Analysis of complex traits using neural networks

Objectives: Some traits, while naturally polychotomous, are routinely dichotomized for genetic analysis. Dichotomization, intuitively, leads to a loss of power to detect linkage, as some phenotypic variability is discarded. This paper examines this power loss in the context of a trichotomous trait. Methods: To examine this power loss, we performed a simulation study where a trichotomous trait was simulated in a sample of 1,000 sib-pairs under various genetic models. The study was replicated 1,000 times. Linkage analysis using a variance components method, as implemented in Mx, was then performed on the trichotomous trait and compared with that on a dichotomized version of the trait. Results: A comparison of the power and false positive rates of the analyses shows that power to detect linkage was increased by up to 22 percentage points simply by examining the trait as a trichotomy instead of a dichotomy. Under all models examined, the trichotomous analysis outperformed the dichotomous version. Conclusions: Comparable levels of false positive rates under both methods confirm that this power gain comes solely from the information lost upon dichotomization. Thus, dichotomizing tri- or poly-chotomous traits can lead to crippling power loss, especially in the case of many loci of small effect.

Section: Discussionmentioning

confidence: 99%

Power Loss for Linkage Analysis due to the Dichotomization of Trichotomous Phenotypes

Corbett

Gu²,

Rice³

et al. 2004

“…This approach was used by Bhat et al [11], where they looked for consistent results across 10 different runs. However, both the results of Bhat et al and our results indicate that it is not necessarily the case that one sees consistently elevated regions from run to run.…”

Section: Discussionmentioning

confidence: 99%

“…As others have pointed out [11,15,16], the definition of the contribution value used here is ad hoc, as there are many different ways to measure the importance of a given input node [17], all of which have some drawbacks. However, CV definition is not directly relevant to lack of repeatability, since the CVs are computed after the weight estimation process is completed.…”

Section: Discussionmentioning

confidence: 99%

The Complexity of Linkage Analysis with Neural Networks

Marinov

Weeks²

2001

As the focus of genome-wide scans for disease loci have shifted from simple Mendelian traits to genetically complex traits, researchers have begun to consider new alternative ways to detect linkage that will consider more than the marginal effects of a single disease locus at a time. One interesting new method is to train a neural network on a genome-wide data set in order to search for the best non-linear relationship between identity-by-descent sharing among affected siblings at markers and their disease status. We investigate here the repeatability of the neural network results from run to run, and show that the results obtained by multiple runs of the neural network method may differ quite a bit. This is most likely due to the fact that training a neural network involves minimizing an error function with a multitude of local minima.

“…For example, methods based on classifi cation are MDR [9] , neural networks [10,11] , decision trees [12] and discriminant analysis [13] , which are designed to uncover the complex relationship without relying on a specifi c genetic mode. Methods generalized from statistical tests are CPM [4] , a set-association method integrating allelic association and Hardy-Weinberg [8] , and the Generalized T 2 Test [14] .…”

Section: Introductionmentioning

confidence: 99%

Multi-Locus Penetrance Variance Analysis Method for Association Study in Complex Diseases

Sun

Zhang²,

Zhang³

et al. 2005

Common heritable diseases often result from the action of several different genes, each of which contributes to the total observed variability in the disease trait. Traditional single-locus association approaches rely heavily on the marginal effects of single-locus and tend to ignore the multigenic nature of complex diseases. The increasing request for localizing genes underlying traits in multi-gene diseases has led to the development of some statistical methods. In this study, we develop a multi-locus analysis method – multi-locus penetrance variance analysis (MPVA), and conduct systematical simulation studies to evaluate its performance. Our results show that compared with other multi-locus methods, MPVA has some advantage in detecting complicated interactions under different epistatic models, and its performance is stable and robust.