RNAsamba: coding potential assessment using ORF and whole transcript sequence information

Antônio Pedro Camargo, Vsevolod Sourkov, Marcelo Falsarella CarazzollebioRxiv 2019 · DOI

Abstract

The advent of high-throughput sequencing technologies made it possible to obtain large volumes of genetic information, quickly and inexpensively. Thus, many efforts are devoted to unveil the biological roles of genomic elements, being one of the main tasks the identification of protein-coding and long non-coding RNAs. We describe RNAsamba, a tool to predict the coding potential of RNA molecules from sequence information using a deep-learning model that processes both the whole sequence and the ORF to look for patterns that distinguish coding and non-coding RNAs. We evaluated the model in the classification of coding and non-coding transcripts of humans and five other model organisms and show that RNAsamba mostly outperforms other state-of-the-art methods. We also show that RNAsamba can identify coding signals in partial-length ORFs and UTR sequences, evidencing that its model is not dependent on the presence of complete coding regions. RNAsamba is a fast and easy tool that can provide valuable contributions to genome annotation pipelines.

Introduction

As high-throughput sequencing continues to flood biological databases, the "functional annotation" bottleneck has become a primary challenge in genomics. Distinguishing between messenger RNAs (mRNAs), which are translated into proteins, and long non-coding RNAs (lncRNAs), which perform regulatory roles without being translated, is essential for understanding cellular function.

Traditional computational tools for this task often suffer from two major drawbacks. First, many rely on sequence homology, meaning they fail to identify novel protein-coding genes that lack close relatives in existing databases. Second, they often depend on human-designed features, such as Open Reading Frame (ORF) length or GC content. These features can be misleading; for instance, some lncRNAs contain long non-functional ORFs, while many coding transcripts are sequenced only partially, leading to truncated ORFs that traditional tools might misclassify. While deep learning models like Recurrent Neural Networks (RNNs) have been proposed to address these issues, they are often computationally expensive and slow to train on long biological sequences.

Methods

To overcome these limitations, we developed RNAsamba, a deep-learning tool that utilizes a unique neural network architecture called IGLOO. Unlike traditional sequential models, IGLOO extracts "global snapshots" of the feature space by processing patches of data, which significantly improves training speed and the ability to capture long-range dependencies within a transcript.

The RNAsamba architecture consists of an IGLOO-based representation learning layer (A) and a dual-branch classification system (B) that integrates whole-sequence and ORF-specific information.

RNAsamba employs a dual-branch approach to maximize its predictive power:

  • Whole Sequence Branch ($B_1$): This branch processes the entire nucleotide sequence, allowing the model to identify coding signals (such as the Kozak consensus sequence) that may reside outside the ORF.
  • Longest ORF Branch ($B_2$): This branch focuses specifically on the longest ORF and its predicted protein translation, analyzing k-mer frequencies, amino acid composition, and length.
  • Attention-based Weighting: The model incorporates a learnable "alpha" ($\alpha$) parameter to dynamically weight the importance of these two branches. This ensures that if ORF information is sparse or unreliable (as in truncated sequences), the model can rely more heavily on patterns found in the whole-sequence branch.

Results and Discussion

RNAsamba was rigorously benchmarked against state-of-the-art tools (including CPAT, CPC2, and mRNN) across multiple datasets.

Performance and Efficiency

The tool achieved exceptional accuracy, reaching an Area Under the Precision-Recall curve (AUPRC) of up to 99.49% on human transcript datasets. Beyond accuracy, RNAsamba demonstrated significant efficiency gains, performing inference approximately 10.5x faster than lncRNAnet and training 14.2x faster than mRNN. This speed makes it feasible to run on standard CPUs, lowering the barrier to entry for researchers without specialized hardware.

Robustness and Generalization

One of the most significant findings was RNAsamba's robustness to sequence truncation. While traditional tools often fail when the start codon is missing, RNAsamba maintained high recall even when only 20% of the ORF was present. Furthermore, a model trained exclusively on human sequences performed remarkably well across diverse species, including Arabidopsis thaliana (plants) and Drosophila melanogaster (insects). This suggests that the model captures fundamental, evolutionarily conserved properties of protein-coding transcripts.

Limitations

Despite its strengths, RNAsamba has a fixed maximum input length of 3,000 nucleotides. While the authors demonstrate that the most critical coding signals are typically located near the start codon, this constraint might limit the detection of distal regulatory patterns in exceptionally long transcripts.

Conclusion

RNAsamba represents a significant step forward in the automated annotation of transcriptomes. By combining the speed of the IGLOO architecture with a flexible dual-branch approach, it provides a robust and accessible solution for distinguishing coding from non-coding RNAs across the tree of life.

Summary updated

Abstract

Abstract Motivation The advent of high-throughput sequencing technologies made it possible to obtain large volumes of genetic information, quickly and inexpensively. Thus, many efforts are devoted to unveil the biological roles of genomic elements, being one of the main tasks the identification of protein-coding and long non-coding RNAs. Results We describe RNAsamba, a tool to predict the coding potential of RNA molecules from sequence information using a deep-learning model that processes both the whole sequence and the ORF to look for patterns that distinguish coding and non-coding RNAs. We evaluated the model in the classification of coding and non-coding transcripts of humans and five other model organisms and show that RNAsamba mostly outperforms other state-of-the-art methods. We also show that RNAsamba can identify coding signals in partial-length ORFs and UTR sequences, evidencing that its model is not dependent on the presence of complete coding regions. RNAsamba is a fast and easy tool that can provide valuable contributions to genome annotation pipelines. Availability and implementation The source code of RNAsamba is freely available at: https://github.com/apcamargo/RNAsamba .
BibTeX
@misc{camargo2019rnasamba,
  author = {Antônio Pedro Camargo and Vsevolod Sourkov and Marcelo Falsarella Carazzolle},
  title = {{RNAsamba: coding potential assessment using ORF and whole transcript sequence information}},
  year = {2019},
  doi = {10.1101/620880},
  howpublished = {bioRxiv},
}
Figure