Amino acid substitutions on protein sequences are generally harmless, but a significant proportion of them can cause diseases. Accurately predicting the effect of these genetic variants can be crucial for clinicians, it can potentially speed up the diagnosis of patients having missense variants that are likely to lead to disease. Today, a variety of computational tools have been developed to predict the pathogenicity of genetic variants using numerous methodologies. The most well-known tools are SIFT and PolyPhen, each of them accumulated more than 10,000 citations. More recently, many tools have been developed using Artificial Intelligence and other innovative approaches. It is important to evaluate and rank the performance of these different computational tools in order to guide future users and clinicians. In this study, we rigorously evaluated 50 tools using quality data and measures for each computational method. In addition, we carried out a detailed analysis of the available data on genetic variants to highlight a problem inherent in public databases: the prediction quality is significantly impacted by the different variant datasets. Our results show that variants from ClinVar appear to be easy to predict, whereas variants from other data sources are more difficult to predict. We show that the predictability of variants can be divided into two distinct categories: (i) Easy and (ii) Difficult to predict. We have therefore developed a neural network model capable of classifying variants into these categories and tested the model on cancer datasets to demonstrate its potential use.