ProstaNet: A Novel Geometric Vector Perceptrons-Graph Neural Network Algorithm for Protein Stability Prediction in...

Area of Science:

• Computational Biology
• Biophysics
• Machine Learning

Background:

• Protein stability prediction is crucial for biology and biopharma but experimentally challenging.
• Deep learning offers an efficient computational approach to predict mutation effects on protein stability.

Purpose of the Study:

• Introduce ProstaNet, a deep learning framework for predicting protein stability changes caused by single- and multiple-point mutations.
• Develop a comprehensive dataset (ProstaDB) and innovative methods to train and validate the deep learning model.

Main Methods:

• Utilized a geometric vector perceptrons-graph neural network for 3D feature processing in ProstaNet.
• Created ProstaDB with 3,784 single-point and 1,642 multiple-point mutations, employing thermodynamic looping and clustering for data enhancement and testing.
• Identified residue scoring as a key encoding method for protein property prediction.

Main Results:

• ProstaNet achieved 0.75 accuracy for single-point mutation prediction, surpassing ThermoMPNN (0.63), PoPMuSiCsym (0.66), MUPRO (0.52), and FoldX (0.71).
• ProstaNet demonstrated a 1.3-fold increase in accuracy over FoldX for multiple-point mutation predictions.
• Experimental validation showed 80% accuracy for single-point and 100% for multiple-point mutations in HuJ3 mutants.

Conclusions:

• ProstaNet accurately predicts thermostability changes for both single- and multiple-point mutations without bias.
• The framework shows significant potential for advancing protein engineering and drug development through accurate mutation effect prediction.