TL;DR
This paper critically re-evaluates a newly proposed set of 1,178 human protein-coding gene predictions from the CHESS database. The authors demonstrate that nearly all of these predictions are false positives, largely stemming from homology to repetitive DNA elements (transposons) and previously misannotated genes, rather than genuine protein-coding potential. This work highlights the critical need for rigorous validation beyond simple homology when identifying novel genes.
What Problem Does This Paper Address?
Accurately identifying all protein-coding genes in a genome is fundamental for understanding biological processes and interpreting genetic variation. While large-scale projects like GENCODE and RefSeq have established high-confidence gene sets, there's ongoing interest in discovering "novel" protein-coding genes. However, distinguishing true novel coding genes from artifacts like transcriptional noise or pseudogenes is challenging. A recent study (Pertea et al.) proposed the CHESS database with 1,178 new human protein-coding predictions, based on transcript evidence and homology. This paper addresses the critical question of whether these new CHESS predictions are indeed functional protein-coding genes, or if they represent a significant source of false positives that could mislead future research.
Methods
The authors employed a multi-pronged approach to re-evaluate the CHESS predictions:
- Homology Analysis: They re-analyzed the homology evidence used by CHESS, comparing it against curated databases (UniProtKB/Swiss-Prot, RefSeq) and the broader GenBank. They specifically investigated the source of homology, identifying a significant reliance on repetitive elements and previously erroneous annotations.
- RepeatMasker Analysis: They used RepeatMasker to identify overlaps between the predicted novel coding transcripts and known interspersed repeats and low-complexity DNA sequences. This is a standard method for identifying regions likely to be non-coding or problematic for gene prediction.
- Evolutionary Conservation Analysis (PhyloCSF): They applied PhyloCSF, a tool that measures evolutionary signatures of protein-coding constraint by analyzing codon substitutions across multiple species. This is a gold-standard method for distinguishing functional coding regions from non-coding sequences. They compared the PhyloCSF scores of CHESS predictions to known coding genes and non-coding RNAs.
- Mass Spectrometry Data: They examined mass spectrometry data for experimental evidence of translation, a direct measure of protein production.
- Manual Curation: They performed manual inspection of genes, following GENCODE annotation guidelines, to assess their validity as protein-coding.
Datasets
- CHESS Database: The primary dataset under investigation, containing 1,178 novel protein-coding gene predictions and 3,819 novel non-coding gene predictions.
- GTEx Project RNA-seq Data: The source data used by Pertea et al. to generate the CHESS transcripts.
- UniProtKB/Swiss-Prot: A curated protein sequence database used for homology comparisons.
- RefSeq Gene Catalog: Another database of gene annotations, including both manual and computational derivations.
- GenBank: A repository of nucleotide sequences, used by CHESS for homology searches.
- RepeatMasker Database: Used to identify repetitive elements in the genome.
- 58 Mammalian Species Alignments: Used for PhyloCSF analysis.
- Human Proteome Project Data: Used for experimental validation of protein evidence.
Key Insights and Ideas
- Transposons are a major source of false positives: The overwhelming overlap (86%) of CHESS predictions with RepeatMasker regions, particularly Alu elements, strongly suggests that many predictions are derived from repetitive DNA sequences that can mimic coding regions.
- Homology to erroneous annotations propagates errors: A significant portion of the CHESS predictions are homologous to sequences that were themselves misclassified as protein-coding, demonstrating how errors in existing databases can be amplified.
- Obsolete Pfam domains are problematic: A substantial number of predictions are homologous to a previously withdrawn Pfam protein domain (GVQW) that was specifically based on Alu elements, highlighting the danger of relying on outdated or flawed annotation resources.
- Lack of evolutionary constraint: PhyloCSF analysis revealed that the evolutionary signatures of the CHESS novel ORFs are indistinguishable from those of known non-coding RNAs, indicating a lack of purifying selection expected for functional protein-coding genes.
- Weak experimental evidence: The mass spectrometry evidence supporting only four predictions was inconclusive, falling far short of the standards required for validating novel coding genes.
Key Results
- 86% of CHESS predictions overlap transposons: This is a primary indicator of false positives.
- Over half are homologous to only nine Alu-derived primate sequences: This points to a highly repetitive and non-diverse source of homology.
- PhyloCSF scores are nearly identical to non-coding RNAs: This strongly suggests a lack of protein-coding potential.
- Only four predictions have mass spectrometry support, and it's inconclusive: This indicates a severe lack of experimental validation.
- Nearly all predictions lack credible experimental or evolutionary evidence of function.
Limitations and Future Work
- Limitations: The paper primarily focuses on re-analyzing the evidence presented by Pertea et al. While they acknowledge that some novel coding genes might exist, their analysis strongly suggests the CHESS set is heavily contaminated with false positives. The paper also notes that PhyloCSF might not be suitable for recently duplicated genes, but argues this is unlikely to explain the vast majority of the CHESS predictions.
- Future Work: The authors implicitly advocate for more rigorous validation pipelines for novel gene discovery. They recommend filtering for transposons, ensuring homology is to manually curated genes, being conservative with proteomics data, and performing manual inspection of all novel predictions. They also emphasize the need for integrating multiple orthogonal datasets (evolutionary signatures, experimental evidence, and expert curation) to achieve high-quality gene annotations, as exemplified by the GENCODE project's ongoing efforts.
Summary updated