BGPT: Paper Review: Facilitating genome annotation using ANNEXA and long-read RNA sequencing

Explore by Goal

Quick Explanation Copied

ANNEXA: long-read genome annotation extension with QC + TSS filtering

ANNEXA is a Nextflow pipeline that extends a user’s reference annotation using long-read RNA-seq transcript reconstruction (Bambu + StringTie2), codespotting with FEELnc, CDS prediction with TransDecoder, and a dedicated deep-learning TSS “full-lengthness” filter (TransforKmer). It outputs both a permissive “full” and a stricter “filtered” annotation, plus a structured QC report inspired by SQANTI. https://github.com/IGDRion/ANNEXA and the preprint are the core sources reviewed here.

Long Explanation

Paper review (skeptical, evidence-based): “Facilitating genome annotation using ANNEXA and long-read RNA sequencing”

Primary reviewed source: preprint DOI 10.1101/2025.04.16.648718

What the paper claims (structure, not hype)

Pipeline purpose: extend reference annotations with novel genes/transcript isoforms from long-read RNA-seq, and simultaneously produce quality-control outputs (full vs filtered annotations).
Transcript reconstruction: integrates Bambu and StringTie2; it merges reconstructions differently for each tool (StringTie per-sample then merged with -merge; Bambu across samples using TPS/NDR).
Coding potential & lncRNA annotation: FEELnc is used for coding potential prediction; TransDecoder.LongOrfs + TransDecoder.Predict are used for CDS/ORF reporting on novel protein-coding transcripts.
TSS-based full-length filtering: the paper describes TransforKmer, a species-specific deep-learning model for TSS validity; it is described as pre-trained with DNABERT and fine-tuned with labeled TSSs.

Benchmarking logic (and the key statistical trap)

The paper’s benchmarking uses reference-based comparisons (GffCompare) and labels reconstructed features as TP/FP/FN relative to the supplied reference annotation. This is legitimate for quantifying “recovery of known features,” but it cannot compute specificity and can misclassify “true but novel” biology as FP, and “true but not expressed in the experiment” biology as FN.

Skeptical bias audit (what could mislead)

Reference completeness bias: changing input reference alters how many “novel” elements are discovered, especially for StringTie vs Bambu behaviors.
Expression/mismatch bias: if the reference has features not present in the specific samples, they can become FNs; if the CAGE validation context differs from LR-RNAseq conditions, TSS validation rates can be underestimated.
QC filter trade-off: producing both full and filtered outputs is good practice, but the union vs intersection choice determines precision–recall; selecting the wrong mode can inflate either artifacts or missed true transcripts.

From the paper’s numbers: tool/reference dependence + filtering impact

The text provides explicit examples of how “novel” counts shift when switching reconstruction tools and reference annotations. For canFam4 with the Ensembl reference, Bambu detects substantially fewer novel genes and transcripts than StringTie2: 3,134 novel genes and 3,600 novel transcripts (Bambu) vs 12,671 novel genes and 20,184 novel transcripts (StringTie).

How to interpret this without overclaiming

Higher “novel counts” may mean higher sensitivity, or lower specificity/artifact tolerance (StringTie retention of more novel transcripts is discussed as a graph/splice-graph-based behavior in comparison to Bambu).
ANNEXA’s TSS validity filter can dramatically reduce novel elements; the paper reports different retention proportions by species and tool (e.g., dogs: ~50% Bambu novel genes retained vs ~16% StringTie after TSS validity check; humans: ~65% Bambu vs ~46% StringTie for CHESS).

Conceptual workflow (reconstruction → classification → TSS filtering → QC)

The workflow schema is based strictly on the ANNEXA module descriptions: reconstruction with Bambu/StringTie2, coding potential with FEELnc and CDS prediction with TransDecoder, filtering with NDR + TransforKmer, and QC outputs.

QC framework transparency

ANNEXA states it computes multiple QC features by comparing known and novel genes/transcripts/exons, inspired by SQANTI. This is helpful: it shifts evaluation from “just counts” to “structure + expression + coverage biases.” However, like many QC pipelines, QC metrics are still proxies unless anchored to orthogonal ground truth across the same biological conditions.

Open scientific questions (what would change the conclusion)

Orthogonal validation breadth: the paper uses CAGE validation as an orthogonal signal, but discusses potential mismatch with LR-RNAseq conditions; systematic validation across matched contexts (same cell lines, matched library protocols, matched TSS capture conditions) could change how strongly TransforKmer’s filter is believed.
Unannotated biology handling: because benchmarking is reference-based, “FP” could include true unannotated transcripts. Adding an endogenous ground truth framework (or independent assays beyond transcript annotations) could recalibrate the claimed precision gains.
Generalization beyond cancer cell lines: ANNEXA is demonstrated on cancer cell lines across two species (human and dog). Generalization to other tissues/cell states and to additional long-read library chemistry protocols is not fully established in the reviewed text.

Author reviews (deep dive links)

Feedback:

Updated: April 25, 2026