Determining the origin of different variants associated with familial mediterranean fever by machine-learning

Abstract

A growing number of familial Mediterranean fever (FMF) patients in Israel do not have a single country of origin for all four grandparents. We aimed to predict the Mediterranean fever gene (MEFV) variant most likely to be found for an individual FMF patient, by a machine learning approach. This study was conducted at the Sheba Medical Center, a referral center for FMF in Israel. All Jewish referrals included in this study carried an FMF associated variant in MEFV as shown by genetic testing performed between 2001 and 2017. We introduced the term 'origin score' to capture the dose and different combinations of the grandparents' origin. A machine learning approach was used to analyze the data. In a total of 1781 referrals included in this study, the p.Met694Val variant was the most common, and the variants p.Glu148Gln and p.Val726Ala second and third most common, respectively. Of 26 countries of origin analyzed, those that increased the likelihood of a referral to carry specific variants were identified in North Africa for p.Met694Val, Europe for p.Val726Ala, and west Asia for p.Glu148Gln. Fourteen of the studied countries did not show a highly probable variant. Based on our results, it is possible to describe an association between modern day origins of the three most common MEFV variant types and a geographical region. A strong geographic association could arise from positive selection of a specific MEFV variant conferring resistance to endemic infectious agents.

Figure 1

Prevalence of the countries of origin most indicative for the p.Met694Val variant: (A) Tunisia, Libya, Morocco, and Algeria are the countries of origin most positively associated with the p.Met694Val variant. (B) Romania, Germany, Iran and Poland are the countries of origin most negatively associated with the p.Met694Val variant. Red and blue bars indicate patients who carry and do not carry the p.Met694Val variant, respectively. The x-axis indicates the amount of grandparents from the four countries (0 being none, 1 being all four).

Figure 2

ROC of multivariate logistic regression variant prediction. The figure represents the performance of the multivariate logistic regression model that uses 27 features (26 countries and sex of the patient) to predict whether a patient: (A) carries the variant p.Met694Val, (B) carries the variant p.Val726Ala, (C) carries the variant p.Glu148Gln, and (D) is homozygous for p.Met694Val.

Figure 3

Prevalence of the countries of origin most indicative for the p.Val726Ala variant: (A) Lebanon, Romania, Hungary, and Poland are the countries of origin most positively associated with the p.Val726Ala variant. (B) Morocco, Libya, Tunisia, and Algeria are the countries of origin most negatively associated with the p.Val726Ala variant. Red and blue bars indicate patients who carry and do not carry the p.Val726Ala variant, respectively. The x-axis indicates the amount of grandparents from the four countries (0 being none, 1 being all four).

Figure 4

Prevalence of the countries of origin most indicative for the p.Glu148Gln variant: Countries that were found to be the most positively (A) and negatively (B) associated origins for a patient who carries a positive p.Glu148Gln variant. Red and blue bars represent the patients who carry and do not carry the p.Glu148Gln variant, respectively. The X-axis indicates the amount of grandparents from the four countries (0 being none, 1 being all four).