Mutation effect estimation on protein-protein interactions using deep contextualized representation learning.

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Protein mutations alter protein-protein interactions (PPIs), impacting biological functions.
• Current computational methods for predicting mutation effects on PPIs often require extensive feature engineering and expert knowledge.
• Existing methods struggle to effectively capture the subtle impacts of point mutations.

Purpose of the Study:

• To develop an end-to-end deep learning framework, MuPIPR, for predicting the effects of mutations on PPI properties.
• To improve the accuracy and efficiency of estimating mutation-induced changes in protein binding affinity and structural characteristics.

Main Methods:

• MuPIPR utilizes a contextualized representation mechanism to propagate mutation effects through amino acid sequences.
• A Siamese residual recurrent convolutional neural encoder is employed to encode wild-type and mutated protein pairs.
• Multi-layer perceptron regressors predict quantifiable changes in PPI properties.

Main Results:

• MuPIPR outperforms state-of-the-art methods in estimating binding affinity changes on the SKEMPI v1 dataset.
• MuPIPR achieves comparable performance to existing methods on the SKEMPI v2 dataset.
• The framework demonstrates state-of-the-art performance in predicting changes in buried surface areas using only sequence data.

Conclusions:

• MuPIPR offers a powerful and efficient deep learning approach for predicting mutation effects on PPIs.
• The framework's ability to leverage sequence information alone simplifies the prediction process.
• MuPIPR advances computational strategies for understanding the functional consequences of protein mutations.