Detecting sequence signals in targeting peptides using deep learning.
Almagro Armenteros JJ, Salvatore M, Emanuelsson O, Winther O, von Heijne G, Elofsson A, Nielsen H. · 2019
Read Detecting sequence signals in targeting peptides using deep learning. by Almagro Armenteros JJ, Salvatore M, Emanuelsson O, Winther O, von Heijne G, Elofsson A, Nielsen H. (2019) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc6769257.txt.
Introduction
Almagro Armenteros and co-authors used deep learning to find signals in protein targeting peptides. Proteins carry address tags that send them to the right cell compartment. While building TargetP 2.0, a leading predictor of these tags, the team noticed a biological signal earlier work had missed. The deep learning output exposed it.
Targeting peptides differ by destination. Some route proteins into the secretion path, others into mitochondria or chloroplasts. The paper explains the regions these tags are built from. It also reports predictions for nearly 289,000 proteins across 13 species, from yeast to humans. All predictions sit on a companion website. Better tag reading sharpens maps of where proteins do their work.
Catalog record
| CREATOR | Almagro Armenteros JJ, Salvatore M, Emanuelsson O, Winther O, von Heijne G, Elofsson A, Nielsen H. |
|---|---|
| YEAR | 2019 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 48 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Detecting sequence signals in targeting peptides using deep learning By Jose Juan Almagro Armenteros, Marco Salvatore, Olof Emanuelsson, Ole Winther, Gunnar von Heijne, Arne Elofsson, Henrik Nielsen Published in: Life Science Alliance DOI: 10.26508/lsa.201900429 ABSTRACT During the development of TargetP 2.0, a state-of-the-art method to predict targeting signal, we find a previously overlooked biological signal for subcellular targeting using the output from a deep learning method. INTRODUCTION The localisation of proteins in the cell is a fundamental determinant of protein function. Specific sorting signals drive the subcellular localisation of proteins. These signals vary in structure, length, and position between the different subcellular compartments. One of the most common types of sorting signals are the N-terminal targeting peptides. These signals are responsible for sorting proteins to the secretory pathway, mitochondria, chloroplasts (or other plastids), and compartments inside the chloroplast such as thylakoids. Signal peptides (SPs) are responsible for transporting proteins to the endoplasmic reticulum to enter the secretory pathway. SPs are composed of three regions: a positively charged domain or n-region, a hydrophobic core or h-region, and a segment before the cleavage site (CS) or c-region (von Heijne, 1990). Mitochondrial transit peptides (mTPs) are responsible for targeting proteins to the mitochondrial matrix. mTPs are usually enriched in arginine, leucine, and serine. Moreover, they tend to form an amphiphilic helical structure to interact with the import receptor on the mitochondrial membrane (von Heijne, 1986). Proteins targeted to the inner mitochondrial membrane or the inter-membrane space often have a bipartite mTP, where the second part is similar to an SP (Stuart & Neupert, 1996) (Fig S1). [Figure: Representation of the attention weights for a few randomly chosen proteins.The height of the letter represents the attention weight in that position and the letter the type of amino acid. The shaded area corresponds to the predicted targeting peptide (SP, mTP, cTP, or luTP).] Chloroplast transit peptides (cTPs) are involved in the transport of proteins to the chloroplast stroma. Most of the cTPs consist of three regions: an uncharged N-terminal region, a central region lacking acidic amino acids but enriched in serine and threonine, and a C-terminal region enriched in arginine that forms an amphiphilic β strand (Heijne et al, 1989a). In addition, chloroplastic proteins targeted to the thylakoid lumen have a bipartite pre-sequence structure (Robinson & Klsgen, 1994). Once the cTP is cleaved and the protein enters the stroma, a luminal transit peptide (luTP) is recognised, and the protein is further transported to the thylakoid, where the luTP is cleaved. The luTP is similar to a bacterial SP, and the thylakoidal processing peptidase belongs to the family of signal peptidases (Packer and Howe, 2013). As these signals direct the transport of proteins within the cell, it is crucial to be able to predict their presence in protein sequences accurately. For this reason, in the last two decades, many tools have been developed. Those adopt various machine learning algorithms, including Grammatical Restrained Hidden Conditional Random Fields, N-to-1 Extreme Learning Machines, Support Vector Machines, Markov chains, profile-hidden Markov models, and neural networks (Emanuelsson et al, 2000; Small et al, 2004; Petsalaki et al, 2006; Fukasawa et al, 2015; Savojardo et al, 2015). One of the most used methods is TargetP 1.1 (Emanuelsson et al, 2000). TargetP uses feed-forward networks and position-weight matrices to process windows of amino acids to predict the presence of SPs, mTPs, and cTPs and the positions of their CSs. However, with the rise of deep learning, new types of networks such as recurrent neural networks (RNNs) have gained popularity. The main reason is their extraordinary ability to work with sequence data and model long-range relationships between inputs in the sequence. RNNs sequentially process sequences of any length, being able to retain information from previous positions in the sequence. Several methods have taken advantage of this type of network to try to better predict signal and transit peptides (Reczko & Hatzigeorgiou, 2004; Boden & Hawkins, 2005). These methods make use of bidirectional RNNs (BiRNN), which are two RNNs, one processing the sequence forwards and another processing the sequence backwards. With this construction, the context around each amino acid is modelled, as the forward RNN processes all the amino acids from the N terminus up to one position and the backward RNN processes all the amino acids from the C terminus up to the same position. However, regular RNNs, the so-called Elman networks, are challenging to train (the so-called exploding/vanishing gradient problem) and often fail to capture dependencies far apart in the sequence (Pascanu et al, 2012 Preprint). Therefore, the ability of the network to hold information from multiple steps back is reduced. A variant of the RNN cell, the long short-term memory (LSTM), solves this problem by a construction akin to a computer memory cell that holds information for multiple steps. This type of RNN cell together with BiRNN has been successfully applied to the prediction of SPs and mTPs (Thireou & Reczko, 2007; Almagro Armenteros et al, 2019). Today, new methods, such as DeepLoc (Almagro Armenteros et al, 2017), use bidirectional LSTM (BiLSTM) to predict the localisation of proteins to a broader range of compartments. DeepLoc accurately predicts the localisation of proteins but not the presence of the N-terminal sorting signals and the position of the CSs. Starting from this architecture, we decided to develop TargetP 2.0 using BiLSTM and a multi-attention mechanism. Using the multi-attention mechanism, the network can predict both the type of peptide and the position of the CS by focusing on particular regions of the sequence. Moreover, we assemble a new protein dataset that we use to train TargetP 2.0 (http://www.cbs.dtu.dk/services/TargetP-2.0/). TargetP 2.0 can jointly predict the presence of SPs, mitochondrial, chloroplast and thylakoid transit peptides, and the corresponding CS positions. TargetP 2.0 is available at http://www.cbs.dtu.dk/services/TargetP-2.0/, and the source code is available under the creative commons CC BY-NC-SA license from https://github.com/JJAlmagro/TargetP-2.0/. When analysing the attention layer from the final version of the network, it became apparent that most information was retrieved from two distinct positions in most sequences. One of these was, as expected, localised close to the CS. However, an equally important signal from position 2 in the sequences was also found. Next, we examined the amino acid frequencies in the second position, after the first methionine, of all proteins. To our surprise, very distinct patterns emerged. In chloroplasts and plastids, the second residue was frequently an alanine, whereas in all targeting peptides in fungi and unicellular eukaryotes amino acids that allow cleavage of the methionine are rare (see Fig 1). [Figure: This figure depicts the frequencies of the second residue in proteins with different targeting peptides.The proteins are divided into their respective type of targeting peptide: signal peptide (SP), mitochondrial transit peptides (mTPs), chloroplast transit peptides (cTPs), luminal transit peptides (luTPs), and noTPs. Furthermore, the proteins were divided into their kingdom: Viridiplantae (P), Metazoa (M), Fungi (F), and other eukaryotic organisms (O) sequences. Inspired by sequence LOGOs, the height of each letter corresponds to the frequency of that amino acid. Only the frequencies for the short side-chained amino acids that allow the cleavage of the N-terminal methionine are shown.] RESULTS AND DISCUSSION Here, we have developed a deep learning model to predict targeting peptides described in