Genetic signature of differentiated thyroid carcinoma susceptibility: a machine learning approach

Abstract

To identify a peculiar genetic combination predisposing to differentiated thyroid carcinoma (DTC), we selected a set of single-nucleotide polymorphisms (SNPs) associated with DTC risk, considering polygenic risk score (PRS), Bayesian statistics, and a machine learning (ML) classifier to describe cases and controls in 3 different datasets. Dataset 1 (649 DTC, 431 controls) has been previously genotyped in a genome-wide association study (GWAS) on Italian DTC. Dataset 2 (234 DTC, 101 controls) and dataset 3 (404 DTC, 392 controls) were genotyped. Associations of 171 SNPs reported to predispose to DTC in candidate studies were extracted from the GWAS of dataset 1, followed by replication of SNPs associated with DTC risk (P<0.05) in dataset 2. The reliability of the identified SNPs was confirmed by PRS and Bayesian statistics after merging the three datasets. SNPs were used to describe the case/control state of individuals by ML classifier. Starting from 171 SNPs associated with DTC, 15 were positive in both the datasets 1 and 2. Using these markers, PRS revealed that individuals in the fifth quintile had a 7-fold increased risk of DTC than those in the first. Bayesian inference confirmed that the selected 15 SNPs differentiate cases from controls. Results were corroborated by ML, finding a maximum AUC of about 0.7. A restricted selection of only 15 DTC-associated SNPs is able to describe the inner genetic structure of Italian individuals and ML allows a fair prediction of case or control status based solely on the individual genetic background.

Genome-wide association studies for thyroid physiology and diseases

“…Future replication studies, including one with Japanese patients, will be required to draw firm conclusions. More recently, an attempt was made to quantify the risk of thyroid cancer by calculating a score based on a large number of SNP information (i.e., polygenic risk score) [44]. Although the area under the ROC curve for thyroid cancer prediction is inadequate (0.7), its performance is expected to improve as more susceptible loci are identified in the future.…”