bioRxivBioinformaticsPreprint
Longest protein, longest transcript or most expression, for accurate gene reconstruction of transcriptomes?
Abstract
Methods of transcript assembly and reduction filters are compared for recovery of reference gene sets of human, pig and plant, including longest coding sequence with EvidentialGene, longest transcript with CD-HIT, and most RNA-seq with TransRate. EvidentialGene methods are the most accurate in recovering reference genes, and maintain accuracy for alternate transcripts and paralogs. In comparison, filtering large over-assemblies by longest RNA measures, and most RNA-seq expression measures, discards a large portion of accurate models, especially alternates and paralogs. Accuracy of protein calculations is compared, with errors found in popular methods, as is accuracy of transcript assemblers. Gene reconstruction accuracy depends upon the underlying measurements, where protein criteria, including homology among species, have the strength of evolutionary biology that other criteria lack. EvidentialGene provides a gene reconstruction algorithm that is consistent with genome biology.
- organism: human
- organism: plant (other)
- organism: multiple species
- assay: bulk RNA-seq
- research focus: gene regulatory network inference
- research focus: variant calling
- research focus: benchmarking
- approach: statistical model
Abstract summary
Claude Opus 5.5 · xhigh effort · 18 Oct 2026 · from the abstract · CC BY-NC-ND 4.0
Across human, pig and plant reference gene sets, keeping the longest coding sequence recovers more true gene models than longest-transcript or highest-expression filters, which drop alternates and paralogs.
Large transcript over-assemblies are collapsed with a single criterion, such as transcript length or expression level, and that choice decides which models survive. Scoring the alternatives against reference gene sets in three species indicates the criterion matters most for alternate transcripts and paralogs, exactly the models these filters discard.
Key results
SuggestedReduction on the longest coding sequence, as implemented in EvidentialGene, recovers reference genes most accurately among the compared criteria, and keeps that accuracy for alternate transcripts and for paralogs.
Comparison against reference gene sets for human, pig and plant; the abstract reports the ranking but no recovery percentages, counts of recovered genes or statistical comparison. · Abstract
SuggestedCollapsing an over-assembly by longest RNA length (CD-HIT) or by most RNA-seq support (TransRate) throws away a large share of correct gene models, with alternate transcripts and paralogs hit hardest.
Same reference-set comparison; the loss is described as a large portion but is not quantified, and no breakdown by species or by model class is given. · Abstract
ClaimedThe comparison also covers protein calling and transcript assemblers, with errors reported in widely used protein calculation methods.
Stated as a finding without naming the affected methods, the kinds of error, their frequency, or the assembler ranking. · Abstract
Methods, data and reproducibility
Transcript assemblies for human, pig and plant are reduced using different criteria - longest coding sequence with EvidentialGene, longest transcript with CD-HIT, and greatest RNA-seq support with TransRate - and the retained models are scored for recovery of the corresponding reference gene sets, with protein calling and assembler accuracy compared alongside.
- Reference gene sets used as the recovery target · human, pig, plant
- RNA-seq transcript over-assemblies passed through each reduction filter · human, pig, plant
- Methods & tools
- EvidentialGene, CD-HIT, TransRate
Limitations (2)
- No quantities are reported: recovery rates, numbers of discarded models, the plant species used and the identity of the protein-calling methods said to contain errors are all missing from the abstract. · noted in this summary
- The benchmark is assembled by one author and ends by endorsing one of the compared tools; the abstract describes no independent or blinded evaluation of the criteria. · noted in this summary
Read this if you assemble transcriptomes and must pick a rule for collapsing an over-assembly without losing alternate transcripts and paralogs.
- Citations
- 59
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- Leveraging multiple transcriptome assembly methods for improved gene structure annotation
Luca Venturini, Shabhonam Caim, … David Swarbreck·GigaSciencePeer-reviewed·Jul 24, 2018
- Strawberry: fast and accurate genome-guided transcript reconstruction and quantification from RNA-seq
Ruolin Liu, Julie Dickerson·bioRxiv·Mar 16, 2016
- Choosing representative proteins based on splicing structure similarity improves the accuracy of gene tree reconstruction
Esaie Kuitche, Marie Degen, … Aïda Ouangraoua·bioRxiv·Apr 9, 2020
- Error, noise and bias in de novo transcriptome assemblies
Adam H. Freedman, Michele E. Clamp, Timothy B. Sackton·Molecular Ecology ResourcesPeer-reviewed·Mar 17, 2020
- Genes of the Pig, Sus scrofa, reconstructed with EvidentialGene
Don Gilbert·PeerJPeer-reviewed·Feb 1, 2019
- Transcript Assembly and Annotations: Bias and Adjustment
Qimin Zhang, Mingfu Shao·PLoS Computational BiologyPeer-reviewed·Dec 21, 2023