4.5 Discussion

Much information about interacting protein residues is typically contained in simple protein sequence features. All popular machine learning and meta-predictors for contact prediction employ sequence derived features as additional source of information besides coevolution scores. In line with this knowledge I developed a random forest classifier for contact prediction that is trained on simple sequence features.

Random forests are a convenient choice for many machine learning applications as they require no input preparation, such as feature scaling, they perform implicit feature selection and provide a robust indicator of feature importance and can handle huge feature space. Furthermore they are quick and straight forward to train and have been shown to perform well for protein contact prediction.

As expected, the random forest model yielded a robust estimator that outperformed coevolution methods for small protein families where they suffer from the low signal-to-noise ratio. Furthermore, I integrated the predictions of the pseudo-likelihood and the constrastive divergence method as additional features for training. Again as expected, the individual methods greatly contribute and improve the predictive performance of the random forest classifier. Even for protein families with many sequences, where coevolutionary methods perform best, the combined random forest model improves over the individual coevolution approaches. Yet, including both coevolution scores as additional features into the random forest model does not help to boost performance further. Apparantly, they do not seem to represent complementary information which was on the other hand already expected from the analysis in chapter 3.