Introduction
Proteins are the workhorses of the cell, and their function depends on their thermodynamic stability, measured by Gibbs free energy ($\Delta G$). A single-point mutation can alter this stability ($\Delta\Delta G$), potentially leading to disease or, in engineering contexts, enabling the design of more robust industrial enzymes. Predicting these changes computationally is a long-standing goal in biophysics.
Existing methods typically fall into two categories: physics-based models, which are accurate but computationally expensive, and machine learning (ML) models, which are fast but often plagued by training bias and data leakage. ThermoNet addresses these issues by formulating $\Delta\Delta G$ prediction as a computer vision problem, using deep 3D-convolutional neural networks (3D-CNNs) to analyze protein structures as multi-channel 3D images.
Methods
The ThermoNet workflow transforms structural data into a format suitable for deep learning, capturing complex chemical interactions directly from spatial arrangements.
Structural Modeling and Voxelization
Because experimental structures for every possible mutant do not exist, the researchers used the Rosetta FastRelax protocol to model the 3D structure of the mutant protein based on the wild-type template. A $16 \times 16 \times 16$ Å cube is defined around the mutation site and discretized into a 3D grid (voxels) with a 1 Å resolution.
Feature Engineering
Each voxel is parameterized with seven biophysical property channels: hydrophobicity, aromaticity, hydrogen bond donors, hydrogen bond acceptors, positive ionizable groups, negative ionizable groups, and occupancy. By stacking the 7 channels of the wild-type and the 7 channels of the mutant, a 14-channel 3D "image" is created.
Architecture and Training
This 14-channel input is fed into an ensemble of ten 3D-CNNs. Each network consists of three convolutional layers followed by a fully connected layer to output the predicted $\Delta\Delta G$. To ensure the model learned physical "anti-symmetry"—the principle that $\Delta\Delta G_{A \to B} = -\Delta\Delta G_{B \to A}$—the training set was augmented with reverse mutations. Furthermore, rigorous homology pruning was applied to the training data to prevent the model from "cheating" on proteins similar to those in the test sets.
Key Results
ThermoNet achieves state-of-the-art performance while eliminating the common bias toward predicting protein destabilization.
- Accuracy: On the balanced $S^{sym}$ test set, ThermoNet achieved a Pearson correlation ($r$) of 0.47 and a Root Mean Square Error ($\sigma$) of ~1.56 kcal/mol, comparable to top-tier methods like DDGun3D.
- Unbiased Predictions: Unlike many competitors, ThermoNet’s average bias was nearly zero (-0.01 kcal/mol), with a near-perfect correlation between direct and reverse predictions ($r_{dir-rev} = -0.96$).
- Clinical Relevance: When applied to ClinVar variants, ThermoNet correctly predicted that benign variants cluster in a "neutral zone" (between -0.5 and 0.5 kcal/mol), while pathogenic variants were significantly more likely to cause large stability shifts.
Discussion
The core innovation of ThermoNet is the use of 3D-CNNs to capture non-linear chemical interactions (like van der Waals forces or hydrogen bonding) directly from spatial atomic arrangements without manual feature engineering. While the model currently requires high-resolution structures, limiting its scope to reliably modeled proteins, it demonstrates the power of deep learning in biophysics when combined with high-quality, unbiased datasets. Future work may expand this architecture to protein-DNA and protein-RNA interactions.
Summary updated