Beyond Homology Transfer: Deep Learning for Automated Annotation of Proteins

Accurate annotation of protein functions is important for a profound understanding of molecular biology. A large number of proteins remain uncharacterized because of the sparsity of available supporting information. For a large set of uncharacterized proteins, the only type of information available is their amino acid sequence. In this paper, we propose DeepSeq – a deep learning architecture – that utilizes only the protein sequence information to predict its associated functions. The prediction process does not require handcrafted features; rather, the architecture automatically extracts representations from the input sequence data. Results of our experiments with DeepSeq indicate significant improvements in terms of prediction accuracy when compared with other sequence-based methods. Our deep learning model achieves an overall validation accuracy of 86.72%, with an F1 score of 71.13%. Moreover, using the automatically learned features and without any changes to DeepSeq, we successfully solved a different problem i.e. protein function localization, with no human intervention. Finally, we discuss how this same architecture can be used to solve even more complicated problems such as prediction of 2D and 3D structure as well as protein-protein interactions. Introduction A biological cell is an intricately arranged chemical factory, the sophisticated organization of which results into multi-cellular species with staggering complexities. At molecular level proteins are the main workhorses of these living cells. The knowledge of protein functions is of paramount importance for an ample understanding of these complex molecular machines that ultimately govern life. Accurate identification of protein function has implications in a wide variety of areas, which includes, discovering new therapeutic interventions, understanding novel diseases as well as designing their cures, better agriculture etc. Experimental methods are generally used to study protein functions but they cannot scale up to the task, due to associated cost and time. On the other hand, high throughput experiments like next generation sequencing technologies are resulting in a large number of new protein sequences that remain uncharacterized1. Figure 1 shows the latest major database statistics, which reflect a rapidly increasing gap between known sequences (GenBank and EMBL curves) and structurally characterized sequences (PDB curve). This tremendous growth of uncharacterized proteins poses a serious challenge at the forefront of molecular biology. To overcome this, computational approaches are generally relied upon for the annotation of protein functions. Many techniques have been proposed in the recent past that exploit a wide variety of data to predict protein functions2–7. A 198 0 198 3 198 6 198 9 199 2 199 5 199 8 200 1 200 4 200 7 201 0 201 3 201 6 102 104 106 108 Year N um be r of En tr ie s in D at ab as e GenBank EMBL KEGG PDB Figure 1. Growth of major sequence, pathway and 3D structure databases1. The amount of available sequence information (GenBank, EMBL) is orders of magnitude greater than that associated with structure information (PDB) peer-reviewed) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not . http://dx.doi.org/10.1101/168120 doi: bioRxiv preprint first posted online Jul. 25, 2017;