1-10% of protein-protein interactions have been identified. Need for prediction. - Roland Krause

Most methods for prediction are genomic or sequence based but not structure based. - Roland Krause

Protein docking: finding and scoring interaction poses between two proteins. Hard problem to distinguish interactors from non-interactors. Use of the Weng benchmark set of 65 complexes. Measure distribution of scores. - Roland Krause

Build decoy set and assume they are not interacting with the benchmark set. Some examples show separation but fails for several. 36 complexes outperform rank better than the 80%. - Roland Krause

Where on the interactors are docking solutions generated? Some AA are involved more frequently than others. - Roland Krause

Still way to go before predictions are possible but a signal is present. - Roland Krause

The problem with methods that try to optimize discrimination between natives and a set of decoys is that they usually find problems in the docking software / energy function. I.e. you'll discover that decoys have very weird electrostatics distribution etc... A good set of non-interactors would be key to develop this field further - Nir London

scoring is indeed a bottleneck in docking. however there is a signal found in this work, it can be very helpful both for docking and prediction of interactions to understand where this signal comes from. - Dina Schneidman