenomic knowledge into optimal efavirenz utilization.

Integrating heterogeneous high-throughput data for meta-dimensional pharmacogenomics and disease-related studies.

The current paradigm of human genetics research is to analyze variation of a single data type (i.e., DNA sequence or RNA levels) to detect genes and pathways that underlie complex traits such as disease state or drug response. While these studies have detected thousands of variations that associate with hundreds of complex phenotypes, much of the estimated heritability, or trait variability due to genetic factors, remain unexplained. We may be able to account for a portion of the missing heritability if we incorporate a systems biology approach into these analyses. Rapid technological advances will make it possible for scientists to explore this hypothesis via the generation of high-throughput omics data - transcriptomic, proteomic and methylomic to name a few. Analyzing this 'meta-dimensional' data will require clever statistical techniques that allow for the integration of qualitative and quantitative predictor variables. For this article, we examine two major categories of approaches for integrated data analysis, give examples of their use in experimental and in silico datasets, and assess the limitations of each method.

Initialization Parameter Sweep in ATHENA: Optimizing Neural Networks for Detecting Gene-Gene Interactions in the Presence of Small Main Effects.

Recent advances in genotyping technology have led to the generation of an enormous quantity of genetic data. Traditional methods of statistical analysis have proved insufficient in extracting all of the information about the genetic components of common, complex human diseases. A contributing factor to the problem of analysis is that amongst the small main effects of each single gene on disease susceptibility, there are non-linear, gene-gene interactions that can be difficult for traditional, parametric analyses to detect. In addition, exhaustively searching all multi-locus combinations has proved computationally impractical. Novel strategies for analysis have been developed to address these issues. The Analysis Tool for Heritable and Environmental Network Associations (ATHENA) is an analytical tool that incorporates grammatical evolution neural networks (GENN) to detect interactions among genetic factors. Initial parameters define how the evolutionary process will be implemented. This research addresses how different parameter settings affect detection of disease models involving interactions. In the current study, we iterate over multiple parameter values to determine which combinations appear optimal for detecting interactions in simulated data for multiple genetic models. Our results indicate that the factors that have the greatest influence on detection are: input variable encoding, population size, and parallel computation.