ISPIP: Improved prediction of epitope binding sites

Abstract

ISPIP is a meta-method designed to improve on previous classifiers by choosing components that use different strategies and using machine learning algorithms to train the model. It is based on Walder’s Meta-DPI, but replacing PredUs 2.0 with SPPIDER, so that the three classifiers included are ISPRED4, SPPIDER, and DockPred. Another significant change is in our strategy for combining results of different technologies. Walder used a logistic regression, taking into account that any residue can only have one of two possible states – interface or not interface – which was an improvement over some previous meta-methods that used linear regression (7). We tested several different algorithms including linear and logistic regression models, and machine learning models including random forest and xgboost, to find the best way to combine each of the classifier’s predictions. The development of ISPIP has been the work of a team led by Dr. Viswanathan and including Moshe Carrol, and Alexandra Roffe. My personal contribution was largely in comparing the results of ISPIP to another predictor, DiscoTope 2.0.