# Deep Recurrent Neural Network for Protein Function Prediction from Sequence

**Authors:** Xueliang Liu  
**arXiv:** [1701.08318](https://arxiv.org/abs/1701.08318)  
**Date:** 2017-01-28  
**Summary page:** https://www.summarizepaper.com/en/paper/1701.08318v1/

## Key points

- High-throughput biological sequencing has advanced in terms of speed and cost-effectiveness.
- Extracting meaningful information from these sequences is a challenge due to low-throughput experimental characterizations.
- Predicting protein functions directly from primary amino-acid sequences is a particular challenge.
- Machine learning techniques using artificial recurrent neural networks (RNN) are used to address this challenge.
- RNN models incorporate long-short-term-memory (LSTM) units and are trained on annotated datasets from UniProt.
- RNN models achieve high performance for in-class prediction of four important protein functions tested.
- RNN models outperform other machine learning algorithms that use sequence-derived protein features.
- RNN models can make out-of-class predictions for phylogenetically distinct protein families with similar functions.
- Trained RNN models predict candidates validated by existing annotations and identify unannotated sequences in the UniRef100 database.
- Some predictions made by the RNN models for the ferritin-like iron sequestering function were experimentally validated.
- This machine learning approach based on RNN holds great potential for discovering and predicting homologues for a wide range of protein functions.

## Summary

The field of high-throughput biological sequencing has seen significant advancements in terms of speed and cost-effectiveness. However, the challenge lies in extracting meaningful information from these sequences, especially due to the limitations of low-throughput experimental characterizations. One particular challenge is accurately predicting protein functions directly from their primary amino-acid sequences. In this study, the authors address this challenge by applying machine learning techniques using artificial recurrent neural networks (RNN). They specifically focus on classifying protein function directly from primary sequence without relying on sequence alignment, heuristic scoring, or feature engineering. The RNN models used in this study incorporate long-short-term-memory (LSTM) units and are trained on publicly available annotated datasets from UniProt. The results demonstrate that the RNN models achieve high performance for in-class prediction of four important protein functions tested. In fact, they outperform other machine learning algorithms that utilize sequence-derived protein features. Additionally, the RNN models are capable of making out-of-class predictions for phylogenetically distinct protein families with similar functions, including proteins belonging to the CRISPR-associated nuclease, ferritin-like iron storage, and cytochrome P450 families. Furthermore, when applied to the partially unannotated UniRef100 database, the trained RNN models not only predict candidates validated by existing annotations but also identify currently unannotated sequences. Some predictions made by the RNN models for the ferritin-like iron sequestering function were experimentally validated. Notably, these validated sequences differ significantly from known characterized proteins and each other, making them challenging to predict using popular bioinformatics methods. Overall, as sequencing and experimental characterization data continue to increase rapidly, this machine learning approach based on RNN holds great potential for discovering and predicting homologues for a wide range of protein functions. By leveraging primary amino acid sequences without relying on traditional bioinformatics methods such as sequence alignment or feature engineering, this approach offers a promising avenue for advancing our understanding of protein function.