Naslov
Comparing parametric and non-parametric Bayesian approaches for genetic prediction of complex traits

Autori
Pleić, Nikolina ; Gunjača, Ivana ; Babić Leko, Mirjana, Polašek, Ozren ; Zemunik, Tatijana

Vrsta, podvrsta i kategorija rada
Sažeci sa skupova, sažetak, znanstveni

Izvornik
Book of Abstracts of the ISCCRO – International Statistical Conference in Croatia, Volume 4, No. 1, 2022 / Berislav Žmuk, Anita Čeh Časni - Zagreb : Hrvatsko statističko društvo, 2022, 12-12

Skup
4th International Statistical Conference in Croatia (ISCCRO 2022)

Mjesto i datum
Opatija, Hrvatska, 05.05.2022. - 06.05.2022

Vrsta sudjelovanja
Predavanje

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
bayesian genomic prediction ; comparison ; polygenic score ; thyroid-stimulating hormone

Sažetak
Predicting phenotypes from genotypes can greatly aid the development of precision medicine by using genetic markers to personalize the treatment and predict the outcome. However, predictions require the development of statistical methods that can accurately model the polygenic architecture of the studied trait. This is achieved by constructing a polygenic score (PGS). The simplest PGS is essentially a weighted sum of genotypes across single nucleotide polymorphisms (SNPs), where weights are the estimated genetic effect sizes. We decided to utilize two Bayesian approaches for genomic prediction of thyroid-stimulating hormone (TSH) levels measured in 2, 566 individuals belonging to the Korčula cohort of the Croatian biobank “10, 001 Dalmatians”. Firstly, we used the parametric Bayesian sparse linear-mixed model (BSLMM), which is a hybrid of a linear mixed model and a sparse regression model. Secondly, we used a latent Dirichlet process regression (DPR), a Bayesian non-parametric model. BSLMM assumes a mixture of two normal distributions for the prior of the SNP effect sizes, while DPR does not assume any fixed parametric distribution for the effect size distribution. We performed 10 Monte Carlo cross-validation data splits. In each data split, we fitted models in a training set with 80% of individuals and evaluated method performance using R2 or MSE in a test set with the remaining 20% of individuals. We compared the performance of the two methods by taking the R2 difference or MSE difference with respect to DPR. These results will provide additional insights into the modeling of the polygenic architecture of complex traits.

Izvorni jezik
Engleski

Znanstvena područja
Matematika, Biologija, Interdisciplinarne biotehničke znanosti, Biotehnologija u biomedicini (prirodno područje, biomedicina i zdravstvo, biotehničko područje)

POVEZANOST RADA