Evolved Matrix Operations for Post-processing Protein Secondary Structure Predictions

Predicting the three-dimensional structure of proteins is a hard problem, so many have opted instead to predict the secondary structural state (usually helix, strand or coil) of each amino acid residue. This should be an easier task, but it now seems that a ceiling of around 76% per-residue three-state accuracy has been reached. Further improvements will require the correct processing of so-called “long-range information”. We present a novel application of genetic programming to evolve high-level matrix operations to post-process secondary structure prediction probabilities produced by the popular, state-of-the-art neural network-based PSIPRED by David Jones. We show that global and long-range information may be used to increase three-state accuracy by at least 0.26 percentage points – a small but statistically significant difference. This is on top of the 0.14 percentage point increase already made by PSIPRED’s built-in filters.