File Name : supp.fig.1.tiff

Caption : deep residual neural network model architecture. layers are represented by rectangles, while connectors are represented by arrows. the dimensionality of the input and output of each layer is specified on the right hand side of each rectangle. 

File Name : si2.tiff

Caption : schematic diagram of mesh layers covering the area around a protein surface. 
water molecules present within the grid are encoded in terms of an array of vectors pointing towards all proximal (<=7.5å) atoms of the protein. 


File Name : fig_si3.tiff

Caption : performance of the final model. 
a) left axis: model accuracy across training epochs (light blue and dark blue line for train and test set, respectively; right axis: binary cross-entropy loss across training epochs (light red and dark red line for train and test set, respectively  b) receiver operating characteristic curve i.e. false-positive versus true positive rate; area under the curve= 0.985.  

File Name : figsi4.tiff

Caption : water prediction for x-ray crystallography. case studies of the three of the six test proteins, for which raw crystallography data was provided, depicting the recall of phenix update waters and the coot find water functions at a variety of resolutions, in comparison to that of the hotwater algorithm.

File Name : figuresi5.tif

Caption : comparison of positive and negative input data and outputs
top: features describing characteristics of positive (water molecules within pdb files) and negative (heuristically chosen non-water positions; see si 1.1) class samples; bottom: features describing the characteristics of water molecules according to the assigned hot-spot prediction scores: a, c) number of interacting protein’s atoms (<7.5 å from the oxygen atom of the water); b,d) mean temperature factor of all protein atoms in the neighbourhood of the position (<7.5 å from the water molecule).