The Random Forests (RF) algorithm has become a commonly used machine learning algorithm for genetic association studies. It is well suited for genetic applications since it is both computationally efficient and models genetic causal mechanisms well. With its growing ubiquity, there has been inconsistent and less than optimal use of RF in the literature. The purpose of this review is to breakdown the theoretical and statistical basis of RF so that practitioners are able to apply it in their work. An emphasis is placed on showing how the various components contribute to bias and variance, as well as discussing variable importance measures. Applications specific to genetic studies are highlighted. To provide context, RF is compared to other commonly used machine learning algorithms.

Variation in major histocompatibility complex genes on chromosome 6p21.3, specifically the human leukocyte antigen HLA-DR2 or DRB1*1501-DQB1*0602 extended haplotype, confers risk for multiple sclerosis (MS). Previous studies of DRB1 variation and both MS susceptibility and phenotypic expression have lacked statistical power to detect modest genotypic influences, and have demonstrated conflicting results. Results derived from analyses of 1339 MS families indicate DRB1 variation influences MS susceptibility in a complex manner. DRB1*15 was strongly associated in families (P=7.8×10⁻³¹), and a dominant DRB1*15 dose effect was confirmed (OR=7.5, 95% CI=4.4–13.0, P<0.0001). A modest dose effect was also detected for DRB1*03; however, in contrast to DRB1*15, this risk was recessive (OR=1.8, 95% CI=1.1–2.9, P=0.03). Strong evidence for under-transmission of DRB1*14 (P=5.7×10⁻⁶) even after accounting for DRB1*15 (P=0.03) was present, confirming a protective effect. In addition, a high risk DRB1*15 genotype bearing DRB1*08 was identified (OR=7.7, 95% CI=4.1–14.4, P<0.0001), providing additional evidence for transDRB1 allelic interactions in MS. Further, a significant DRB1*15 association observed in primary progressive MS families (P=0.0004), similar to relapsing-remitting MS families, suggests that DRB1-related mechanisms are contributing to both phenotypes. In contrast, results obtained from 2201 MS cases argue convincingly that DRB1*15 genotypes do not modulate age of onset, or significantly influence disease severity measured using expanded disease disability score and disease duration. These results contribute substantially to our understanding of the DRB1 locus and MS, and underscore the importance of using large sample sizes to detect modest genetic effects, particularly in studies of genotype–phenotype relationships.

In monkeys performing a visual spatial attention task, stimulation of neurons in the lateral geniculate nucleus and recording of shock-evoked responses from monosynaptically connected primary-visual-cortex neurons shows that attention enhances neuronal communication by increasing the efficacy of presynaptic input, increasing synchronous responses, and by decreasing redundant signals. Attention enhances neural processing at many points in sensory systems, but how it influences interactions between different brain areas is less well understood. Martin Usrey and colleagues probe how attention alters the functional connection between lateral geniculate nucleus and primary visual cortex, two areas modulated by attention. They find that attention not only enhances the efficacy of transmission between the areas, but also increases synchronization and reduces redundancy in primary visual cortex responses, thereby potentially reducing noise in the system. Attention is a critical component of perception1. However, the mechanisms by which attention modulates neuronal communication to guide behaviour are poorly understood. To elucidate the synaptic mechanisms of attention, we developed a sensitive assay of attentional modulation of neuronal communication. In alert monkeys performing a visual spatial attention task, we probed thalamocortical communication by electrically stimulating neurons in the lateral geniculate nucleus of the thalamus while simultaneously recording shock-evoked responses from monosynaptically connected neurons in primary visual cortex. We found that attention enhances neuronal communication by increasing the efficacy of presynaptic input in driving postsynaptic responses, by increasing synchronous responses among ensembles of postsynaptic neurons receiving independent input, and by decreasing redundant signals between postsynaptic neurons receiving common input. The results demonstrate that attention finely tunes neuronal communication at the synaptic level by selectively altering synaptic weights, enabling enhanced detection of salient events in the noisy sensory environment.

Folic acid taken in early pregnancy reduces risks for delivering offspring with several congenital anomalies. The mechanism by which folic acid reduces risk is unknown. Investigations into genetic variation that influences transport and metabolism of folate will help fill this data gap. We focused on 118 SNPs involved in folate transport and metabolism. Using data from a California population-based registry, we investigated whether risks of spina bifida or conotruncal heart defects were influenced by 118 single nucleotide polymorphisms (SNPs) associated with the complex folate pathway. This case-control study included 259 infants with spina bifida and a random sample of 359 nonmalformed control infants born during 1983–86 or 1994–95. It also included 214 infants with conotruncal heart defects born during 1983–86. Infant genotyping was performed blinded to case or control status using a designed SNPlex assay. We examined single SNP effects for each of the 118 SNPs, as well as haplotypes, for each of the two outcomes. Few odds ratios (ORs) revealed sizable departures from 1.0. With respect to spina bifida, we observed ORs with 95% confidence intervals that did not include 1.0 for the following SNPs (heterozygous or homozygous) relative to the reference genotype: BHMT (rs3733890) OR = 1.8 (1.1–3.1), CBS (rs2851391) OR = 2.0 (1.2–3.1); CBS (rs234713) OR = 2.9 (1.3–6.7); MTHFD1 (rs2236224) OR = 1.7 (1.1–2.7); MTHFD1 (hcv11462908) OR = 0.2 (0–0.9); MTHFD2 (rs702465) OR = 0.6 (0.4–0.9); MTHFD2 (rs7571842) OR = 0.6 (0.4–0.9); MTHFR (rs1801133) OR = 2.0 (1.2–3.1); MTRR (rs162036) OR = 3.0 (1.5–5.9); MTRR (rs10380) OR = 3.4 (1.6–7.1); MTRR (rs1801394) OR = 0.7 (0.5–0.9); MTRR (rs9332) OR = 2.7 (1.3–5.3); TYMS (rs2847149) OR = 2.2 (1.4–3.5); TYMS (rs1001761) OR = 2.4 (1.5–3.8); and TYMS (rs502396) OR = 2.1 (1.3–3.3). However, multiple SNPs observed for a given gene showed evidence of linkage disequilibrium indicating that the observed SNPs were not individually contributing to risk. We did not observe any ORs with confidence intervals that did not include 1.0 for any of the studied SNPs with conotruncal heart defects. Haplotype reconstruction showed statistical evidence of nonrandom associations with TYMS, MTHFR, BHMT and MTR for spina bifida. Our observations do not implicate a particular folate transport or metabolism gene to be strongly associated with risks for spina bifida or conotruncal defects.