<i>Lite-SeqCNN:</i> A Light-Weight Deep CNN Architecture for Protein Function Prediction
Vikash Kumar, Akshay Deepak, Ashish Ranjan, Aravind Prakash
Abstract
The <i>short-and-long</i> range interactions amongst amino-acids in a protein sequence are primarily responsible for the function performed by the protein. Recently convolutional neural network (CNN)s have produced promising results on sequential data including those of NLP tasks and protein sequences. However, CNN's strength primarily lies at capturing <i>short</i> range interactions and are not so good at <i>long</i> range interactions. On the other hand, dilated CNNs are good at capturing both <i>short-and-long</i> range interactions because of varied – <i>short-and-long</i> – receptive fields. Further, CNNs are quite light-weight in terms of trainable parameters, whereas most existing deep learning solutions for protein function prediction (PFP) are based on multi-modality and are rather complex and heavily parametrized. In this paper, we propose a (sub-sequence + <i>dilated</i> -CNNs)-based simple, light-weight and sequence-only PFP framework <i>Lite-SeqCNN</i> . By varying <i>dilation-rates</i> , <i>Lite-SeqCNN</i> efficiently captures both <i>short-and-long</i> range interactions and has (0.50–0.75 times) fewer trainable parameters than its contemporary deep learning models. Further, <i>Lite-SeqCNN</i> <inline-formula><tex-math notation="LaTeX">$^+$</tex-math></inline-formula> is an ensemble of three <i>Lite-SeqCNN</i> s developed with different segment-sizes that produces even better results compared to the individual models. The proposed architecture produced improvements upto 5% over state-of-the-art approaches <i>Global-ProtEnc Plus</i> , <i>DeepGOPlus</i> , and <i>GOLabeler</i> on three different prominent datasets curated from the UniProt database.