The Neural Network properties are mathematical combinations of the individual physical properties computed by HotPatch, i.e. concavity, roughness, electrostatic potential and charge, hydrophobicity by-residue and by-atom. These are combined with other simpler properties (e.g. residue exposed area) into "super-predictor" combination properties. Each neural network (NN) is optimized specifically for a particular protein function; that is, if you choose protein function=hydrolase, and property=neural network, HotPatch will use a specific NN optimized and tested only on hydrolases.

When HotPatch makes predictions using individual properties (e.g. hydrophobicity), in the first step, a value of the property (score) is assigned to each atom. However, for Neural Network predictions, the first step is modified. For simplicity, here we average each individual property over all atoms in each residue. Then from averaged values of all properties of the residue, we compute one NN score for it, assigning that to all atoms in the residue.

Next, we find patches of atoms high in NN prediction scores, and compute Functional Confidences of the patches, in the usual HotPatch way. In the final step, residues are assigned to a patch according to the patches to which their atoms belong.

The network networks were trained using all exposed residues in sets of protein structures of the same function. These were subsets of proteins drawn from the SFR Database. The network had two layers and 267 hidden variables. Network weights were optimized via the NevProp statistical package (Goodman, P. (1996). NevProp Software Version, 3rd. ed., University of Nevada, Reno, NV.) We attempted to avoid overtraining the weights.

We are indebted to Philip Goodman for help with NevProp.