ThermoNet

ThermoNet predicts changes in protein thermodynamic stability (ΔΔG) caused by single-point amino acid substitutions by applying deep 3D convolutional neural networks to protein structures represented as multi-channel 3D voxel grids derived from raw atom coordinates.

Key Features:

• Deep Learning Architecture: An ensemble of deep 3D-CNNs processes multi-channel voxel grids to capture complex spatial relationships within protein structures.
• Structural Representation: Protein structures are encoded as multi-channel 3D images that represent biophysical properties computed from raw atom coordinates.
• Data Handling and Bias Mitigation: Training and evaluation use a curated dataset that controls for protein homology and includes both direct and reverse mutations to mitigate biases from homologous sequences in sets such as S2648 and VariBench.
• Performance: Demonstrates performance comparable to state-of-the-art methods on the Ssym test set and predicts both stabilizing and destabilizing mutations without the typical destabilization bias.
• Applications in Practice: Applied to predict ΔΔGs for clinically relevant proteins including p53 and myoglobin and to assess pathogenic and benign missense variants from ClinVar.

Scientific Applications:

• Protein Engineering: Predicts thermodynamic effects of mutations to guide design of proteins with altered stability.
• Variant Interpretation: Provides thermodynamic impact predictions to aid interpretation of missense variants and differentiate pathogenic versus benign alterations.
• Protein Biophysics: Enables investigation of mutation-induced structural and energetic changes at the molecular level.

Methodology:

Proteins are represented as multi-channel 3D voxel grids of biophysical properties from atomic coordinates, an ensemble of deep 3D-CNNs is trained on a curated, homology-controlled dataset including direct and reverse mutations, and performance is evaluated on the Ssym test set.