Enhancing Missense Variant Pathogenicity Prediction with MissenseNet: Integrating Structural Insights and...

Area of Science:

• Human genetics
• Computational biology
• Bioinformatics

Background:

• Classifying missense variant pathogenicity is vital for genetic disease diagnosis and personalized medicine.
• Traditional methods face limitations in feature selection and generalizability.
• Accurate prediction requires advanced computational models integrating diverse data.

Purpose of the Study:

• To develop and validate an enhanced deep learning model, MissenseNet, for accurate missense variant pathogenicity classification.
• To improve upon existing methods by incorporating structural protein information.
• To optimize the prediction of functional impacts for clinical applications.

Main Methods:

• Developed MissenseNet, a deep learning model based on the ShuffleNet architecture.
• Incorporated an encoder-decoder framework and a Squeeze-and-Excitation (SE) module for adaptive feature weighting.
• Utilized structural insights from AlphaFold2 protein predictions to enhance feature representation.

Main Results:

• MissenseNet demonstrated superior accuracy compared to conventional pathogenicity prediction methods.
• Achieved the highest Area Under the Receiver Operating Characteristic (ROC) curve and Area Under the Precision-Recall (PR) curve on an independent test set.
• Validated the model's effectiveness in classifying variant pathogenicity and functional impact.

Conclusions:

• MissenseNet offers a significant advancement in missense variant pathogenicity prediction.
• The integration of AlphaFold2 structural data enhances model performance.
• This model holds potential for improving genetic diagnostics and guiding personalized treatment strategies.