Predicting changes in protein thermodynamic stability upon point mutation with deep 3D convolutional neural networks.

Area of Science:

• Computational biology
• Biophysics
• Structural biology

Background:

• Predicting mutation-induced changes in protein thermodynamic stability (ΔΔG) is crucial for protein engineering and variant interpretation.
• Existing methods may suffer from biases due to protein homology and training data imbalances.

Purpose of the Study:

• Introduce ThermoNet, a deep, 3D-convolutional neural network (3D-CNN) for structure-based ΔΔG prediction.
• To develop a robust method that overcomes limitations of previous ΔΔG prediction approaches.

Main Methods:

• Treated protein structures as multi-channel 3D images (voxel grids) based on biophysical properties.
• Utilized a curated dataset accounting for protein homology and balanced mutations.
• Employed deep, 3D-convolutional neural networks (3D-CNNs) for prediction.

Main Results:

• ThermoNet achieved performance comparable to state-of-the-art methods on the Ssym test set.
• Accurately predicted both stabilizing and destabilizing mutations, unlike biased methods.
• Demonstrated utility in predicting ΔΔGs for clinically relevant proteins and variants from ClinVar.

Conclusions:

• 3D-CNNs can effectively model complex interactions perturbed by mutations directly from atomic biophysical properties.
• ThermoNet offers a promising tool for accurate ΔΔG prediction, aiding protein engineering and clinical variant analysis.
• Highlights potential overestimation of previous methods due to dataset homology.