Item

Abstract

Coiled-coils domains are present in approximately 15% of proteins and they are involved in a plethora of biological functions such as signal transduction, molecular transport and mediation of oligomerization processes [1]. Thus, their reliable annotation is crucial for studies of protein structure and function. Here, we report DeepCoil [2], a novel neural network-based tool for the detection and localization of coiled-coil domains in protein sequences. DeepCoil predictions are based either on a sequence information alone (DeepCoil_SEQ) or on a sequence and a profile derived from homologous sequences (DeepCoil_PSSM). In a rigorous benchmark both DeepCoil variants outperformed current state-of-the-art methods and detected many coiled coils that remained undetected by other methods. This higher sensitivity of DeepCoil opens up a possibility of an accurate, genome-wide annotation of coiled-coil domains without the time-consuming profile generation. We will also present strategies for improvement of predictor performance through large-scale distributed optimization of hyper-parameters and network architecture.