A primer on the use of machine learning to distil knowledge from data in biological psychiatry.

Applications of machine learning in the biomedical sciences are growing rapidly. This growth has been spurred by diverse cross-institutional and interdisciplinary collaborations, public availability of large datasets, an increase in the accessibility of analytic routines, and the availability of powerful computing resources. With this increased access and exposure to machine learning comes a responsibility for education and a deeper understanding of its bases and bounds, borne equally by data scientists seeking to ply their analytic wares in medical research and by biomedical scientists seeking to harness such methods to glean knowledge from data. This article provides an accessible and critical review of machine learning for a biomedically informed audience, as well as its applications in psychiatry. The review covers definitions and expositions of commonly used machine learning methods, and historical trends of their use in psychiatry. We also provide a set of standards, namely Guidelines for REporting Machine Learning Investigations in Neuropsychiatry (GREMLIN), for designing and reporting studies that use machine learning as a primary data-analysis approach. Lastly, we propose the establishment of the Machine Learning in Psychiatry (MLPsych) Consortium, enumerate its objectives, and identify areas of opportunity for future applications of machine learning in biological psychiatry. This review serves as a cautiously optimistic primer on machine learning for those on the precipice as they prepare to dive into the field, either as methodological practitioners or well-informed consumers.

A review of neuroeconomic gameplay in psychiatric disorders.

Abnormalities in social interaction are a common feature of several psychiatric disorders, aligning with the recent move towards using Research Domain Criteria (RDoC) to describe disorders in terms of observable behaviours rather than using specific diagnoses. Neuroeconomic games are an effective measure of social decision-making that can be adapted for use in neuroimaging, allowing investigation of the biological basis for behaviour. This review summarises findings of neuroeconomic gameplay studies in Axis 1 psychiatric disorders and advocates the use of these games as measures of the RDoC Affiliation and Attachment, Reward Responsiveness, Reward Learning and Reward Valuation constructs. Although research on neuroeconomic gameplay is in its infancy, consistencies have been observed across disorders, particularly in terms of impaired integration of social and cognitive information, avoidance of negative social interactions and reduced reward sensitivity, as well as a reduction in activity in brain regions associated with processing and responding to social information.

Genome-wide association study of antidepressant treatment resistance in a population-based cohort using health service prescription data and meta-analysis with GENDEP.

Antidepressants demonstrate modest response rates in the treatment of major depressive disorder (MDD). Despite previous genome-wide association studies (GWAS) of antidepressant treatment response, the underlying genetic factors are unknown. Using prescription data in a population and family-based cohort (Generation Scotland: Scottish Family Health Study; GS:SFHS), we sought to define a measure of (a) antidepressant treatment resistance and (b) stages of antidepressant resistance by inferring antidepressant switching as non-response to treatment. GWAS were conducted separately for antidepressant treatment resistance in GS:SFHS and the Genome-based Therapeutic Drugs for Depression (GENDEP) study and then meta-analysed (meta-analysis n = 4213, cases = 358). For stages of antidepressant resistance, a GWAS on GS:SFHS only was performed (n = 3452). Additionally, we conducted gene-set enrichment, polygenic risk scoring (PRS) and genetic correlation analysis. We did not identify any significant loci, genes or gene sets associated with antidepressant treatment resistance or stages of resistance. Significant positive genetic correlations of antidepressant treatment resistance and stages of resistance with neuroticism, psychological distress, schizotypy and mood disorder traits were identified. These findings suggest that larger sample sizes are needed to identify the genetic architecture of antidepressant treatment response, and that population-based observational studies may provide a tractable approach to achieving the necessary statistical power.

A major role for common genetic variation in anxiety disorders.

Anxiety disorders are common, complex psychiatric disorders with twin heritabilities of 30-60%. We conducted a genome-wide association study of Lifetime Anxiety Disorder (ncase = 25 453, ncontrol = 58 113) and an additional analysis of Current Anxiety Symptoms (ncase = 19 012, ncontrol = 58 113). The liability scale common variant heritability estimate for Lifetime Anxiety Disorder was 26%, and for Current Anxiety Symptoms was 31%. Five novel genome-wide significant loci were identified including an intergenic region on chromosome 9 that has previously been associated with neuroticism, and a locus overlapping the BDNF receptor gene, NTRK2. Anxiety showed significant positive genetic correlations with depression and insomnia as well as coronary artery disease, mirroring findings from epidemiological studies. We conclude that common genetic variation accounts for a substantive proportion of the genetic architecture underlying anxiety.

The role of polygenic risk score gene-set analysis in the context of the omnigenic model of schizophrenia.

A recent development in the genetic architecture of schizophrenia suggested that an omnigenic model may underlie the risk for this disorder. The aim of our study was to use polygenic profile scoring to quantitatively assess whether a number of experimentally derived sets would contribute to the disorder above and beyond the omnigenic effect. Using the PGC2 secondary analysis schizophrenia case-control cohort (N = 29,125 cases and 34,836 controls), a robust polygenic signal was observed from gene sets based on TCF4, FMR1, upregulation from MIR137 and downregulation from CHD8. Additional analyses revealed a constant floor effect in the amount of variance explained, consistent with the omnigenic model. Thus, we report that putative core gene sets showed a significant effect above and beyond the floor effect that might be linked with the underlying omnigenic background. In addition, we demonstrate a method to quantify the contribution of specific gene sets within the omnigenic context.

Pharmaco-epidemiology of antidepressant exposure in a UK cohort record-linkage study.

OBJECTIVES: Antidepressants are the most commonly prescribed psychiatric medication but concern has been raised about significant increases in their usage in high income countries. We aimed to quantify antidepressant prevalence, incidence, adherence and predictors of use in the adult population. METHODS: The study record-linked administrative prescribing and morbidity data to the Generation Scotland cohort ( N = 11,052), between 2009 and 2016. Prevalence and incidence of any antidepressant use was determined. Antidepressant adherence was measured using Proportion of Days Covered and Medication Possession Ratio. Time-to-event analysis for incident antidepressant use within 5 years of Generation Scotland: Scottish Family Health Study (GS:SFHS) recruitment was performed to reveal patient-level predictors of use. RESULTS: Almost one-third (28.0%, 95%CI 26.9-29.1) of the adults in our sample were prescribed at least one antidepressant in the 5-year period 2012-2016. There was a 36.2% increase in annual prevalence between 2010 and 2016. Incidence was 2.4(2.1-2.7)% per year. The majority of antidepressant episodes (57.6%) were greater than 9 months duration and adherence was generally high (69.0% with Proportion of Days Covered >80%). Predictors of new antidepressant use included history of affective disorder, being female, physical comorbidities, higher neuroticism scores, and lower cognitive function scores. CONCLUSIONS: Antidepressant prevalence is greater than previously reported but incidence remains relatively stable. We found the majority of antidepressant episodes to be of relatively long duration with good estimated adherence. Our study supports the hypothesis that increased long-term use among existing (and returning) users, along with wider ranges of indications for antidepressants, has significantly increased the prevalence of these medications.

Using tree-based methods for detection of gene-gene interactions in the presence of a polygenic signal: simulation study with application to educational attainment in the Generation Scotland Cohort Study.

Motivation: The genomic architecture of human complex diseases is thought to be attributable to single markers, polygenic components and epistatic components. No study has examined the ability of tree-based methods to detect epistasis in the presence of a polygenic signal. We sought to apply decision tree-based methods, C5.0 and logic regression, to detect epistasis under several simulated conditions, varying strength of interaction and linkage disequilibrium (LD) structure. We then applied the same methods to the phenotype of educational attainment in a large population cohort. Results: LD pruning improved the power and reduced the type I error. C5.0 had a conservative type I error rate whereas logic regression had a type I error rate that exceeded 5%. Despite the more conservative type I error, C5.0 was observed to have higher power than logic regression across several conditions. In the presence of a polygenic signal, power was generally reduced. Applying both methods on educational attainment in a large population cohort yielded numerous interacting SNPs; notably a SNP in RCAN3 which is associated with reading and spelling and a SNP in NPAS3, a neurodevelopmental gene. Availability and implementation: All methods used are implemented and freely available in R. Supplementary information: Supplementary data are available at Bioinformatics online.

Unravelling the GSK3ß-related genotypic interaction network influencing hippocampal volume in recurrent major depressive disorder.

OBJECTIVE: Glycogen synthase kinase 3ß (GSK3ß) has been implicated in mood disorders. We previously reported associations between a GSK3ß polymorphism and hippocampal volume in major depressive disorder (MDD). We then reported similar associations for a subset of GSK3ß-regulated genes. We now investigate an algorithm-derived comprehensive list of genes encoding proteins that directly interact with GSK3ß to identify a genotypic network influencing hippocampal volume in MDD. PARTICIPANTS AND METHODS: We used discovery (N=141) and replication (N=77) recurrent MDD samples. Our gene list was generated from the NetworKIN database. Hippocampal measures were derived using an optimized Freesurfer protocol. We identified interacting single nucleotide polymorphisms using the machine learning algorithm Random Forest and verified interactions using likelihood ratio tests between nested linear regression models. RESULTS: The discovery sample showed multiple two-single nucleotide polymorphism interactions with hippocampal volume. The replication sample showed a replicable interaction (likelihood ratio test: P=0.0088, replication sample; P=0.017, discovery sample; Stouffer's combined P=0.0007) between genes associated previously with endoplasmic reticulum stress, calcium regulation and histone modifications. CONCLUSION: Our results provide genetic evidence supporting associations between hippocampal volume and MDD, which may reflect underlying cellular stress responses. Our study provides evidence of biological mechanisms that should be further explored in the search for disease-modifying therapeutic targets for depression.

Enhancing Psychosis-Spectrum Nosology Through an International Data Sharing Initiative.

The latent structure of schizotypy and psychosis-spectrum symptoms remains poorly understood. Furthermore, molecular genetic substrates are poorly defined, largely due to the substantial resources required to collect rich phenotypic data across diverse populations. Sample sizes of phenotypic studies are often insufficient for advanced structural equation modeling approaches. In the last 50 years, efforts in both psychiatry and psychological science have moved toward (1) a dimensional model of psychopathology (eg, the current Hierarchical Taxonomy of Psychopathology [HiTOP] initiative), (2) an integration of methods and measures across traits and units of analysis (eg, the RDoC initiative), and (3) powerful, impactful study designs maximizing sample size to detect subtle genomic variation relating to complex traits (the Psychiatric Genomics Consortium [PGC]). These movements are important to the future study of the psychosis spectrum, and to resolving heterogeneity with respect to instrument and population. The International Consortium of Schizotypy Research is composed of over 40 laboratories in 12 countries, and to date, members have compiled a body of schizotypy- and psychosis-related phenotype data from more than 30000 individuals. It has become apparent that compiling data into a protected, relational database and crowdsourcing analytic and data science expertise will result in significant enhancement of current research on the structure and biological substrates of the psychosis spectrum. The authors present a data-sharing infrastructure similar to that of the PGC, and a resource-sharing infrastructure similar to that of HiTOP. This report details the rationale and benefits of the phenotypic data collective and presents an open invitation for participation.

Phenotypic and genetic analysis of cognitive performance in Major Depressive Disorder in the Generation Scotland: Scottish Family Health Study.

Lower performances in cognitive ability in individuals with Major Depressive Disorder (MDD) have been observed on multiple occasions. Understanding cognitive performance in MDD could provide a wider insight in the aetiology of MDD as a whole. Using a large, well characterised cohort (N = 7012), we tested for: differences in cognitive performance by MDD status and a gene (single SNP or polygenic score) by MDD interaction effect on cognitive performance. Linear regression was used to assess the association between cognitive performance and MDD status in a case-control, single-episode-recurrent MDD and control-recurrent MDD study design. Test scores on verbal declarative memory, executive functioning, vocabulary, and processing speed were examined. Cognitive performance measures showing a significant difference between groups were subsequently analysed for genetic associations. Those with recurrent MDD have lower processing speed versus controls and single-episode MDD (ß = -2.44, p = 3.6 × 10-04; ß = -2.86, p = 1.8 × 10-03, respectively). There were significantly higher vocabulary scores in MDD cases versus controls (ß = 0.79, p = 2.0 × 10-06), and for recurrent MDD versus controls (ß = 0.95, p = 5.8 × 10-05). Observed differences could not be linked to significant single-locus associations. Polygenic scores created from a processing speed meta-analysis GWAS explained 1% of variation in processing speed performance in the single-episode versus recurrent MDD study (p = 1.7 × 10-03) and 0.5% of variation in the control versus recurrent MDD study (p = 1.6 × 10-10). Individuals with recurrent MDD showed lower processing speed and executive function while showing higher vocabulary performance. Within MDD, persons with recurrent episodes show lower processing speed and executive function scores relative to individuals experiencing a single episode.

Self-reported medication use validated through record linkage to national prescribing data.

OBJECTIVES: Researchers need to be confident about the reliability of epidemiologic studies that quantify medication use through self-report. Some evidence suggests that psychiatric medications are systemically under-reported. Modern record linkage enables validation of self-report with national prescribing data as gold standard. Here, we investigated the validity of medication self-report for multiple medication types. STUDY DESIGN AND SETTING: Participants in the Generation Scotland population-based cohort (N = 10,244) recruited 2009-2011 self-reported regular usage of several commonly prescribed medication classes. This was matched against Scottish NHS prescriptions data using 3- and 6-month fixed time windows. Potential predictors of discordant self-report, including general intelligence and psychological distress, were studied via multivariable logistic regression. RESULTS: Antidepressants self-report showed very good agreement (κ = 0.85, [95% confidence interval (CI) 0.84-0.87]), comparable to antihypertensives (κ = 0.90 [CI 0.89-0.91]). Self-report of mood stabilizers showed moderate-poor agreement (κ = 0.42 [CI 0.33-0.50]). Relevant past medical history was the strongest predictor of self-report sensitivity, whereas general intelligence was not predictive. CONCLUSION: In this large population-based study, we found self-report validity varied among medication classes, with no simple relationship between psychiatric medication and under-reporting. History of indicated illness predicted more accurate self-report, for both psychiatric and nonpsychiatric medications. Although other patient-level factors influenced self-report for some medications, none predicted greater accuracy across all medications studied.

Data science for mental health: a UK perspective on a global challenge.

Data science uses computer science and statistics to extract new knowledge from high-dimensional datasets (ie, those with many different variables and data types). Mental health research, diagnosis, and treatment could benefit from data science that uses cohort studies, genomics, and routine health-care and administrative data. The UK is well placed to trial these approaches through robust NHS-linked data science projects, such as the UK Biobank, Generation Scotland, and the Clinical Record Interactive Search (CRIS) programme. Data science has great potential as a low-cost, high-return catalyst for improved mental health recognition, understanding, support, and outcomes. Lessons learnt from such studies could have global implications.

An examination of the language construct in NIMH's research domain criteria: Time for reconceptualization!

The National Institute of Mental Health's Research Domain Criteria (RDoC) Initiative "calls for the development of new ways of classifying psychopathology based on dimensions of observable behavior." As a result of this ambitious initiative, language has been identified as an independent construct in the RDoC matrix. In this article, we frame language within an evolutionary and neuropsychological context and discuss some of the limitations to the current measurements of language. Findings from genomics and the neuroimaging of performance during language tasks are discussed in relation to serious mental illness and within the context of caveats regarding measuring language. Indeed, the data collection and analysis methods employed to assay language have been both aided and constrained by the available technologies, methodologies, and conceptual definitions. Consequently, different fields of language research show inconsistent definitions of language that have become increasingly broad over time. Individually, they have also shown significant improvements in conceptual resolution, as well as in experimental and analytic techniques. More recently, language research has embraced collaborations across disciplines, notably neuroscience, cognitive science, and computational linguistics and has ultimately re-defined classical ideas of language. As we move forward, the new models of language with their remarkably multifaceted constructs force a re-examination of the NIMH RDoC conceptualization of language and thus the neuroscience and genetics underlying this concept. © 2016 The Authors. American Journal of Medical Genetics Part B: Neuropsychiatric Genetics Published by Wiley Periodicals, Inc.

Variability in working memory performance explained by epistasis vs polygenic scores in the ZNF804A pathway.

IMPORTANCE: We investigated the variation in neuropsychological function explained by risk alleles at the psychosis susceptibility gene ZNF804A and its interacting partners using single nucleotide polymorphisms (SNPs), polygenic scores, and epistatic analyses. Of particular importance was the relative contribution of the polygenic score vs epistasis in variation explained. OBJECTIVES: To (1) assess the association between SNPs in ZNF804A and the ZNF804A polygenic score with measures of cognition in cases with psychosis and (2) assess whether epistasis within the ZNF804A pathway could explain additional variation above and beyond that explained by the polygenic score. DESIGN, SETTING, AND PARTICIPANTS: Patients with psychosis (n = 424) were assessed in areas of cognitive ability impaired in schizophrenia including IQ, memory, attention, and social cognition. We used the Psychiatric GWAS Consortium 1 schizophrenia genome-wide association study to calculate a polygenic score based on identified risk variants within this genetic pathway. Cognitive measures significantly associated with the polygenic score were tested for an epistatic component using a training set (n = 170), which was used to develop linear regression models containing the polygenic score and 2-SNP interactions. The best-fitting models were tested for replication in 2 independent test sets of cases: (1) 170 individuals with schizophrenia or schizoaffective disorder and (2) 84 patients with broad psychosis (including bipolar disorder, major depressive disorder, and other psychosis). MAIN OUTCOMES AND MEASURES: Participants completed a neuropsychological assessment battery designed to target the cognitive deficits of schizophrenia including general cognitive function, episodic memory, working memory, attentional control, and social cognition. RESULTS: Higher polygenic scores were associated with poorer performance among patients on IQ, memory, and social cognition, explaining 1% to 3% of variation on these scores (range, P = .01 to .03). Using a narrow psychosis training set and independent test sets of narrow phenotype psychosis (schizophrenia and schizoaffective disorder), broad psychosis, and control participants (n = 89), the addition of 2 interaction terms containing 2 SNPs each increased the R2 for spatial working memory strategy in the independent psychosis test sets from 1.2% using the polygenic score only to 4.8% (P = .11 and .001, respectively) but did not explain additional variation in control participants. CONCLUSIONS AND RELEVANCE: These data support a role for the ZNF804A pathway in IQ, memory, and social cognition in cases. Furthermore, we showed that epistasis increases the variation explained above the contribution of the polygenic score.

Category fluency, latent semantic analysis and schizophrenia: a candidate gene approach.

BACKGROUND: Category fluency is a widely used task that relies on multiple neurocognitive processes and is a sensitive assay of cortical dysfunction, including in schizophrenia. The test requires naming of as many words belonging to a certain category (e.g., animals) as possible within a short period of time. The core metrics are the overall number of words produced and the number of errors, namely non-members generated for a target category. We combine a computational linguistic approach with a candidate gene approach to examine the genetic architecture of this traditional fluency measure. METHODS: In addition to the standard metric of overall word count, we applied a computational approach to semantics, Latent Semantic Analysis (LSA), to analyse the clustering pattern of the categories generated, as it likely reflects the search in memory for meanings. Also, since fluency performance probably also recruits verbal learning and recall processes, we included two standard measures of this cognitive process: the Wechsler Memory Scale and California Verbal Learning Test (CVLT). To explore the genetic architecture of traditional and LSA-derived fluency measures we employed a candidate gene approach focused on SNPs with known function that were available from a recent genome-wide association study (GWAS) of schizophrenia. The selected candidate genes were associated with language and speech, verbal learning and recall processes, and processing speed. A total of 39 coding SNPs were included for analysis in 665 subjects. RESULTS AND DISCUSSION: Given the modest sample size, the results should be regarded as exploratory and preliminary. Nevertheless, the data clearly illustrate how extracting the meaning from participants' responses, by analysing the actual content of words, generates useful and neurocognitively viable metrics. We discuss three replicated SNPs in the genes ZNF804A, DISC1 and KIAA0319, as well as the potential for computational analyses of linguistic and textual data in other genomics tasks.

People of the British Isles: preliminary analysis of genotypes and surnames in a UK-control population.

There is a great deal of interest in a fine-scale population structure in the UK, both as a signature of historical immigration events and because of the effect population structure may have on disease association studies. Although population structure appears to have a minor impact on the current generation of genome-wide association studies, it is likely to have a significant part in the next generation of studies designed to search for rare variants. A powerful way of detecting such structure is to control and document carefully the provenance of the samples involved. In this study, we describe the collection of a cohort of rural UK samples (The People of the British Isles), aimed at providing a well-characterised UK-control population that can be used as a resource by the research community, as well as providing a fine-scale genetic information on the British population. So far, some 4000 samples have been collected, the majority of which fit the criteria of coming from a rural area and having all four grandparents from approximately the same area. Analysis of the first 3865 samples that have been geocoded indicates that 75% have a mean distance between grandparental places of birth of 37.3 km, and that about 70% of grandparental places of birth can be classed as rural. Preliminary genotyping of 1057 samples demonstrates the value of these samples for investigating a fine-scale population structure within the UK, and shows how this can be enhanced by the use of surnames.

Letter to the editor: on the stability and ranking of predictors from random forest variable importance measures.

A recent study examined the stability of rankings from random forests using two variable importance measures (mean decrease accuracy (MDA) and mean decrease Gini (MDG)) and concluded that rankings based on the MDG were more robust than MDA. However, studies examining data-specific characteristics on ranking stability have been few. Rankings based on the MDG measure showed sensitivity to within-predictor correlation and differences in category frequencies, even when the number of categories was held constant, and thus may produce spurious results. The MDA measure was robust to these data characteristics. Further, under strong within-predictor correlation, MDG rankings were less stable than those using MDA.

Biological validation of increased schizophrenia risk with NRG1, ERBB4, and AKT1 epistasis via functional neuroimaging in healthy controls.

CONTEXT: NRG1 is a schizophrenia candidate gene and plays an important role in brain development and neural function. Schizophrenia is a complex disorder, with etiology likely due to epistasis. OBJECTIVE: To examine epistasis between NRG1 and selected N-methyl-d-aspartate-glutamate pathway partners implicated in its effects, including ERBB4, AKT1, DLG4, NOS1, and NOS1AP. DESIGN: Schizophrenia case-control sample analyzed using machine learning algorithms and logistic regression with follow-up using neuroimaging on an independent sample of healthy controls. PARTICIPANTS: A referred sample of schizophrenic patients (n = 296) meeting DSM-IV criteria for schizophrenia spectrum disorder and a volunteer sample of controls for case-control comparison (n = 365) and a separate volunteer sample of controls for neuroimaging (n = 172). MAIN OUTCOME MEASURES: Epistatic association between single-nucleotide polymorphisms (SNPs) and case-control status; epistatic association between SNPs and the blood oxygen level-dependent physiological response during working memory measured by functional magnetic resonance imaging. RESULTS: We observed interaction between NRG1 5' and 3' SNPs rs4560751 and rs3802160 (likelihood ratio test P = .00020) and schizophrenia, which was validated using functional magnetic resonance imaging of working memory in healthy controls; carriers of risk-associated genotypes showed inefficient processing in the dorsolateral prefrontal cortex (P = .015, familywise error corrected). We observed epistasis between NRG1 (rs10503929; Thr286/289/294Met) and its receptor ERBB4 (rs1026882; likelihood ratio test P = .035); a 3-way interaction with these 2 SNPs and AKT1 (rs2494734) was also observed (odds ratio, 27.13; 95% confidence interval, 3.30-223.03; likelihood ratio test P = .042). These same 2- and 3-way interactions were further biologically validated via functional magnetic resonance imaging: healthy individuals carrying risk genotypes for NRG1 and ERBB4, or these 2 together with AKT1, were disproportionately less efficient in dorsolateral prefrontal cortex processing. Lower-level interactions were not observed between NRG1 /ERBB4 and AKT1 in association or neuroimaging, consistent with biological evidence that NRG1 × ERBB4 interaction modulates downstream AKT1 signaling. CONCLUSION: Our data suggest complex epistatic effects implicating an NRG1 molecular pathway in cognitive brain function and the pathogenesis of schizophrenia.

Linkage disequilibrium and age of HLA region SNPs in relation to classic HLA gene alleles within Europe.

The HLA region on chromosome 6 is gene-rich and under selective pressure because of the high proportion of immunity-related genes. Linkage disequilibrium (LD) patterns and allele frequencies in this region are highly differentiated across broad geographical populations, making it a region of interest for population genetics and immunity-related disease studies. We examined LD in this important region of the genome in six European populations using 166 putatively neutral SNPs and the classical HLA-A, -B and -C gene alleles. We found that the pattern of association between classic HLA gene alleles and SNPs implied that most of the SNPs predated the origin of classic HLA gene alleles. The SNPs most strongly associated with HLA gene alleles were in some cases highly predictive of the HLA allele carrier status (misclassification rates ranged from <1 to 27%) in independent populations using five or fewer SNPs, a much smaller number than tagSNP panels previously proposed and often with similar accuracy, showing that our approach may be a viable solution to designing new HLA prediction panels. To describe the LD within this region, we developed a new haplotype clustering method/software based on r(2), which may be more appropriate for use within regions of strong LD. Haplotype blocks created using this proposed method, as well as classic HLA gene alleles and SNPs, were predictive of a northern versus southern European population membership (misclassification error rates ranged from 0 to 23%, depending on which independent population was used for prediction), indicating that this region may be a rich source of ancestry informative markers.

The behaviour of random forest permutation-based variable importance measures under predictor correlation.

BACKGROUND: Random forests (RF) have been increasingly used in applications such as genome-wide association and microarray studies where predictor correlation is frequently observed. Recent works on permutation-based variable importance measures (VIMs) used in RF have come to apparently contradictory conclusions. We present an extended simulation study to synthesize results. RESULTS: In the case when both predictor correlation was present and predictors were associated with the outcome (HA), the unconditional RF VIM attributed a higher share of importance to correlated predictors, while under the null hypothesis that no predictors are associated with the outcome (H0) the unconditional RF VIM was unbiased. Conditional VIMs showed a decrease in VIM values for correlated predictors versus the unconditional VIMs under HA and was unbiased under H0. Scaled VIMs were clearly biased under HA and H0. CONCLUSIONS: Unconditional unscaled VIMs are a computationally tractable choice for large datasets and are unbiased under the null hypothesis. Whether the observed increased VIMs for correlated predictors may be considered a "bias" - because they do not directly reflect the coefficients in the generating model - or if it is a beneficial attribute of these VIMs is dependent on the application. For example, in genetic association studies, where correlation between markers may help to localize the functionally relevant variant, the increased importance of correlated predictors may be an advantage. On the other hand, we show examples where this increased importance may result in spurious signals.