Figure 2.

3D convolutional neural network architecture of DeepKa. The crystal structure of the protein as an example is visualized with the NewCartoon model by VMD.⁵⁹ The titratable carboxyal oxygens of an Asp residue are highlighted with red balls. The input structure is the 20 Å open red cubic box centralized at the center of mass of the two carboxyl oxygens. The four solid boxes represent 4D tensors with the grid sizes of 21 × 21 × 21, 11 × 11 × 11, 6 × 6 × 6, and 3 × 3 × 3, respectively, and the channel counts per grid point are 20, 64, 128, and 256, respectively. In the dense block, the input layer includes 6912 neurons and the following three hidden layers have 1000, 500, and 200 neurons, respectively. The output layer with one single node (green) is the predicted pK_a.