Facilitating genome annotation using ANNEXA and long-read RNA sequencing

This study introduces an updated, open-source ANNEXA pipeline that leverages long-read RNA sequencing data, integrates multiple reconstruction tools and deep learning models, and provides rigorous quality control to enhance genome annotation and identify novel coding and non-coding transcripts across species.

The Gist

Imagine your genome (your DNA) as a massive, ancient library containing the instructions for building and running a human or a dog. For a long time, this library was a bit messy: some books were missing pages, some chapters were out of order, and some stories were completely unwritten.

Genome annotation is the process of organizing this library, labeling every book, and writing summaries so scientists can actually understand what the instructions mean.

Recently, scientists got a new, super-powerful tool called Long-Read RNA Sequencing. Think of this as a high-speed camera that can photograph entire books from cover to cover in one shot, rather than taking blurry snapshots of just a few sentences. This helps us see the full stories (transcripts) much better.

However, there's a catch: even with this amazing camera, the photos can still be a bit blurry, cut off at the edges, or sometimes the camera accidentally invents a story that doesn't exist. We need a way to check the photos and make sure the library catalog is accurate.

Enter ANNEXA: The Super Librarian

This paper introduces ANNEXA, a new computer program (a pipeline) designed to be the ultimate "Super Librarian" for this new type of data. Here is how it works, using some everyday analogies:

1. The Double-Check System (Two Tools in One)

ANNEXA doesn't just trust one method. It uses two different "detectives" to read the library: StringTie2 and Bambu.

• StringTie2 is like a creative writer who looks at the raw photos and tries to piece together every possible story, even the weird ones. It finds a lot of new stories but might sometimes invent a few fake ones.
• Bambu is like a strict editor who only writes down stories that fit perfectly with the known rules of the library. It finds fewer new stories, but the ones it finds are very likely to be real.
• ANNEXA runs both detectives, compares their notes, and combines the best parts of both lists.

2. The "Start-Button" Test (Quality Control)

One common problem with these photos is that they often get cut off at the beginning (the start of the story).

• ANNEXA uses a special Deep Learning AI (a "smart robot") to look at the very first few words of every new story.
• If the robot sees that the story starts in a weird, unnatural place (like a sentence starting in the middle of a word), it flags it as a "cut-off photo" and throws it out. This ensures that only complete, full-length stories make it into the final catalog.

3. The "Story Type" Sorter (Coding vs. Non-Coding)

Not all stories in the library are the same. Some are instruction manuals for building proteins (the "machines" of the cell), and others are "long non-coding RNAs" (lncRNAs), which are more like the library's management notes or regulatory signs.

• ANNEXA has a special sorter (called FEELnc) that looks at every new story and asks: "Is this a machine manual or a management note?"
• This is crucial because scientists often struggle to find the "management notes" because they are harder to spot. ANNEXA makes sure these are found and labeled correctly.

4. The "Dog vs. Human" Connection (Comparative Oncology)

To prove this new librarian works, the authors tested it on a real-world mystery: Cancer.

• They looked at cancer cells from 8 dogs and 2 humans.
• Dogs and humans are like cousins; they often get the same types of cancer (like melanoma and bone cancer), but their "libraries" are written in slightly different languages.
• ANNEXA found brand new stories (genes) in both species. Even better, it found 5 specific stories that were new in dogs and new in humans, and they were located in the same spot in the library.
• This is a huge deal! It suggests these new stories might be important keys to understanding cancer in both species.

Why Does This Matter?

Before ANNEXA, scientists had to manually check thousands of these "photos" to see if they were real, which took forever and was prone to error.

ANNEXA is like an automated quality-control factory:

1. It takes the raw photos.
2. It runs them through two different assembly lines.
3. It uses a smart robot to cut off the blurry edges.
4. It sorts the stories into the right bins.
5. It prints a final, clean report with a visual map so scientists can see exactly what changed.

The Bottom Line

This paper gives scientists a free, open-source tool (ANNEXA) that makes the messy job of organizing the genome's library much faster and more accurate. By ensuring we have the right "books" on the shelf, we can better understand how diseases like cancer work in both humans and our furry friends, leading to better treatments for everyone.

Technical Summary

1. Problem Statement

With the advent of complete, telomere-to-telomere (T2T) genome assemblies, the focus of genomics has shifted from assembly completeness to refining genome annotation. While Long-Read RNA Sequencing (LR-RNAseq) technologies (PacBio and Oxford Nanopore) have revolutionized transcriptome reconstruction by capturing full-length isoforms, they introduce specific challenges:

• Artifacts: LR-RNAseq data often produces truncated transcript models and incomplete isoform representations, particularly at Transcription Start Sites (TSSs).
• Tool Limitations: Existing pipelines often lack robust Quality Control (QC) strategies, fail to distinguish effectively between protein-coding mRNAs and long non-coding RNAs (lncRNAs), or are restricted to specific sequencing technologies (e.g., only Nanopore or only PacBio).
• Annotation Gaps: Current reference annotations are incomplete, especially for lncRNAs, and there is a need for tools that can systematically extend these annotations while filtering out false positives.

2. Methodology: The ANNEXA Pipeline

The authors developed ANNEXA, a scalable, open-source pipeline built on Nextflow (compatible with Docker/Singularity) designed to extend reference annotations using LR-RNAseq data. The pipeline is technology-agnostic (supporting both Nanopore and PacBio) and consists of four integrated modules:

1. Transcriptome Reconstruction & Quantification:

   • Integrates two leading reconstruction tools: Bambu (reference-guided machine learning) and StringTie2 (splice-graph assembly).
   • Allows users to choose between tools or run both to compare sensitivity vs. precision.
   • Handles multi-sample pooling to produce a unified, non-redundant transcript set.

2. Coding Potential & Classification:

   • Uses FEELnc to systematically classify novel transcripts as either protein-coding (mRNA) or long non-coding (lncRNA).
   • Employs TransDecoder to identify the longest Open Reading Frames (ORFs) and predict Coding Sequences (CDS) for protein-coding candidates.
   • Utilizes GffCompare to assign class codes (e.g., j for novel isoforms, u for intergenic, x for antisense) relative to the reference.

3. Transcript Filtering & Full-Length Assessment:

   • Implements a dual-filtering strategy to remove fragmented or biologically invalid models:
     • NDR Cutoff: Uses the Novel Discovery Rate from Bambu to filter low-confidence predictions.
     • TransforKmer Cutoff: A novel deep learning module using a DNABERT-based transformer model (fine-tuned on labeled TSSs) to evaluate the biological validity of Transcription Start Sites.
   • Offers "Union" (retain if passing one filter) or "Intersection" (retain only if passing both) filtering modes to balance sensitivity and precision.

4. Quality Control (QC):

   • Generates comprehensive QC reports comparing known vs. novel elements (genes, transcripts, exons).
   • Metrics include structural features (length distribution, exon count) and quantification metrics (read counts, expression breadth).
   • Includes optional gene body coverage analysis using RSeQC to detect 5'/3' bias.

3. Key Contributions

• ANNEXA Pipeline: A unified, reproducible framework that combines reconstruction, classification, deep-learning-based TSS validation, and QC.
• Deep Learning for TSS Validation: The integration of TransforKmer provides a species-specific, computational method to filter out truncated transcripts, significantly improving annotation reliability without requiring additional wet-lab data.
• Comparative Oncology Application: Demonstrated the pipeline's utility in a cross-species study (Human vs. Canine) involving mucosal melanomas, histiocytic sarcomas, and osteosarcomas.
• Open Source & Reproducibility: The tool is publicly available on GitHub, with all code, input files, and extended annotations (Human/Canine) shared via Zenodo and the ENA database.

4. Results

The authors benchmarked ANNEXA using 10 LR-RNAseq samples (8 canine, 2 human) against multiple reference annotations (Ensembl, RefSeq, UU for dogs; Gencode, CHESS for humans).

• Tool Performance (Bambu vs. StringTie):
  • Bambu consistently reconstructed more known genes and transcripts with higher precision, particularly in well-annotated genomes.
  • StringTie identified more novel elements but with a higher rate of fragmentation.
  • Filtering Impact: Applying ANNEXA's TSS filters drastically reduced the number of novel elements (e.g., retaining only ~50% of Bambu's novel genes and ~16% of StringTie's in dogs), significantly increasing the validation rate against orthogonal CAGE data (from ~9% to >50%).
• lncRNA vs. mRNA Annotation:
  • Both tools reconstructed more mRNAs than lncRNAs from reference annotations.
  • Classification Differences: Bambu was superior at detecting antisense lncRNAs (class x), whereas StringTie predominantly classified novel lncRNAs as intergenic (u) or extensions of known genes. This suggests Bambu is less dependent on reference strand information.
• Cross-Species Conservation:
  • The pipeline identified 9,612 novel canine genes.
  • Orthology mapping revealed that 38.6% of these mapped to the human genome.
  • Crucially, five novel canine lncRNA genes were found to be conserved in humans (mapping to novel human genes), exhibiting high expression in both species, suggesting functional relevance.

5. Significance

• Enhanced Annotation Reliability: ANNEXA addresses the "noise" problem in LR-RNAseq by combining statistical filtering (NDR) with deep learning (TSS validation), ensuring that extended annotations are biologically plausible.
• Flexibility for Diverse Genomes: The pipeline is effective across species with varying annotation completeness (e.g., the highly annotated human genome vs. the less complete canine genome), making it valuable for non-model organisms.
• Functional Discovery: By successfully identifying conserved, novel lncRNAs across species, the study demonstrates that ANNEXA can uncover functional elements that standard short-read pipelines or unfiltered long-read analyses might miss.
• Standardization: Provides a structured, Nextflow-based standard for integrating multiple reconstruction tools and QC metrics, moving the field toward more rigorous transcriptome annotation practices.