A decision tree to improve identification of pathogenic mutations in clinical practice

Abstract

Background: A variant of unknown significance (VUS) is a variant form of a gene that has been identified through genetic testing, but whose significance to the organism function is not known. An actual challenge in precision medicine is to precisely identify which detected mutations from a sequencing process have a suitable role in the treatment or diagnosis of a disease. The average accuracy of pathogenicity predictors is 85%. However, there is a significant discordance about the identification of mutational impact and pathogenicity among them. Therefore, manual verification is necessary for confirming the real effect of a mutation in its casuistic. Methods: In this work, we use variables categorization and selection for building a decision tree model, and later we measure and compare its accuracy with four known mutation predictors and seventeen supervised machinelearning (ML) algorithms. Results: The results showed that the proposed tree reached the highest precision among all tested variables: 91% for True Neutrals, 8% for False Neutrals, 9% for False Pathogenic, and 92% for True Pathogenic. Conclusions: The decision tree exceptionally demonstrated high classification precision with cancer data, producing consistently relevant forecasts for the sample tests with an accuracy close to the best ones achieved from supervised ML algorithms. Besides, the decision tree algorithm is easier to apply in clinical practice by non-IT experts. From the cancer research community perspective, this approach can be successfully applied as an alternative for the determination of potential pathogenicity of VOUS.

Germline Compound Heterozygous Variants Identified in the STXBP2 Gene Leading to a Familial Hemophagocytic Lymphohistiocytosis Type 5: A Case Report

“…Pathogenicity prediction algorithms are controversial for the impact of this missense change (SIFT: “Deleterious”; PolyPhen-2: “Benign”; Align-GVGD: “Class C0”). We have applied a decision tree based on a machine learning technique for data classification, and the algorithm predicted the VUS as pathogenic or damaging ( 15 ).…”

AmazonForest: In-silico Meta-Prediction of Pathogenic Variants

“…Some previously proposed meta-prediction approaches were proposed each containing different machine learning or statistical methods, and differ in training datasets [9][10][11]13]. Futher, most of the reviewed meta-predictors used decision tree methods, [9,11,13]. Decision tree-based models deal with categorical predictors without the need to transform them [32].…”

“…To the best of our knowledge, there are no similar works that offer or provide tools for user interaction with the user. [9][10][11]13]. Thus, here we have provided an online tool for a better understanding of the training process and variants reclassification.…”

“…Distinct machine learning (ML) meta-prediction models have been proposed for pathogenicity prediction of new genetic variants. Generally, each meta-prediction models have been suggested for the analysis of a single variant class (synonymous or non-synonymous variants) [9][10][11][12][13] and most meta-predictors were used for the pathogenicity prediction of VUS and CI variants [9,10,12,13]. Interestingly,the majority of the recently proposed meta-predictors are based on decision trees or ensembles of decision trees, which constitute models with clear interpretations.…”

“…Interestingly,the majority of the recently proposed meta-predictors are based on decision trees or ensembles of decision trees, which constitute models with clear interpretations. Ensemble-based methods, such as Random Forest, are promising for pathogenicity prediction of coding and non-coding variants [9][10][11]13]. However, these models have shown differences regarding data training methods, specifically on the number of features used to build each classification models.…”