A Neural Network Approach for the Analysis of Reproducible Ribo-Seq Profiles

Giorgia Giacomini, Caterina Graziani, Veronica Lachi, Pietro Bongini, Niccolò Pancino, Monica Bianchini, Davide Chiarugi, Angelo Valleriani, Paolo AndreiniAlgorithms 2022 · DOI · OpenReview

TL;DR

This paper introduces a novel statistical and machine learning approach to identify reproducible Ribo-seq profiles, addressing the inherent variability in this powerful technique for studying translation. By developing a method to extract reliable translation speed signals, the authors demonstrate that both nucleotide composition and sequence order play crucial roles in modulating ribosome dynamics.

What Problem Does This Paper Address?

Ribosome profiling (Ribo-seq) is a cutting-edge technique that provides single-nucleotide resolution insights into how ribosomes translate mRNA into proteins. This allows researchers to study gene expression regulation at a fundamental level. However, a significant challenge with Ribo-seq data is its poor reproducibility across different experiments and laboratories. This variability stems from complex experimental protocols and a lack of standardized data analysis methods. Consequently, extracting reliable and consistent translation speed information from Ribo-seq data has been difficult, hindering our ability to understand the factors that influence translation dynamics.

Methods

The authors propose a multi-step approach to tackle the reproducibility issue. First, they develop a statistical method to identify "highly reproducible" Ribo-seq profiles by comparing data from multiple independent experiments. This involves:

  • Preprocessing: Standardizing Ribo-seq data from various sources.
  • Signal Digitalization: Converting Ribo-seq coverage profiles into binary sequences (+1 for slow elongation/high ribosome density, -1 for fast elongation/low ribosome density) by comparing nucleotide-level coverage to the ORF median.
  • Digital Profile Comparison: Quantifying similarity between these digital profiles using a "matching score."
  • Statistical Significance Testing: Employing a rigorous statistical framework, including null distributions and False Discovery Rate (FDR) correction, to identify profiles with statistically significant similarity across datasets.
  • Reproducible Sub-sequence Extraction: Once reproducible profiles are identified, they extract sub-sequences with consistent translation speed labels.

To validate the quality and informative content of these reproducible profiles, they employ machine learning models:

  • Multilayer Perceptron (MLP): Used to predict translation speed based solely on the nucleotide composition (relative frequencies of A, T, G, C) of sub-sequences.
  • Convolutional Neural Networks (CNNs): Applied to process the sub-sequences as sequential data, allowing them to capture information from both nucleotide composition and order.
  • Ensemble CNN: A more robust approach combining predictions from seven individual CNNs to improve stability and accuracy.

Datasets

The study primarily utilizes publicly available Ribo-seq datasets for E. coli from the Gene Expression Omnibus (GEO) repository. Specifically, they analyzed a subset of eight samples (Dataset 1 to Dataset 8) from experiments using the K-12 MG1655 genotype.

A new benchmark dataset, termed "SubsequencesDataset," was constructed from the identified reproducible Ribo-seq profiles. This dataset consists of mRNA sub-sequences labeled with their translation speed (slow or fast). It was further split into training (413 sub-sequences) and testing (46 sub-sequences) sets, with a validation set created from 15% of the training data for unbiased evaluation.

Key Insights and Ideas

  • Nucleotide Composition Matters: The statistical analysis revealed that the relative frequencies of nucleotides are significantly different between slow and fast translation regions, suggesting that nucleotide composition is not random and carries information about translation speed.
  • Sequence Order is Crucial: By employing CNNs, which can process sequential data, the authors demonstrated that the order of nucleotides within a sub-sequence also provides valuable information for predicting translation speed, outperforming models that only consider nucleotide frequencies.
  • Reproducibility is Achievable: The core contribution is a robust methodology for identifying reproducible Ribo-seq profiles, which is a critical step towards reliable downstream analysis of translation dynamics.
  • Biological Significance of Clusters: Visualizing sub-sequences based on nucleotide pair frequencies showed distinct clusters for slow and fast translation regions, highlighting how specific nucleotide combinations are associated with different translation speeds.

Key Results

  • Identification of Reproducible Profiles: Out of 3588 common genes across eight E. coli datasets, 40 genes were identified as having significantly reproducible Ribo-seq profiles.
  • MLP Performance: The MLP model achieved an average accuracy of 82.24% in predicting translation speed based on nucleotide frequencies alone, significantly outperforming models trained on random sub-sequences.
  • CNN Performance: The 1-D CNN model, which considers sequence order, achieved an average accuracy of 89.39%, an improvement of approximately 8 percentage points over the MLP.
  • Ensemble CNN Performance: The ensemble of seven CNNs further improved accuracy to 91.02% with a reduced standard deviation, demonstrating a more stable and reliable prediction of translation speed.

Limitations and Future Work

  • Human Data Complexity: The authors acknowledge that applying their method to more complex human data (e.g., from liver tumors) requires further analysis, particularly in defining ad-hoc neural architectures and specialized hyperparameter searches.
  • Generalizability: While the method is presented as effective for "any kind of Ribo-seq data," further validation across diverse organisms and experimental conditions would strengthen this claim.
  • Further Biological Insights: The paper suggests that their method opens "new exciting frontiers" for analyzing ribosome translation dynamics in different organisms and investigating which specific features influence translation speed. They also mention a preliminary analysis on human liver tumor data that showed promising results, indicating a potential future direction.

Summary updated

Abstract

In recent years, the Ribosome profiling technique (Ribo–seq) has emerged as a powerful method for globally monitoring the translation process in vivo at single nucleotide resolution. Based on deep sequencing of mRNA fragments, Ribo–seq allows to obtain profiles that reflect the time spent by ribosomes in translating each part of an open reading frame. Unfortunately, the profiles produced by this method can vary significantly in different experimental setups, being characterized by a poor reproducibility. To address this problem, we have employed a statistical method for the identification of highly reproducible Ribo–seq profiles, which was tested on a set of E. coli genes. State-of-the-art artificial neural network models have been used to validate the quality of the produced sequences. Moreover, new insights into the dynamics of ribosome translation have been provided through a statistical analysis on the obtained sequences.
BibTeX
@article{giacomini2022neural,
  author = {Giorgia Giacomini and Caterina Graziani and Veronica Lachi and Pietro Bongini and Niccolò Pancino and Monica Bianchini and Davide Chiarugi and Angelo Valleriani and Paolo Andreini},
  title = {{A Neural Network Approach for the Analysis of Reproducible Ribo-Seq Profiles}},
  journal = {Algorithms},
  year = {2022},
  doi = {10.3390/a15080274},
}

Loading related content...

Figure