MutFormer: A context-dependent transformer-based model to predict pathogenic missense mutations
Year: 2021
Type: preprint
Abstract: A missense mutation is a point mutation that results in a substitution of an amino acid in a protein sequence. Currently, missense mutations account for over half of the known variants responsible for Mendelian diseases, but accurate prediction of the pathogenicity of missense variants remains challenging. Recent advances show transformer models - a type of deep neural network - to be particularly powerful at modeling sequence information with context dependence, such as language translation tasks. In this study, we introduce MutFormer, a transformer-based model for the prediction of pathogenic missense mutations, using reference and mutated amino acid sequences from the human genome as the features. We first pre-trained MutFormer on reference protein sequences and mutated protein sequences resulting from common genetic variants (assumed to be benign). We next tested different fine-tuning methods for pathogenicity prediction, and determined that utilizing paired-protein sequences (concatenation of the wildtype and mutated protein sequence) yields the most optimal results. Our results show that MutFormer outperforms a variety of existing tools in pathogenicity prediction, including those using conventional machine-learning approaches (e.g., support vector machine, convolutional neural network, recurrent neural network, etc.), and may complement existing computational predictions and empirically generated functional scores to improve our understanding of disease variants. MutFormer and pre-computed variant scores are publicly available on GitHub at https://github.com/WGLab/mutformer.
Source: arXiv (Cornell University)
Authors: Theodore Jiang, Fang Li, Kai Wang
Language: English
Topic: Genomics and Rare Diseases
Subfield: Genetics
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W3209370467