Detecting Novel Sequence Signals in Targeting Peptides Using Deep Learning

José Juan Almagro Armenteros, Marco Salvatore, Olof Emanuelsson, Ole Winther, Gunnar von Heijne, Arne Elofsson, Henrik NielsenbioRxiv 2019 · DOI

TL;DR

This paper introduces TargetP 2.0, a deep learning model that significantly improves the prediction of N-terminal protein targeting signals. Beyond enhanced accuracy, the model reveals that the second amino acid after the initial methionine plays a surprisingly crucial role in protein localization, offering new biological insights.

What Problem Does This Paper Address?

Cells are incredibly organized, with proteins needing to be delivered to specific compartments to function correctly. This precise delivery is orchestrated by "sorting signals" – short sequences of amino acids, often at the protein's beginning (N-terminus). These signals act like molecular zip codes, directing proteins to destinations like the secretory pathway, mitochondria, or chloroplasts. Accurately identifying these signals and their precise cleavage sites is crucial for understanding protein function and cellular processes. While many computational tools exist, there's a continuous need for more accurate and insightful methods, especially as new types of targeting signals are discovered.

Methods

The core of TargetP 2.0 is a deep learning architecture built upon bidirectional Long Short-Term Memory (BiLSTM) networks combined with a multi-attention mechanism.

  • BiLSTM: This recurrent neural network architecture processes the protein sequence in both forward and backward directions. This allows it to capture long-range dependencies and contextual information within the sequence, which is vital for understanding complex biological signals.
  • Multi-Attention Mechanism: This component allows the model to "focus" on specific parts of the protein sequence that are most informative for prediction. By using multiple attention vectors, the model can simultaneously consider different aspects of the sequence, leading to more nuanced predictions.
  • Input Encoding: Amino acids are encoded using BLOSUM62 substitution matrices, a standard practice in bioinformatics that captures evolutionary relationships between amino acids.
  • Output: The model predicts both the type of targeting peptide (e.g., signal peptide, mitochondrial transit peptide) and the position of its cleavage site.

This approach builds upon previous deep learning methods for protein localization but specifically enhances the prediction of N-terminal sorting signals and their cleavage sites.

Datasets

The authors trained and evaluated TargetP 2.0 using data extracted from the UniProt database (release 2018_04).

  • Positive Set: Proteins with experimentally verified signal peptides, mitochondrial transit peptides (mTPs), chloroplast transit peptides (cTPs), and thylakoid transit peptides (ITPs).
  • Negative Set: Proteins without targeting peptides, localized to the nucleus, cytoplasm, or plasma membrane.

The dataset was further partitioned into four eukaryotic kingdoms: Viridiplantae, Metazoa, Fungi, and other eukaryotes. To ensure robust evaluation, a stringent homology partitioning was applied using PSI-CD-HIT to prevent similar proteins from appearing in both training and testing sets.

Key Insights and Ideas

  • The Second Residue Matters (A Lot!): The most striking finding is the significant influence of the amino acid at position 2 (immediately following the initial methionine) on protein localization. This was identified by analyzing the attention layer's focus.
  • Alanine's Role in Plastid Targeting: For chloroplast and thylakoid targeting peptides in plants, the presence of alanine at position 2 is very common (two-thirds of cases). This is in contrast to other plant proteins, where it's much less frequent.
  • Methionine Cleavage and Fungal/Eukaryotic Targeting: In fungi and single-celled eukaryotes, the N-terminal methionine is less likely to be cleaved in targeting peptides compared to non-targeted proteins. This suggests that the ability to cleave the initial methionine is linked to the protein's destination.
  • Deep Learning Uncovers Biological Nuances: The paper highlights how advanced machine learning, beyond just improving prediction accuracy, can reveal underlying biological mechanisms and patterns that were previously overlooked.

Key Results

  • State-of-the-Art Performance: TargetP 2.0 outperforms existing methods in identifying all four types of N-terminal targeting peptides (Signal Peptide, mTP, cTP, ITP) and their cleavage sites.
  • Improved Discrimination: It significantly improves the discrimination between mTPs and cTPs, a common point of confusion for older methods.
  • Cleavage Site Prediction: While signal peptide cleavage site prediction is highly accurate (83% recall), mTP and cTP cleavage site prediction shows substantial improvement over previous methods, reaching 46% and 49% recall, respectively.
  • ITP Prediction: TargetP 2.0 introduces reliable prediction for thylakoid transit peptides (ITPs), a feature not well-handled by previous tools.
  • Speed: The model is fast, taking only about 0.20 seconds per protein on an 8-core machine, making it suitable for large-scale proteome annotation.
  • Agreement with UniProt: TargetP 2.0 shows high agreement (around 80-90%) with UniProt annotations for well-annotated proteomes, suggesting it can significantly improve the annotation of less-studied proteomes.

Limitations and Future Work

  • ITP Underprediction: The model underpredicts ITPs by about 25% compared to UniProt annotations, suggesting room for improvement in this specific category.
  • Fungal Kingdom Accuracy: Prediction accuracy is slightly lower in Fungi, potentially due to lower GC content in fungal genomes affecting amino acid frequencies.
  • mTP Cleavage Site Prediction: While improved, mTP cleavage site prediction still has room for enhancement, particularly regarding arginines at position -10. The authors suggest further exploring peptidase recognition patterns.
  • Importance of Position 2: The significant role of the second amino acid warrants further experimental investigation to fully understand its biological implications, especially concerning methionine aminopeptidases and co-translational targeting.
  • Further Experimental Validation: Additional experimental studies are needed to confirm the importance of the second position and its interaction with methionine aminopeptidases.

Summary updated

Abstract

Abstract In bioinformatics, machine learning methods have been used to predict features embedded in the sequences. In contrast to what is generally assumed, machine learning approaches can also provide new insights into the underlying biology. Here, we demonstrate this by presenting TargetP 2.0, a novel state of art method to identify N-terminal sorting signals, which direct proteins to the secretory pathway, mitochondria and chloroplasts or other plastids. By examining the strongest signals from the attention layer in the network, we find that the second residue in the protein, i.e. the one following the initial methionine, has a strong influence on the classification. When subsequently examining all targeting peptides, we observe that two-thirds of chloroplast and thylakoid transit peptides have an alanine in position two, but only 20% of other plant proteins. Further highlighting the importance of the second residue, we also note that in fungi and single-celled eukaryotes, less than 30% of the targeting peptides have an amino acid that allows the removal of the N-terminal methionine compared with 60% for the proteins without targeting peptide. TargetP 2.0 is available at http://www.cbs.dtu.dk/services/TargetP-2.0/index.php
BibTeX
@misc{armenteros2019detecting,
  author = {José Juan Almagro Armenteros and Marco Salvatore and Olof Emanuelsson and Ole Winther and Gunnar von Heijne and Arne Elofsson and Henrik Nielsen},
  title = {{Detecting Novel Sequence Signals in Targeting Peptides Using Deep Learning}},
  year = {2019},
  doi = {10.1101/639203},
  howpublished = {bioRxiv},
}
Figure