Ensembling Diverse Neural Networks for Improved Protein Contact Prediction

Predicting residue-residue contacts from protein sequence is an important task, as protein contacts are relevant to the process of predicting folded protein structures. We propose that the deep learning practice of ensembling has the potential to improve protein contact prediction by combining the outputs of discrete neural networks. We show that ensembling the predictions made by different groups in the recent Critical Assessment of Protein Structure Prediction (CASP13) outperforms all individual groups. Further, we show that contacts derived from the distance predictions of three additional deep neural networks – AlphaFold, trRosetta, and our own ProSPr – can be substantially improved by ensembling all three networks. Finally, we demonstrate that ensembling these recent deep neural networks with the best CASP13 group creates a superior contact prediction tool. These results indicate that combining the predictions of diverse, high quality neural networks can improve protein contact prediction and outperform the best individual models. We call for increased availability of protein contact prediction methods and the creation of a better contact benchmark set, in order to create a community-based ensemble approach to superior protein contact prediction.