Methods of designing and predicting proteins

A method of predicting or designing a protein, the method comprising:

• providing a protein comprising a target residue;
• centering a gridded box on a selected atom of the target residue, wherein the gridded box has sides that are about 10 angstroms to about 30 angstroms in length;
• collecting local structural information adjacent the selected atom and the target residue;
• determining the target residue of the protein based on the local structural information, wherein the determining of the target residue comprises providing the local structural information as input data to a convolutional neural network;
• and
• modifying (i) the target residue to achieve a new function of the protein, or (ii) a designed protein to include the target residue to (a) achieve a new function of the designed protein, or (b) facilitate the folding of the designed protein into a known three-dimensional structure;
• wherein the convolutional neural network has an architecture comprising—
• (i) a layer comprising three parallel three dimensional convolution layers with rectified linear unit activation function, wherein (a) a first of the three parallel three dimensional convolution layers is configured to capture information on covalent bonds having lengths of from about 1.2 angstroms to about 1.8 angstroms, (b) a second of the three parallel three dimensional convolution layers is configured to recognize bond angles formed by two connected bonds, and (c) a third of the three parallel three dimensional convolution layers is configured to capture information on dihedral angles;
• (ii) a 2×2×2 max pooling layer configured to produce an output of dimensions 9×9×9×80,
• (iii) a three dimensional convolution layer comprising 64 channels of 3×3×3 filter, and configured to detect secondary structure motifs,
• (iv) a 3×3×3 max pooling layer configured to produce an output of dimensions 3×3×3×64,
• (v) a flatten layer configured to convert an output of layer (iv) to a simple vector,
• (vi) a fully connected layer with rectified linear unit activation function,
• (vii) a dropout layer with a drop rate of 0.2, and
• (viii) a 20 dimensional activation layer.