MACHINE LEARNING-BASED PATHOGENICITY PREDICTION AND PRIORITIZATION OF HBB GENE VARIANTS USING CLINVAR AND POPULATION-SCALE GENOMIC DATA

Abstract

Background:

The advent of advanced techniques in genomic sequencing has led to the discovery of several genetic variants, where most of them are still uncharacterized. The determination of the pathogenicity of these genetic variants is crucial in understanding the genetic basis of diseases such as β-thalassemia, caused by mutations in the HBB gene.

Objective:

The study aims to develop a machine learning-based model for pathogenicity prediction and prioritizing novel variations in the HBB gene.

Methodology:

A dataset of 7,819 ClinVar variants, was compiled and labelled as benign or pathogenic. Machine learning models such as Random Forest and XGBoost were implemented and evaluated using techniques for handling class imbalance. The models were evaluated using metrics such as accuracy, ROC-AUC, and precision-recall AUC (PR-AUC). The model that outperformed was then applied to the dataset of 36,099 variants in the HBB gene region.

Result:

The XGBoost model was better than the Random Forest model in the prediction of the pathogenicity of the genetic variants. The ROC-AUC and PR-AUC were 0.81 and 0.61, respectively. The model showed better precision in the prediction of the pathogenicity of the genetic variants, improving from 0.35 to 0.49 for pathogenic variations. XGBoost was further applied to the dataset of the HBB gene and showed a right-skewed distribution of the probabilities. The model showed that the highly confident predictions (>0.97) were in the highly conserved regions (>7).

Conclusion:

The proposed model is effective in the prediction of the pathogenicity of the genetic variants and the prioritization of novel variations in the HBB gene. The model is applicable in the future in the field of genomic medicine and in the study of diseases.