DeepShape: estimating isoform-level ribosome abundance and distribution with Ribo-seq data

Hailin Hu, Hongfei Cui, Hailin Hu 0002, Jianyang Zeng, Ting ChenBMC Bioinformatics 2019 · DOI

Introduction

Ribosome profiling (Ribo-seq) provides a high-resolution snapshot of protein synthesis by sequencing ribosome-protected fragments (RPFs). However, the short length of these fragments (~30 nucleotides) often leads to "multi-mapping," where a single read aligns to multiple transcript isoforms. Traditional approaches either discard these reads, assign them randomly, or rely on RNA-seq data as a proxy for translation levels—a problematic assumption given that translation is independently regulated and mRNA abundance does not always correlate with protein synthesis.

Isoform-level mapping challenges: Reads from shared exons (e.g., Exon C) can originate from multiple transcript variants.

The DeepShape Framework

To address these limitations, the authors developed DeepShape, a computational framework that leverages deep learning to accurately assign RPFs to specific isoforms using Ribo-seq data alone. DeepShape operates on the principle that local genetic sequences influence ribosome "traffic," creating predictable patterns of stalling or acceleration based on sequence context.

Methodology

DeepShape consists of two primary components:

  1. The Shape Model (CNN): A Convolutional Neural Network trained to learn the relationship between codon sequences and ribosome distribution. The model analyzes a 121-codon window (60 upstream, 60 downstream, and the target P-site) to predict the "instructive ribosome shape."
  2. Iterative Allocation: The framework employs an iterative process. It begins with a uniform distribution of reads and progressively refines these estimates. The CNN's predicted shape guides the allocation of multi-mapped reads, which in turn provides better data to further train the model until convergence.
The DeepShape architecture: (A-B) Iterative read allocation and shape modeling; (C) CNN architecture for predicting ribosome distribution from sequence context.

Validation and Performance

The model was validated using synthetic data from human HeLa cells and benchmarked against diverse organisms including S. cerevisiae, E. coli, mice, and humans.

  • Superior Accuracy: On synthetic datasets, DeepShape achieved a Pearson Correlation Coefficient (PCC) for ribosome abundance exceeding 0.99, significantly outperforming existing state-of-the-art tools like Ribomap (PCC ~0.93).
  • High-Resolution Shape Prediction: The CNN achieved an average PCC of 0.60 for predicting ribosome profiles, a substantial improvement over previous wavelet-based methods like RiboShape.
  • Independence from RNA-seq: Notably, DeepShape demonstrated higher accuracy using only Ribo-seq data than competing methods that utilize both Ribo-seq and RNA-seq, making it a valuable tool for researchers with limited library types.

Biological Discovery: The Codon Residence Index (CRI)

The researchers introduced the Codon Residence Index (CRI) to quantify the relative speed of ribosomes on specific codons compared to their synonymous versions. Applying DeepShape to PC3 human prostate cancer cells treated with the mTOR inhibitor PP242 revealed that certain "fast" codons (e.g., GCC, GGC) slow down significantly under drug treatment. Furthermore, the model identified isoform-specific responses in metastasis-related genes like VIM and CD44, highlighting translational regulation that gene-level analyses typically overlook.

Conclusion

DeepShape represents a significant advancement in translatomics, providing a robust method for isoform-level analysis without the need for supplemental RNA-seq data. While the model currently depends on accurate transcriptome annotations and a fixed sequence window, its ability to untangle complex ribosome traffic patterns offers a powerful new lens for studying translational control in health and disease.

Summary updated

BibTeX
@article{hu2019deepshape,
  author = {Hailin Hu and Hongfei Cui and Hailin Hu and Jianyang Zeng and Ting Chen},
  title = {{DeepShape: estimating isoform-level ribosome abundance and distribution with Ribo-seq data}},
  journal = {BMC Bioinformatics},
  year = {2019},
  doi = {10.1186/s12859-019-3244-0},
}
Figure