A Survey of Variant Effect Prediction Methods and a Proposed Framework for Pathogenicity Prediction

Abstract

Everything in life is coded in DNA and hence the distinction between pathogenic and benign variants has been a critical task in clinical genetics. This survey discusses how variant effect prediction methods, which began as heuristics and machine learning methods, evolved into deep learning, disease specific, large language model-based methods. Initial genomics studies used heuristic methods that used algorithms such as SIFT to assist in the distinction between mutations and phenotypic changes by predicting the possibility of an amino acid substitution to be detrimental. Further advancements led to unsupervised learning, in which Multiple Sequence Alignments were used to train generative models of natural sequence variation within each gene. Disease specific and context aware models have also been developed more recently which are trained on variants of a particular disease or tissue; this leads to higher clinical applicability and interpretability. Recently there has been an emergence of deep learning methods, which learn hierarchical properties of raw data without requiring labeled data, and are capable of predicting variant effects more accurately. We have classified the existing approaches into rule-based, supervised, unsupervised, deep learning, disease-specific and large language model approaches which are analysed in terms of their methodologies, usability and accessibility. None of these methods, however, provide with the reference databases like ClinVar that is a necessary tool and plays a significant role in clinical research and decision making. This study proposes an integrated system constituting a modern transformer-based model such as Evo2 and a reference database like ClinVar for better reliability and decision making.