Applications of machine learning and data mining methods to detect associations of rare and common variants with complex traits.

Machine learning methods (MLMs), designed to develop models using high-dimensional predictors, have been used to analyze genome-wide genetic and genomic data to predict risks for complex traits. We summarize the results from six contributions to our Genetic Analysis Workshop 18 working group; these investigators applied MLMs and data mining to analyses of rare and common genetic variants measured in pedigrees. To develop risk profiles, group members analyzed blood pressure traits along with single-nucleotide polymorphisms and rare variant genotypes derived from sequence and imputation analyses in large Mexican American pedigrees. Supervised MLMs included penalized regression with varying penalties, support vector machines, and permanental classification. Unsupervised MLMs included sparse principal components analysis and sparse graphical models. Entropy-based components analyses were also used to mine these data. None of the investigators fully capitalized on the genetic information provided by the complete pedigrees. Their approaches either corrected for the nonindependence of the individuals within the pedigrees or analyzed only those who were independent. Some methods allowed for covariate adjustment, whereas others did not. We evaluated these methods using a variety of metrics. Four contributors conducted primary analyses on the real data, and the other two research groups used the simulated data with and without knowledge of the underlying simulation model. One group used the answers to the simulated data to assess power and type I errors. Although the MLMs applied were substantially different, each research group concluded that MLMs have advantages over standard statistical approaches with these high-dimensional data.

Detection of macular ganglion cell loss in glaucoma by Fourier-domain optical coherence tomography.

PURPOSE: To map ganglion cell complex (GCC) thickness with high-speed Fourier-domain optical coherence tomography (FD-OCT) and compute novel macular parameters for glaucoma diagnosis. DESIGN: Observational, cross-sectional study. PARTICIPANTS: One hundred seventy-eight participants in the Advanced Imaging for Glaucoma Study, divided into 3 groups: 65 persons in the normal group, 78 in the perimetric glaucoma group (PG), and 52 in the preperimetric glaucoma group (PPG). METHODS: The RTVue FD-OCT system was used to map the macula over a 7 x 6 mm region. The macular OCT images were exported for automatic segmentation using software we developed. The program measured macular retinal (MR) thickness and GCC thickness. The GCC was defined as the combination of nerve fiber, ganglion cell, and inner plexiform layers. Pattern analysis was applied to the GCC map and the diagnostic powers of pattern-based diagnostic parameters were investigated. Results were compared with time-domain (TD) Stratus OCT measurements of MR and circumpapillary nerve fiber layer (NFL) thickness. MAIN OUTCOME MEASURES: Repeatability was assessed by intraclass correlation, pooled standard deviation, and coefficient of variation. Diagnostic power was assessed by the area under the receiver operator characteristic (AROC) curve. Measurements in the PG group were the primary measures of performance. RESULTS: The FD-OCT measurements of MR and GCC averages had significantly better repeatability than TD-OCT measurements of MR and NFL averages. The FD-OCT GCC average had significantly (P = 0.02) higher diagnostic power (AROC = 0.90) than MR (AROC = 0.85 for both FD-OCT and TD-OCT) in differentiating between PG and normal. One GCC pattern parameter, global loss volume, had significantly higher AROC (0.92) than the overall average (P = 0.01). The diagnostic powers of the best GCC parameters were statistically equal to TD-OCT NFL average. CONCLUSIONS: The higher speed and resolution of FD-OCT improved the repeatability of macular imaging compared with standard TD-OCT. Ganglion cell mapping and pattern analysis improved diagnostic power. The improved diagnostic power of macular GCC imaging is on par with, and complementary to, peripapillary NFL imaging. Macular imaging with FD-OCT is a useful method for glaucoma diagnosis and has potential for tracking glaucoma progression.

Reproducibility of tear meniscus measurement by Fourier-domain optical coherence tomography: a pilot study.

BACKGROUND AND OBJECTIVE: To study the reproducibility of tear meniscus measurement with high-speed high-resolution Fourier-domain optical coherence tomography (FD-OCT). PATIENTS AND METHODS: Twenty normal participants were enrolled in this prospective study. The lower tear meniscus in the right eye of each subject was imaged by vertical scans centered on the inferior cornea and the lower eyelid using an FD-OCT system (RTVue; Optovue, Inc., Fremont, CA) with a corneal adaptor. The system performs 26,000 axial scans per second and has a 5-micron axial resolution. Each subject was examined at two visits 30 to 60 days apart. Each eye was scanned twice on each visit. The scans were taken 2 seconds after a blink. The lower meniscus height, depth, and cornea-meniscus angle were measured with a computer caliper. The cross-sectional area was calculated using a two-triangle approximation. RESULTS: The between-visits coefficient of variation was 17.5%, 18.0%, 35.5%, and 12.2% for meniscus height, depth, area, and angle, respectively. The intraclass correlations for these parameters were 0.605, 0.558, 0.567, and 0.367, respectively. CONCLUSION: FD-OCT measures lower tear meniscus dimensions and area with higher between-visits reproducibility than previous OCT instruments. FD-OCT may be a useful way to measure dry eye severity and treatment effectiveness.

Mapping of macular substructures with optical coherence tomography for glaucoma diagnosis.

PURPOSE: To use optical coherence tomography (OCT) to identify the specific retinal layers and macular regions damaged in glaucoma. DESIGN: Observational cross-sectional study. PARTICIPANTS: One hundred forty-nine participants in the Advanced Imaging for Glaucoma Study, divided into 3 groups: normal (N) perimetric glaucoma (PG), and glaucoma suspect and preperimetric glaucoma (GSPPG) with 44, 73, and 29 persons, respectively. METHODS: The Zeiss Stratus OCT system (Carl Zeiss Meditec, Inc., Dublin, CA) was used to map the macula over a 6-mm diameter and to scan the circumpapillary nerve fiber layer (cpNFL). The macular OCT images were exported for automatic segmentation using software developed by the authors. The thickness of the macular nerve fiber layer (mNFL), ganglion cell layer (mGCL), inner plexiform layer (mIPL), inner nuclear layer (mINL), outer retinal layer (mORL), and total retinal thickness were measured. Thickness measurements of GSPPG and PG eyes were compared with those of N eyes. The ability to differentiate between GSPPG and PG eyes against N eyes was assessed by fractional loss, standardized deviation, and the area under the receiver operating characteristic curve. MAIN OUTCOME MEASURES: Area-weighted average thicknesses of retinal sublayers in the macula. RESULTS: The mNFL, mGCL, mIPL, and mINL were significantly (P<0.001) thinner in both the GSPPG and PG eyes than in the N eyes. In PG eyes, mNFL, mGCL, and mIPL thinning was most severe (approximately 20%), mINL thinning was intermediate (7%), and mORL thinning was minimal (3%). The repeatability (coefficient of variation and intraclass correlation) of thickness measurements was improved by combining the mNFL, mGCL, and mIPL measurements as the inner retinal layer (mIRL). The mIRL was the best macular parameter for glaucoma diagnosis and had discriminant power comparable with that of the cpNFL. The fractional loss of mIRL thickness was most severe in the inferior perifoveal region for both the PG and GSPPG groups. CONCLUSIONS: Glaucoma leads to thinning of the mNFL, mGCL, mIPL, and mINL, even before detectable visual field changes occur. A combination of the 3 innermost layers seems to provide optimal glaucoma detection. Increasing the sampling of peripheral macula with a new OCT scan pattern may improve glaucoma diagnosis further.

Combining nerve fiber layer parameters to optimize glaucoma diagnosis with optical coherence tomography.

PURPOSE: To identify the best combination of Stratus optical coherence tomography (OCT) retinal nerve fiber layer (RNFL) thickness parameters for the detection of glaucoma. DESIGN: Observational cross-sectional study. PARTICIPANTS: Eighty-nine age-matched normal and perimetric glaucoma participants enrolled in the Advanced Imaging for Glaucoma Study. METHODS: The Zeiss Stratus OCT system was used to obtain the circumpapillary RNFL thickness in both eyes of each participant. Right and left eye clock-hour data are analyzed together, assuming mirror-image symmetry. The RNFL diagnostic parameters were combined using either or-logic or and-logic approaches. MAIN OUTCOME MEASURES: Area under the receiver operating characteristic curve (AROC), sensitivity, and specificity are used to evaluate diagnostic performance. RESULTS: Overall average RNFL thickness has the highest AROC value (0.89) of all single parameters evaluated, followed by the inferior and superior quadrants (0.88 and 0.86, respectively). The clock hours with the best AROC values are in the inferior and superior quadrants. The highest AROC (0.92) was achieved by the or-logic combination of overall, inferior, and superior quadrant RNFL thicknesses. The 3-parameter combination was significantly better than the overall average alone (P = 0.01). The addition of more quadrants or clock hours to the combination reduced diagnostic performance. CONCLUSIONS: The best stand-alone diagnostic strategy for Stratus OCT RNFL data is to classify an eye as glaucomatous if the overall, inferior quadrant, or superior quadrant RNFL thickness average is below normal.

Correlation of Fourier domain optical coherence tomography retinal nerve fiber layer maps with visual fields in nonarteritic ischemic optic neuropathy.

BACKGROUND AND OBJECTIVE: To correlate peripapillary retinal nerve fiber layer (NFL) loss and visual field defects in nonarteritic ischemic optic neuropathy (NAION). PATIENTS AND METHODS: Patients with NAION and control subjects were enrolled in a case-control study. Participants were scanned with a Fourier domain optical coherence tomography (OCT) system. Peripapillary NFL thickness was averaged in hemispheric, quadrant, and octant divisions. Standard achromatic static perimetry was used to assess visual fields. RESULTS: The reproducibility of peripapillary NFL parameters was excellent in both the healthy and NAION groups. Eyes in the NAION group showed a significant decrease of peripapillary NFL thickness in terms of the overall average, all quadrant averages, and all octants. There were statistically significant correlations between the peripapillary NFL and visual fields in terms of both overall averages and superior-inferior differences. CONCLUSION: In NAION, the visual field and peripapillary NFL losses are correlated in both severity and location. Fourier domain OCT provides reproducible measurement of the peripapillary NFL and may be useful in the assessment of NAION.

Combining information from 3 anatomic regions in the diagnosis of glaucoma with time-domain optical coherence tomography.

PURPOSE: To improve the diagnosis of glaucoma by combining time-domain optical coherence tomography (TD-OCT) measurements of the optic disc, circumpapillary retinal nerve fiber layer (RNFL), and macular retinal thickness. PATIENTS AND METHODS: Ninety-six age-matched normal and 96 perimetric glaucoma participants were included in this observational, cross-sectional study. Or-logic, support vector machine, relevance vector machine, and linear discrimination function were used to analyze the performances of combined TD-OCT diagnostic variables. RESULTS: The area under the receiver-operating curve (AROC) was used to evaluate the diagnostic accuracy and to compare the diagnostic performance of single and combined anatomic variables. The best RNFL thickness variables were the inferior (AROC=0.900), overall (AROC=0.892), and superior quadrants (AROC=0.850). The best optic disc variables were horizontal integrated rim width (AROC=0.909), vertical integrated rim area (AROC=0.908), and cup/disc vertical ratio (AROC=0.890). All macular retinal thickness variables had AROCs of 0.829 or less. Combining the top 3 RNFL and optic disc variables in optimizing glaucoma diagnosis, support vector machine had the highest AROC, 0.954, followed by or-logic (AROC=0.946), linear discrimination function (AROC=0.946), and relevance vector machine (AROC=0.943). All combination diagnostic variables had significantly larger AROCs than any single diagnostic variable. There are no significant differences among the combination diagnostic indices. CONCLUSIONS: With TD-OCT, RNFL and optic disc variables had better diagnostic accuracy than macular retinal variables. Combining top RNFL and optic disc variables significantly improved diagnostic performance. Clinically, or-logic classification was the most practical analytical tool with sufficient accuracy to diagnose early glaucoma.

Support for calcium channel gene defects in autism spectrum disorders.

BACKGROUND: Alternation of synaptic homeostasis is a biological process whose disruption might predispose children to autism spectrum disorders (ASD). Calcium channel genes (CCG) contribute to modulating neuronal function and evidence implicating CCG in ASD has been accumulating. We conducted a targeted association analysis of CCG using existing genome-wide association study (GWAS) data and imputation methods in a combined sample of parent/affected child trios from two ASD family collections to explore this hypothesis. METHODS: A total of 2,176 single-nucleotide polymorphisms (SNP) (703 genotyped and 1,473 imputed) covering the genes that encode the α1 subunit proteins of 10 calcium channels were tested for association with ASD in a combined sample of 2,781 parent/affected child trios from 543 multiplex Caucasian ASD families from the Autism Genetics Resource Exchange (AGRE) and 1,651 multiplex and simplex Caucasian ASD families from the Autism Genome Project (AGP). SNP imputation using IMPUTE2 and a combined reference panel from the HapMap3 and the 1,000 Genomes Project increased coverage density of the CCG. Family-based association was tested using the FBAT software which controls for population stratification and accounts for the non-independence of siblings within multiplex families. The level of significance for association was set at 2.3E-05, providing a Bonferroni correction for this targeted 10-gene panel. RESULTS: Four SNPs in three CCGs were associated with ASD. One, rs10848653, is located in CACNA1C, a gene in which rare de novo mutations are responsible for Timothy syndrome, a Mendelian disorder that features ASD. Two others, rs198538 and rs198545, located in CACN1G, and a fourth, rs5750860, located in CACNA1I, are in CCGs that encode T-type calcium channels, genes with previous ASD associations. CONCLUSIONS: These associations support a role for common CCG SNPs in ASD.

Does optic nerve head size variation affect circumpapillary retinal nerve fiber layer thickness measurement by optical coherence tomography?

PURPOSE: To determine the relationship between retinal nerve fiber layer (RNFL) thickness, optic disc size, and image magnification. METHODS: The cohort consisted of 196 normal eyes of 101 participants in the Advanced Imaging for Glaucoma Study (AIGS), a multicenter, prospective, longitudinal study to develop advanced imaging technologies for glaucoma diagnosis. Scanning laser tomography was used to measure disc size. Optical coherence tomography (OCT) was used to perform circumpapillary RNFL thickness measurements using the standard fixed 3.46-mm nominal scan diameter. A theoretical model of magnification effects was developed to relate RNFL thickness (overall average) with axial length and magnification. RESULTS: Multivariate regression showed no significant correlation between RNFL thickness and optic disc area (95% confidence interval [CI] = -0.9 to 4.1 µm/mm², P = 0.21). Linear regression showed that RNFL thickness depended significantly on axial length (slope = -3.1 µm/mm, 95% CI = -4.9 to -1.3, P = 0.001) and age (slope = -0.3 µm/y, 95% CI = -0.5 to -0.2, P = 0.0002). The slope values agreed closely with the values predicted by the magnification model. CONCLUSIONS: There is no significant association between RNFL thickness and optic disc area. Previous publications that showed such an association may have been biased by the effect of axial length on fundus image magnification and, therefore, both measured RNFL thickness and apparent disc area. The true diameter of the circumpapillary OCT scan is larger for a longer eye (more myopic eye), leading to a thinner RNFL measurement. Adjustment of measured RNFL thickness by axial length, in addition to age, may lead to a tighter normative range and improve the detection of RNFL thinning due to glaucoma.

Prediction of serotonin transporter promoter polymorphism genotypes from single nucleotide polymorphism arrays using machine learning methods.

BACKGROUND: The serotonin transporter gene (SLC6A4) and its promoter (5-HTTLPR) polymorphism have been the focus of a large number of association studies of behavioral traits and psychiatric disorders. However, large-scale genotyping of the polymorphism has been very difficult. We report the development and validation of a 5-HTTLPR genotype prediction model. METHODS: The single nucleotide polymorphisms (SNPs) from the 2000 kb region surrounding 5-HTTLPR were used to construct a prediction model through a newly developed machine learning method, multicategory vertex discriminant analysis with 2147 individuals from the Northern Finnish Birth Cohort genotyped with the Illumina 370K SNP array and manually genotyped for 5-HTTLPR polymorphism. The prediction model was applied to SNP genotypes in a Dutch/German schizophrenia case-control sample of 3318 individuals to test the association of the polymorphism with schizophrenia. RESULT: The prediction model of eight SNPs achieved a 92.4% accuracy rate and a 0.98±0.01 area under the receiving operating characteristic. Evidence for an association of the polymorphism with schizophrenia was observed (P=0.05, odds ratio=1.105). CONCLUSION: This prediction model provides an effective substitute of manually genotyped 5-HTTLPR alleles, providing a new approach for large scale association studies of this polymorphism.

Weighted variance FBAT: a powerful method for including covariates in FBAT analyses.

The family-based association test (FBAT), an extension of transmission/disequilibrium test, capitalizes on linkage disequilibrium to assess the association of genetic markers and traits in nuclear families. It does not permit a formal inclusion of covariates, although an offset under the FBAT -o option allows for an overall trait-intercept adjustment. The PBAT software provides additional features and permits the inclusion of covariates in the FBAT test statistic, but does not account for the parental genotype information when the traits are adjusted for the covariates. We propose the weighted variance FBAT (WVF) method to generate trait values adjusted for both parental genotypes and covariate values. WVF is expected to be more powerful, because the variance is minimized considering both of these factors simultaneously using a weighted Gauss-Newton algorithm. Two simulated parent/child trio data sets, both with a covariate and the second with a gene by covariate interaction, were simulated to compare WVF power with FBAT and PBAT for a quantitative trait. WVF is most powerful when levels of significance are greater and covariates have a larger influence, indicating WVF may be especially effective when multiple comparisons are an important consideration, such as with whole genome association studies. WVF will also improve the cost of an association study when environmental covariates are considered. A SAS program (www.SAS.com) for generating WVF residuals that can be input to the current versions of the FBAT (www.biostat.harvard.edu/fbat/fbat.html) and PBAT (www.biostat.harvard.edu/clange/default.htm) software is provided.