Deciphering "the language of nature": A transformer-based language model for deleterious mutations in proteins

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Machine learning models, including deep neural networks, are used to predict the deleteriousness of missense mutations.
• Transformer models, advanced deep neural networks, excel at modeling context-dependent sequence information, offering potential for improved mutation prediction.

Purpose of the Study:

• To introduce MutFormer, a novel transformer-based model for predicting deleterious missense mutations.
• To leverage self-attention and convolutional layers for comprehensive analysis of protein sequence dependencies.

Main Methods:

• MutFormer was pre-trained on human reference and mutated protein sequences from common genetic variants.
• The model was fine-tuned for missense mutation deleteriousness prediction.
• Performance was evaluated on multiple independent testing datasets.

Main Results:

• MutFormer demonstrated comparable or superior performance against existing mutation prediction tools.
• The model effectively captures both long-range and short-range dependencies in protein sequences.
• Novel sequence features were identified and utilized by MutFormer.

Conclusions:

• MutFormer offers a powerful new approach to predicting deleterious missense mutations.
• The model's ability to analyze previously unexplored sequence features enhances understanding of disease variants.
• MutFormer can complement existing computational and experimental methods for variant interpretation.