ClinVar is a web platform that stores ∼789,000 genetic associations with complex diseases. A partial set of these cataloged genetic associations has challenged clinicians and geneticists, often leading to conflicting interpretations or uncertain clinical impact significance. In this study, we addressed the (re)classification of genetic variants by AmazonForest, which is a random-forest-based pathogenicity metaprediction model that works by combining functional impact data from eight prediction tools. We evaluated the performance of representation learning algorithms such as autoencoders to propose a better strategy. All metaprediction models were trained with ClinVar data, and genetic variants were annotated with eight functional impact predictors cataloged with SnpEff/SnpSift. AmazonForest implements the best random forest model with a one hot data-encoding strategy, which shows an Area Under ROC Curve of ≥0.93. AmazonForest was employed for pathogenicity prediction of a set of ∼101,000 genetic variants of uncertain significance or conflict of interpretation. Our findings revealed ∼24,000 variants with high pathogenic probability ($R{F}_{prob}\ge 0.9$). In addition, we show results for Alzheimer’s Disease as a demonstration of its application in clinical interpretation of genetic variants in complex diseases. Lastly, AmazonForest is available as a web tool and R object that can be loaded to perform pathogenicity predictions.

ClinVar is a web platform that stores around 774k curated entries, which allows exploring genetic variants and their associations with complex phenotypes. A partial set of ClinVar’s genetic associations were reported with conflict of interpretation or uncertain clinical impact significance, which currently challenges clinicians and geneticists. Here, we evaluate the performance of data pre-processing methods combined with classical prediction methods, such as Naive Bayes, Random Forest, and Support Vector Machine to build a meta-prediction model aiming to improve genetic pathogenicity interpretation. Models were trained with ClinVar data (September 2020), and genetic variants were annotated with eight functional impact predictors catalogued with SnpEff/SnpSift (v4.3). A 10-fold cross-validation strategy was performed for evaluation by accuracy, F1-Score, Receiver Operating Characteristic, Area Under Curve. The best meta-prediction model raises by combining one-hot encoding with tree-based classifiers as Random Forest, which shows Area Under Curve ≥ 0,93. We predict pathogenicity for 109k genetic variants, which were found labeled as uncertain significance or conflict of interpretation. Additionally, we implemented AmazonForest (https://www.lghm.ufpa.br/amazonforest), a web tool to query data for a set of 5k variants that were predicted with high pathogenic probability (RFprob >= 0.9).