Genomic sequences and resources

We downloaded all assemblies and annotations of vertebrate organisms available in Ensembl version v.108 [16], comprising of 309 genomes, each corresponding to a different species or strain. For the data at the National Center for Biotechnology Information (NCBI), we used the tool “NCBI datasets” to obtain all genome assemblies corresponding to vertebrates, then considered only those with an available gene annotation [17,18]. Because some annotations were available but limited to the mitochondrial genome or scaffold structure, we only considered GTF files whose archive size was > 3MB. This resulted in a set of 1,172 NCBI vertebrate genomes, each representing a distinct species. Selenoprofiles (see below) was successfully run on all genomes, except for 25 NCBI and 3 Ensembl assemblies. These assemblies consistently produced non-recoverable errors and were consequently excluded from our dataset. The list of genome assemblies is available in S1 Table. For analysis, we differentiated between RefSeq (ids starting with “GCF_”) from GenBank (“GCA_”) entries. The phylogenetic tree of Ensembl species we downloaded from Ensembl Compara [19]. The analogous NCBI tree was obtained from NCBI Taxonomy [20].

Selenoprotein gene prediction

Selenoprotein gene prediction was performed using Selenoprofiles [14,15] version v4.5.7, available at https://github.com/marco-mariotti/selenoprofiles4. Selenoprofiles is a computational pipeline to identify members of the known selenoprotein families and related proteins. It uses several homology-based gene prediction programs, and employs a set of manually curated multiple sequence alignments of selenoprotein families to scan genomes for homologues. The profiles we used are available at https://github.com/marco-mariotti/selenoprotein_profiles (v1.4.1). We scanned genome assemblies for all known selenoprotein families in metazoa and focused our analyses on the predictions labelled as “selenocysteine”, i.e., the genes encoding for selenoproteins, with no apparent pseudogene features (e.g., frameshifts). Our results in vertebrate genomes originated from 21 profiles representing 19 protein families. Multimember families are represented by a single profile that can correctly predict all its paralogs (e.g., the GPX profile can predict Sec-encoding GPX1, GPX2, GPX3, GPX4, GPX6 and Cys-paralogs GPX5, GPX7, GPX8). SELENOW and SELENOV proteins share high homology and are therefore considered as groups within the same protein family, represented by a single profile (SELENOW). On the other hand, though we consider SELENOF and SELENOE proteins to be part of the same homologous family, they are dissimilar enough to warrant two different profiles, for better prediction performance. Lastly, SELENOK genes are highly divergent in insects, which led to the development of two distinct profiles (SELENOK and SelKi) whose results are joined in post-processing. Whenever a profile encompasses multiple groups or paralogs, the orthology routine (described below) allows to classify predicted genes in the relevant subfamilies.

Annotation assessment

To assess the annotation of selenoproteins in public databases, we developed a new utility called Selenoprofiles assess, now integrated in Selenoprofiles v4.5.7. This tool takes as input two GTF/GFF files, one from Selenoprofiles and one from another source, and matches gene annotations based on their overlap in genomic coordinates. Overlapping annotations must occur on the same strand, and on the same frame (except for the “Out of frame” classification). After analysis of the Sec-encoding codon and the rest of coding sequence, each Selenoprofiles prediction is assigned an annotation status label, as shown in Fig 1 and described in Results. To develop this utility, we took advantage of PyRanges, a package to represent and manipulate genomic data in Python [21]. Plots were built using the R package ggplot2 [22].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Classification of selenoprotein annotation/misannotations.

Segments represent coding sequence (CDS) regions of selenoprotein genes. Selenoprofiles predicted CDSs are colored in dark grey, and annotated Ensembl/NCBI CDSs are in light grey. Dotted vertical lines represent start and end positions of the Sec encoding codon. Introns were shrunk to a fixed size for visualization purposes. The classes in blue, i.e., all other than “Well annotated” and “Absent”, are collectively referred to with the term “Misannotation” in this manuscript.

https://doi.org/10.1371/journal.pcbi.1013885.g001

Phylogenetic reconstruction and orthology

The protein sequences of all Selenoprofiles predictions across all genomes were collected using the Selenoprofiles join utility. These were used as input to phylogenetic reconstruction workflows, available in the ETE3 v3.1.3 build framework [23]. We ultimately analyzed trees obtained with the “eggnog41” workflow [24], which internally uses PhyML v20160115.patched for phylogenetic reconstruction [25]. We visualized gene trees with the show_tree program in the biotree_tools package v0.0.6 (https://github.com/marco-mariotti/biotree_tools.git), based on the ETE3 python module [23].

For every selenoprotein family with multiple paralogs in vertebrates (GPX, TXNRD, DIO, SELENOW/V), we inspected phylogenetic trees and manually partitioned them into orthologous groups, here referred to as subfamilies, by applying the species overlap method and following maximum parsimony criteria [26]. To assign subfamily names, we performed similarity searches against Uniprot and consulted relevant literature [6,27,28].

Next, we built a new utility called Selenoprofiles orthology, now available in v4.5.7, constituting a fast alternative to tree building for subfamily assignment. To this aim, we first generated reference alignments (“anchors”) for each multimember selenoprotein family including representatives of each subfamily, which were assigned to orthologous groups beforehand via phylogenetic reconstruction as detailed above. Alignments were performed using Mafft v7.45331 [29]. Anchor alignments were reduced to a maximum of 4, 8, or 12 representative sequences for each subfamily for each lineage, automatically selected using TrimAl [30] to preserve sequence diversity. In the Selenoprofiles orthology procedure, candidate sequences are aligned to reference alignments using Mafft, and Pyaln v0.1.4 (https://github.com/marco-mariotti/pyaln.git) is used for fast calculation of similarity scores of candidate sequences against the representative sequences of each subfamily. These scores are used for subfamily assignment of each candidate. Similarity score parameters are displayed in S2 Table. Results of our benchmark with 3,952 Selenoprofiles predictions in Ensembl genomes are shown in S3 Table, which led to choosing anchor alignments with size of 12 sequences, and this Pyaln sequence comparison method: gap positions are excluded (gaps = n), and the average weighted sequence identity “AWSI” (metrics = ‘w’) is computed with weights based on the maximum frequency of non-gap characters (weights = ‘m’). The same similarity metrics is used for ranking gene predictions in the novel filtering procedure Selenoprofiles lineage (see Results). Sequence and tree manipulations throughout the project were carried out using the programs alignment_tools and tree_tools from the biotree_tools package.

Evaluation of selenoprotein annotation

We set to evaluate the status of selenoprotein annotations in Ensembl and NCBI. The identification of all selenoprotein genes in a genome is not trivial, and it is likely that many selenoprotein families remain undiscovered in less studied organisms, e.g., unicellular eukaryotes [8]. To elude this problem, we focused on vertebrates: their selenoproteome has been thoroughly studied [4,6,7,31–36], so that, we argue, it can be reliably identified by homology to known selenoprotein families. Our dataset consisted of paired genome assemblies and genome annotations, 306 from Ensembl and 1,147 from NCBI (Methods). We subdivided NCBI genomes into two categories: GenBank (413 entries), which includes genome and annotation data submitted directly by researchers without further curation, and RefSeq (734 entries), which comprises a non-redundant selection of genomes consistently annotated using the NCBI Eukaryotic Genome Annotation Pipeline (EGAP) [37]. As bona fide selenoprotein annotation, we ran Selenoprofiles, which yielded 8,827 and 31,572 “raw” selenoprotein predictions in Ensembl and NCBI genomes, respectively, corresponding to 19 protein families (see Methods, Data Availability). Our choice is justified by the high accuracy of Selenoprofiles, whose benchmark demonstrated sensitivity and specificity above 90% at the nucleotide level when compared to manually annotated gene structures [15]. Later on, we critically evaluated and refined this method by adding a novel Selenoprofiles lineage utility, detailed in later sections of this manuscript, which yielded a higher-quality set of “filtered” predictions (8,085 and 29,242 in Ensembl and NCBI, respectively).

We inferred the status of selenoprotein annotation by matching genomic coordinates from Selenoprofiles predictions with annotated coding sequence (CDS) regions and evaluating whether the Sec codon was included in the annotated CDS. If it was not the case, we further classified them into different misannotation cases depending on the position of Sec and annotated CDS (Fig 1). CDS annotations that ended precisely at the Sec UGA were classified as “Stop codon”. The misannotation classes “Upstream”, “Downstream”, and “Skipped” were applied when there were portions of annotated CDS at the 5’, 3’, or both 5’ and 3’ of the Sec UGA, respectively. Some selenoprotein annotations exhibited a distinct frame compared to Selenoprofiles predictions, and were classified as “Out of frame” cases. Transcripts which contained a splicing site in the Sec codon were classified as “Spliced”. Instances where no annotation was overlapping to a Selenoprofiles selenoprotein CDS were classified as “Absent”.

The great majority of selenoprotein families contain a single Sec residue. Notable exceptions in vertebrates are SELENOP, which contains 10 Secs in human and up to 37 in fish [38], and SELENOL, which contains two nearby Sec forming a diselenide bond [36]. While the first Sec residue of SelenoP is embedded in a conserved thioredoxin-like motif, the rest are located in a Sec-rich C-terminal tail, which is particularly challenging to identify in genomes [38]. For simplicity, here we considered only the first Sec residue of each selenoprotein gene to define its annotation status.

The status of selenoprotein annotation in Ensembl

Our results show that most selenoprotein genes have defective annotations in Ensembl: only ~11% of predicted selenoproteins present at least one “Well annotated” entry (Fig 2A). Around 8% of selenoproteins have no annotation at all (“Absent”). The majority (~81%) are “Misannotations”, i.e., there are only flawed annotations which lack the Sec-encoding UGA. We next examined annotation at transcript level, wherein a single Selenoprofiles gene prediction may match multiple annotated transcripts, particularly in well studied model organisms. This analysis allowed us to distinguish among misannotation types (Fig 1). We observed that almost 22% of transcripts skip the Sec-UGA-containing exon, or at least its relevant portion (Fig 2B). There are also numerous cases where the annotated CDS ends at the Sec codon (33%), or further upstream (10%), or starts downstream of it (18%). Other cases were more rare.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Status of selenoprotein annotation in Ensembl, RefSeq and GenBank.

a. Bar plot showing the annotation status of selenoproteins (Selenoprofiles predictions) at gene level in different databases. Three main annotation cases are shown, with “Misannotation” aggregating all classes shown in Fig 1 other than “Well annotated” and “Absent”. Bar shading differentiates the raw (i.e., unfiltered) Selenoprofiles predictions and those filtered with Selenoprofiles lineage as explained later in the manuscript. Bar height represents percentages, and numbers on top of bars show actual counts. Bar colors correspond to each database. b. Bar plot showing the annotation status at transcript level in Ensembl, RefSeq and GenBank, with all possible annotation types. c. Heatmap representing the distribution of annotation/misannotation classes for each selenoprotein family, calculated at transcript level using filtered predictions. The right panel shows the gene structures of representatives of each family: the CDS is represented in blue, and Sec residues are displayed as red lines. Families are sorted by the position of Sec residue, with those carrying it at the C-terminus placed at the bottom. Note that the SELENOW family contains both SELENOW and SELENOV genes, since they are highly homologous and included in the same Selenoprofiles profile. (*) SELENOP: Annotation status was determined using only the first Sec residue of each selenoprotein.

https://doi.org/10.1371/journal.pcbi.1013885.g002

Next, we examined the annotation status for each selenoprotein family separately. Surprisingly, the most well-known selenoprotein families do not exhibit the best annotation status in Ensembl (Fig 2C); rather, SELENOH and SELENOP (first Sec) exhibit the highest percentage of “Well annotated” cases (29% and 24%, respectively). Consistent with intuition, we observed a clear correlation between misannotation type and the position of Sec in the gene structure. In protein families that carry Sec in close proximity to the C-terminus, such as TXNRDs and selenoproteins S, O, K, and I, the Sec residue tends to be annotated as a “Stop codon” signal, or the CDS ends further “Upstream” (Fig 2C). On the other hand, ”Downstream” cases are more abundant in families containing Sec in their N-terminal part, such as GPX, DIO, SEPHS2, and selenoproteins H, M, N, T, V, W. Unsurprisingly, the selenoproteins SELENOE [32], SELENOL [36], and SELENOJ [31], that are not present in human and mouse, exhibit the worst status in Ensembl, with <1% “Well annotated” genes.

We noticed that misannotations are not homogenously distributed across species. Instead, “Well annotated” selenoprotein genes are concentrated in few model species (e.g., human, mouse, zebrafish, pig). In contrast, almost all selenoproteins are misannotated in the rest of species. The full tree of Ensembl species with the complete classification of their selenoprotein genes is shown in S1 Fig, while Fig 3A shows a reduced version with a representative set of species. Altogether, we attribute the presence of well annotated selenoprotein genes in model organisms to the work of manual curators, while the pervasive mistakes in remaining taxa stem from the lack of a dedicated automated pipeline at Ensembl. In some cases, this led to the paradoxical situation of closely related species exhibiting selenoproteome annotations of very different quality.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Distribution of misannotations across species.

The species tree of six representative organisms is shown with the annotation of selenoprotein genes in Ensembl (a), RefSeq (b) and GenBank (c), color-coded by annotation/misannotation labels grouped by subfamily. Gene-level annotation classes are used (Fig 1), wherein “Absent” (grey) means that a selenoprotein predicted by Selenoprofiles has no overlapping annotation in Ensembl/NCBI. Moreover, the yellow color is used to mark “Missing” genes, i.e., expected in this genome based on lineage but not detected by Selenoprofiles. Also, red crosses mark those genes “Filtered out” by Selenoprofiles lineage, i.e., unexpected genes. Subfamilies with C-terminal Sec are highlighted in green. In the species tree, stars denote model organisms with well annotated selenoproteomes, and a triangle denote NCBI genomes with well annotated selenoproteins except for the C-terminal Sec families.

https://doi.org/10.1371/journal.pcbi.1013885.g003

The status of selenoprotein annotation in NCBI GenBank and RefSeq

Our analysis reveals that selenoprotein annotation is flawed in NCBI GenBank, too: only ~5% of selenoprotein genes are “Well annotated”, while 84% fall into the “Misannotation” category and 11% are completely “Absent”. At the transcript level, the most frequent error remains treating the Sec residue as a stop codon, but the other classes are also common. Well annotated cases are spread across diverse species with no apparent underlying pattern. Correct models appear sporadically across a wide taxonomic range with no discernible trend, reflecting the fact that GenBank records deposited by the research community are generated through diverse pipelines that do not incorporate dedicated selenoprotein routines.

In contrast, vertebrate selenoprotein genes in RefSeq genome annotations have “Well annotated” Sec residues in 77% of cases. The rest of cases are either “Absent” (6%) or “Misannotations” (17%). By analyzing the distribution of misannotation types at transcript-level and by family (Fig 2), it appears evident that the main flaw of RefSeq annotation of selenoproteins concerns the protein families with C-terminal Sec residues. Indeed, the “Stop codon” class constitutes 75% of “Misannotation” cases. It also appears that, in RefSeq, the annotation status of fish-specific selenoprotein families is not worse than mammalian families. Analyzing of the distribution of misannotations across species (representative tree in Fig 3; full set in S2 Fig), we notice that we can divide RefSeq genome annotations in two groups: 1. A handful of model species have selenoproteomes that are completely well annotated; 2. Some species have well annotated selenoproteins, except for a few C-terminal Sec protein families, namely SELENOS, TXNRD, and, in a lesser extent, SELENOI, SELENOK and SELENOO. Taken together, our analysis shows that RefSeq relies on manual curated entries for a few model species, while it deploys an automated pipeline for the rest. The RefSeq pipeline evidently has a procedure in-place for selenoproteins that is overall effective, but falters when the Sec residue is followed by a very short C-terminal extension.

To better illustrate the status of selenoprotein annotation in vertebrates, we summarized our results in 18 organisms commonly used in research (Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Selenoprotein annotation status in common vertebrate models in Ensembl and NCBI RefSeq. The table reports selenoprotein gene counts for each species in each of these categories: W = well annotated; Ma = misannotated; A = absent from annotation; Mi = missing from assembly. Expected = the total number of selenoproteins expected in that genome.

https://doi.org/10.1371/journal.pcbi.1013885.t001

Improving selenoprofiles for fully automated reliable annotation of selenoproteins

We set to provide an fully automated computational tool to produce reliable selenoprotein annotations. Previously, we developed Selenoprofiles, a pipeline for selenoprotein gene prediction in genomes featuring remarkable gene prediction accuracy [15]. Yet, Selenoprofiles was not integrated in the annotation pipeline of public databases, to our knowledge. We ascribe this at least partially to the characteristics of the early versions of Selenoprofiles, which were hard to install and run. Here, we present a new version which remedied these shortcomings and introduce various other improvements. First, Selenoprofiles can now be installed via a single conda command, or by pulling a ready-made Docker image, as explained in the new comprehensive online documentation (https://selenoprofiles4.readthedocs.io/). Moreover, code consistency and reproducibility are ensured through a robust continuous integration system implemented on GitHub, which automatically runs validation tests at every code change.

Selenoprofiles orthology: Automatic assignment of selenoprotein subfamilies

Second, we created a new utility called Selenoprofiles orthology to assign gene predictions to specific subfamilies, i.e., orthologous groups/ paralogs. Indeed, Selenoprofiles predicts all paralogous genes of a given family (e.g., DIO1, DIO2, DIO3) using the same profile (DIO), so that, in previous versions, the user could not differentiate among paralogs without a dedicated posterior analysis. The Selenoprofiles orthology utility addresses this task by adopting a fast alternative to phylogenetic tree reconstruction, the state-of-the-art for orthology assignment (Fig 4). Our procedure is based on curated reference alignments referred to as “anchors”, containing selected representatives for all subfamilies for each multimember selenoprotein family. New selenoprotein predictions are aligned to anchors, and their protein sequence similarity is computed against each subfamily, ultimately assigning each prediction to its most similar subfamily.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. New Selenoprofiles utilities for subfamily classification pipeline and prediction refinement.

Initially, we built a phylogenetic tree for every multimember family to produce a bona-fide classification of orthologous groups (a). Next, we devised an automated, fast similarity-based procedure that does not require tree building, now implemented as the Selenoprofiles orthology utility (b). Finally, we developed a filter based on the expected selenoproteome for each lineage, now implemented as the Selenoprofiles lineage utility (c).

https://doi.org/10.1371/journal.pcbi.1013885.g004

To develop this functionality, we built gene trees and manually partitioned them into orthologous groups (Methods), which replicated well our previous phylogenetic analysis of vertebrate selenoproteins [6]. From these, we derived anchor alignments with reliable subfamily labels, which we then reduced for faster processing by automatic selection of sequence representatives. For subfamily assignment of candidates, we tested various metrics of sequence similarity, differing by how gaps are scored and how different columns are aggregated. We then selected the method that maximized the accuracy of orthology assignment in a set of 3,952 selenoprotein predictions from four multimember families (GPX, TXNRD, DIO and SELENOW/V) in Ensembl genomes (Methods; S2 and S3 Tables).

Selenoprofiles lineage: Filtering based on expected selenoproteomes per lineage

Lastly, we decided to improve Selenoprofiles so that its predictions roughly match the accuracy of experienced manual curation. While our pipeline already features high sensitivity, it can lead to some false positives, typically in presence of abundant retrotransposed pseudogenes in mammalian genomes [15]. In Selenoprofiles, most pseudogenes are already labelled and discarded due to the presence of in-frame stop codons or frameshifts (e.g., 888 predictions across 306 Ensembl genomes, an average of 2.9 per genome). Yet, some happen to lack these features and are output. At manual inspection, these may be identified by three criteria: 1. they are typically intronless, 2. most lack evidence of transcription, and 3. they are not conserved. Because some functional genes are also intronless (e.g., Sephs2 [6,39]), the first criterion is not universally applicable and cannot be easily employed in an automated pipeline. The second criterion is also not suitable for Selenoprofiles, because it requires additional data (transcriptomics) not universally available. We thus decided to implement the conservation criteria in a novel utility called Selenoprofiles lineage, which consists in a filter based on the set of selenoproteins previously observed in other species related to the one under analysis. This functionality depend on a manually curated table defining the precise selenoprotein content of all major vertebrate lineages (Data Availability). We compiled this data based on our recent analysis of vertebrate selenoproteomes [7].

Selenoprofiles lineage (Fig 4C) is built on top of the orthology utility: gene expectations are defined at the subfamily level, for improved precision. First, the species under analysis is mapped into a vertebrate lineage present in the expectation table (e.g., placentals, birds), either manually or automatically via NCBI taxonomy. Next, the program checks for extra predicted genes: for example, two copies of GPX4 are found in the mouse genome, but a single one is expected in placentals (Fig 3). Finally, genes are selected based on sequence similarity to the subfamily anchor alignment: only the N top-scoring candidates are kept, where N is the number of expected genes. Besides marking some genes as “Filtered out”, Selenoprofiles lineage may also report some subfamilies as “Missing”, if a given genome contains fewer predictions than expected.

Our new utility allow to refine predicted selenoproteomes in full automation. Results are shown in Fig 2, which shows the annotation status of Ensembl and NCBI before and after lineage-based filtering, and in Fig 3, S1 Fig and S2 Fig, wherein filtered and missing genes are shown for each species. Mostly, filtered out predictions corresponded to “Absent” and “Misannotation” rather than “Well annotated” cases, consistent with the notion that functional genes are carefully annotated more often than pseudogenes. Altogether, the lineage utility yields predictions with quality comparable to manual analysis. On the other hand, it may miss out on lineage-specific duplications that are not reported in the expectation table, although this was compiled via a thorough analysis of hundreds of vertebrates [7]. We also note that this scored-based approach based may occasionally favour recent retrotransposed pseudogenes over functional copies, when these are highly similar. In the future, we may incorporate further selection criteria, such as exon–intron structure expectations.

The new orthology and lineage functionalities are available in Selenoprofiles v4.5.7. In the future, we may expand their use to non-vertebrate genomes.