Modeling the contribution of genetic variation to cognitive gains following training with artificial neural networks

CASCALLAR, E. C.; MUSSO, M. F.; RUEDA, M. R.; CÓMBITA, L. M.

Conferencia; 29th Annual Convention Boston 2017; 2017

Association for Psychological Sciences (APS)

Introduction: Evidence in the field of cognitive training has shown that certain individual differences can modulate the extent to which children benefit from training. The main goal of the present study was to develop predictive models of the gains in working memory (WM) and fluid intelligence (IQ-f) after an executive attention training program for children. In addition, this study aims to understand the contribution of specific genetic markers and to compare particular patterns predicting gains in both cognitive functions.Method: A sample of 66 children (56% males), ages between 50.9 and 75.9 months (M= 63.41; SD= 7.47) participated in a computerized training program during four weeks (10 sessions; 45 minutes per session). The program consisted of 14 computerized exercises divided into 6 categories: (1)Tracking/Anticipatory; (2)Attention Focusing/Discrimination; (3)Conflict Monitoring/Resolution; (4) Inhibitory control; (5)Task Switching; and (6)Sustained Attention. Variation in genes involved in the regulation of dopamine (COMT, DAT1, DRD4), serotonin (5HTT), norepinephrine (MAOA3), acetylcholine (CHRNA4), and others brain factors (SNAP, DBH) was collected, together with several behavioral, cognitive and brain function measures. The present study focuses on the prediction of gains in cognitive dependent measures by analyzing the patterns of genetic markers, gender and age variables. The standardized pre- to post-training gains of two dependent measures were considered: Working Memory Span Subtest of the WISC, backwards condition (Wechsler, 1991), and the fluid intelligence- sub-scale of the K-BIT (Kaufman & Kaufman, 1990).Analysis procedure: The approach followed in this research was a machine-learning methodology, utilizing multilayer perceptron artificial neural networks (ANN), with a backpropagation algorithm to develop precise models of the contribution of all independent variables, for each dependent measure. ANN have been shown to be very effective to study problems consisting of a large number of variables in complex, non-linear, and poorly understood interactions. In addition to being powerful classifiers, they build plausible architectures to explore the participation of variables involved in the modeling of a problem.Results: Both ANN reached high overall accuracy (90-100%) in the predictive classification of two levels of standardized gains (?Zero or less? and ?Higher than zero? in IQ-f and WM). Both networks showed interesting differences in the pattern of predictive weights (importance) of particular genetic markers showing the highest participation in the predictive model. Age, presence or absence of the 3r allele of the monoamine oxidase A (MAOA), the SNP rs6269 and a haplotype including the SNPs rs6269, rs4633, rs4818, rs4680 of the catechol O-methyltransferase (COMT) gene were particularly informative in the modeling of the pre- to post-training change in WM. COMT haplotype Group, age, gender and the presence or absence of 10r allele of the dopamine transporter (DAT1) gene were the top four predictors with the most significant importance in modeling IQ-f change. Findings suggest that among other individual differences, particular variations in genes involved in dopamine and norepinephrine neurotransmission affect children?s susceptibility to benefit from executive attention training, a pattern that is consistent with previous studies. The study also shows the usefulness of a machine-learning approach for studies in this area of research.