The Complexity of Linkage Analysis with Neural Networks

Abstract

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.

Linkage Disequilibrium in Genetic Association Studies Improves the Performance of Grammatical Evolution Neural Networks

“…The NN approach is a commonly used pattern recognition technique for data mining. NNs have been successful in a variety of fields, though they have met with mixed success in genetic epidemiology [9][10][11][12][13][14]. Unsuccessful applications may be attributed to the paramount importance of choosing the correct NN architecture for each individual dataset.…”

Joint Tests of Linkage and Association for Quantitative Traits

“…There is increasing recognition of the need for and benefits of modeling the joint effects, both additive and epistatic, of multiple loci (Frankel and Schork, 1996;Goldgar and Easton, 1997;Marinov and Weeks, 2001). In most TDTs for quantitative traits, multiple loci can be easily accommodated.…”

Comparison of approaches for machine‐learning optimization of neural networks for detecting gene‐gene interactions in genetic epidemiology

“…Also, they are model free in that no assumption has to be made about the genetic architecture that results in a particular phenotype. Many groups, including our own, have applied NN to genetic data, with varying results [Lucek et al, 1998;Lucek and Ott, 1997;Marinov and Weeks, 2001; Motsinger et al, 2006a,d; Ritchie et al, 2003b. For an in-depth review of NN applications in genetic epidemiology, see Motsinger et al [2008].…”

“…Many groups, including our own, have applied NN to genetic data, with varying results [Lucek et al, 1998;Lucek and Ott, 1997;Marinov and Weeks, 2001; Motsinger et al, 2006a,d; Ritchie et al, 2003b. For an in-depth review of NN applications in genetic epidemiology, see Motsinger et al [2008].NNs have been used successfully for pattern recognition in many fields, including medical research [Ripley, 1996]; however, the results in genetic epidemiology have not been as consistent [Curtis et al, 2001;Lucek et al, 1998;Lucek and Ott, 1997;Marinov and Weeks, 2001;North et al, 2003; Saccone et al, 1999]. One possible reason for this inconsistency is the fact that training a standard back-propagation NN minimizes the mean-squared error and in a complex fitness landscape (such as the one expected for complex human diseases) [Moore and Parker, 2001] there may be several minima.…”

“…Because of the non-linear properties of the NN activation functions, it is easy to stall in a local minimum, rather than finding the global minimum. Therefore, on any given run of a traditional NN using a back-propagation learning algorithm, the NN may stall on a different local minimum for each subsequent run of the network [Marinov and Weeks, 2001]. Another possible explanation for the general lack of success in many previous NN applications is the use of less than optimal NN architecture.…”