Step 1

In the first phase of the BLAST-based screening, our goal was to determine whether previously identified orphan genes in one species, identified two years ago, were still unique to that species. A large proportion of the putative orphan genes were not unique to the original species as orphans in TRGdb shared detectable homology when re-evaluated (Fig 2b). The number of target sequences was kept at the default BLAST setting (100) to optimize computational efficiency and reduce processing time. This reduced the number of hits considered. Only those genes that showed no similarity at the 100-target threshold were taken to Step 2.

Step 2

After identifying initial orphan gene candidates, we next sought to determine whether some of these genes had simply escaped detection due to conservative BLAST search parameters. Step 2 re-examined all remaining candidates by altering the “maximum target sequences.” The increase in the “maximum target sequences” to the maximum limit allowed by NCBI allowed detection of homologs for some genes that exhibited no similarity in the first round. The count of orphan genes declined once more in Step 2 (Fig 2b). We opted not to perform Step 2 in the first place due to the intense computational demands associated with the high number of target sequences. Step 1 facilitated quicker elimination of genes that lost their status as “orphans” by reducing processing time. The longer processing time required for Step 2 made it a more suitable follow-up step after narrowing down the orphan genes. The difference in results when increasing the “maximum target sequences” in BLASTp is because of how BLAST processes its search set. BLAST does not merely return the first N hits that exceed the specified E-value threshold [15]. The BLAST search is divided into two phases. First, the database search phase, where the query sequence is divided into subsequences (smaller words), and these words are then compared to target sequences in the database. Second, the alignment phase, where BLAST executes detailed alignment of these selected target sequences with the query [14]. The former step is designed to rapidly identify the most promising target sequences that are likely to produce meaningful alignments. When the “maximum target sequences” parameter is set to a lower value, the search gets restricted to a subset of the most promising sequences, optimizing speed but missing potential homologs. By increasing the value to 5,000, BLAST could detect additional homologs that may not have been identified in an initial, smaller search. Although this leads to an increase in computational time, it enables a more comprehensive comparison, thereby enhancing the sensitivity of the alignment. This emphasizes the trade-off between computational efficiency and sensitivity in database searches.

Step 3

To filter out low-confidence predictions, we applied a BLASTp thresholding approach in which any ORF detected in three or fewer genomes of the target species was classified as a low-confidence orphan gene (Fig 2a). This filtering step serves several important purposes.

First, many bacterial gene entries in NCBI, including coding sequences (CDSs) and open reading frames (ORFs), originate from automated annotation pipelines that classify any sufficiently long ORF as a “gene.” As a result, sequencing or assembly artifacts (e.g., misassembled contigs or repetitive regions) can yield spurious ORFs that (without supporting RNA-seq or proteomic evidence) remain hypothetical and may never correspond to true expressed proteins.

Second, mis-assembly, contamination, and mis-annotation remain pervasive issues in bacterial genome data, frequently leading to the prediction of false “genes” that disappear following resequencing or improved assembly. In several cases, re-sequenced genomes show a substantial reduction in annotated genes due to the removal of contaminant sequences (most commonly originating from human or laboratory sources). Recent studies have demonstrated that draft bacterial assemblies can contain hundreds of contaminant contigs derived from foreign species, resulting directly from sequencing and assembly errors [16]. Similarly, human contamination has been shown to generate thousands of spurious protein-coding predictions in bacterial assemblies, which were subsequently propagated into public databases [17].

More recently, the NCBI FCS-GX cleanup effort removed thousands of contaminated prokaryotic genomes and reported that most contaminant sequences (and consequently, most erroneous genes) were short in length [18]. It was also reported that predicted orphan genes are often shorter than non-orphan genes [19]. This overlap between short sequence length and higher error susceptibility underscores why many predicted orphan genes may, in fact, reflect annotation or assembly artifacts rather than true novel genes.

Because the current databases are not free of false positives, predicted small orphan genes could include spurious entries. Step 3 helps minimize the likelihood of such false positives. Although this filtering may eliminate a small number of true genes, ongoing genome sequencing efforts will enable future corrections and refinements. Overall, Step 3 resulted in the exclusion of only seven potential orphan genes from the entire dataset, representing a minimal reduction in the total number of candidates (Fig 2b).

Step 4

To ensure that no distant homologs were missed in our initial screens, Step 4 employed a reciprocal BLASTp analysis [20,21] with an E-value cutoff of 1E-3. The E-value represents the expected number of alignments with a given score that could occur by chance when searching a query against the database; thus, a threshold of 1E-3 retains only hits with ≤ 0.001 probability of arising randomly. This commonly used cutoff [22] was also applied in the TRGdb database for gene identification [13], allowing direct comparison between studies. We deliberately used the weakest Step 3 match to cast a wider net for remote homologs for our reciprocal BLAST. Any gene that returned a hit with an E-value cutoff of 1E-3 to another organism in this reciprocal search was removed from our candidate list. This extra layer is vital; without it, highly divergent genes or those with low-complexity regions can masquerade as true orphans. Applying this strategy further refined our dataset, leading to an even greater reduction in orphan genes (Fig 2b).

Steps 1–4

Through our screening method, the four intracellular pathogens we examined (M. tuberculosis, H. pylori, L. pneumophila, and C. trachomatis) had reduced orphan genes, but this decline was smaller compared to bacteria from broader environmental niches (Fig 2b). The percentage decline in orphan gene count for bacteria with a broad environmental habitat ranged from 90% to 100%. In contrast, for four intracellular pathogens, the decline was less, ranging from 28% to 66%.

We do not know why intracellular pathogens had a lower decline. This could be due to their isolated niches, which limit opportunities for horizontal gene transfer (HGT) and genome rearrangements, as well as the smaller number of available sequenced genomes. Intracellular bacteria are known to undergo reductive evolution [23,24], a process where non-essential genes are lost while highly specialized genes are retained, resulting in greater genetic stability. In contrast, bacteria inhabiting diverse environments are more prone to gene acquisitions and losses via HGT, leading to frequent changes in their orphan gene status. Although the isolation of these pathogens may contribute to the observed trend, we argue that the limited availability of sequenced genomes is likely the primary factor, as even restricted genomes showed fewer orphan genes as more genome sequencing data became available.

General Trend: Increasing Genome Sequencing Reduces Orphan Genes. We investigated how the number of available genomes in NCBI influences the count of orphan genes across bacterial species. Overall, species with a greater number of sequenced genomes tended to exhibit fewer candidate orphan genes. The 10 bacterial species analyzed were ranked in descending order based on the number of available sequenced genomes (NCBI) and then divided into two groups (Fig 3, Left). The top five species, with the highest number of sequenced genomes, had the fewest candidate orphan genes. However, the bottom five species, with fewer sequenced genomes, showed the highest number of candidate orphan genes (Fig 3, Right). This pattern suggests that the number of sequenced genomes strongly affects orphan gene identification, likely because increased sampling (more sequenced genomes) improves annotation accuracy and enables broader genomic comparisons, leading to the reclassification of genes previously labeled orphans.

Download:

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

Fig 3. A Higher Number of Sequenced Genomes Generally Leads to a Reduction in Orphan Genes.

The left panel demonstrates the number of orphan genes in 2023 and 2025, alongside the number of sequenced genomes available for each bacterial species in NCBI (S1 Table, individual data points without log transformation). Labels: xx: data points corresponding to a single orphan gene (Log 1 = 0); x: data point corresponding to zero orphan genes (Log 0 = undefined). The 10 bacterial species studied were divided into 2 groups and arranged in descending order based on the number of genomes available (NCBI). The right panel compares the top five High Genome Count species to the bottom five Low Genome Count species using a violin plot.

https://doi.org/10.1371/journal.pone.0338891.g003

Earth harbors immense microbial diversity with an estimated 10E12 microbial taxa [25], yet scientific attention is disproportionately focused on organisms of medical or industrial relevance. The majority of microbial taxa are uninvestigated. Indeed, half of all microbiology publications center on just 10 species (E. coli dominates the literature), and only 26% of known bacterial species have been studied, leaving over 90% vastly underinvestigated [26]. This uneven research landscape highlights the significant gap in scientific knowledge in microbial research and helps explain why poorly sampled species generate more apparent orphan genes.

While the number of sequenced genomes is a crucial factor, there are other factors that can contribute to the differences observed in the data. For example, E. faecalis and M. tuberculosis have similar numbers of genomes available in NCBI, yet they differ markedly in the number of candidate orphan genes identified in each species. We do not know why this is, but it could be due to the global prevalence and extent of genomic sampling. This is highlighted best by the World Health Organization (WHO) 2024 tuberculosis (TB) report. This report shows that much of the data on M. tuberculosis exhibits a limited geographical distribution, with high prevalence in specific regions. More than 66% of worldwide TB cases originated from just eight countries: India, Indonesia, China, Philippines, Pakistan, Nigeria, Bangladesh, and the Democratic Republic of the Congo [27]. However, M. tuberculosis is not confined to these regions, nor to human hosts. For instance, M. tuberculosis has been isolated from free‐ranging wildlife in South Africa [28] and from wild chimpanzee in West Africa [29]. The focus on human clinical isolates from select locales likely limits the detection of homologs in more diverse strains, thereby inflating the number of candidate orphan genes. In contrast, E. faecalis, a remarkably versatile organism, is commonly found as a commensal in the gastrointestinal tract of humans and animals [30], and is also frequently isolated from extra-enteric environments such as blood, urine, soil, water, plants, and a wide range of fermented foods (particularly in traditional dairy products) [31–34]. Furthermore, it can also be an opportunistic pathogen [30]. This widespread distribution promotes its inclusion in a broad range of studies, spanning microbiome research to clinical, environmental, and food-related research. Hence, comprehensive geographical and ecological sampling may be responsible for strengthening genome annotation and thereby leading to fewer candidate orphan gene identifications in E. faecalis.

Given the immense, underexplored microbial diversity, we predict two key trends:

1. New genes classified as orphans are likely to emerge over time.
2. As more genomes are sequenced, many genes once labeled as orphans will be reassigned, perpetuating a cycle of orphan designation and subsequent reclassification. To break this cycle, we must revise both our classification criteria and our terminology.

Revising the Terminology for Orphan Genes with Descriptors That Better Reflect Current Understanding. Our results demonstrate that genes once labeled as “orphan” (or taxonomically restricted) often lose their apparent uniqueness as more genomes are sequenced and annotated. Following the precedent established for newly proposed bacterial taxa, which are designated as “candidate species,” we suggest adopting similar provisional descriptors, such as “candidate” or “putative,” for newly identified genes lacking known homologs. These descriptors more accurately convey the provisional and potentially temporary nature of their apparent novelty. Using the term candidate is clearer and better reflects the underlying uncertainty when no homologs have yet been detected in nature. Another useful term is cryptogenic, a word commonly used in medicine to describe conditions of obscure or unknown origin, which is particularly apt as a descriptor for genes whose evolutionary origins remain uncertain. Regardless of whether the term is orphan, taxonomically restricted, cryptogenic, or some other term, we argue that a provisional descriptor is required.

The biological origins of these genes help clarify why a more flexible terminology is warranted. Such genes could have emerged through de novo gene birth (evolution from noncoding DNA) or through horizontal gene transfer (acquisition from another organism). While orphan genes have often been assumed to evolve unusually [35–37] rapidly, our proposed adjustment in terminology acknowledges an alternative explanation: their apparent uniqueness may stem not from accelerated evolution, but rather from insufficient sequence sampling. We also outline five additional reasons to employ descriptive qualifiers when referring to orphans, emphasizing the need for terminology that captures both uncertainty and the potential for future revisability.

1. Bioinformatic Artifacts: Many predicted genes might not be genuine but are artifacts of computational algorithms. The descriptor, such as candidate, emphasizes the provisional nature of these findings, recognizing that further evidence would be required to confirm their uniqueness.
2. Incomplete Genomic Sampling: Our understanding of genetic diversity is still tremendously inadequate. The genomes sequenced to date represent only a minute fraction of the total diversity on Earth. It is estimated that the Earth is home to approximately one trillion bacterial species. Fewer than one million bacterial genomes have been sequenced, with the majority belonging to the same species [38]. A recent study reported that, despite analyzing approximately 1.5 million microbial genomes, including Metagenome-Assembled Genomes and more than 18,000 metagenomic samples, a large fraction of prokaryotic diversity remains unsequenced, with about 42% of bacterial and 36% of archaeal phylogenetic diversity still lacking genomic representation [39]. Considering predictions of up to 1 trillion microbial species and that the NCBI database currently houses approximately 330,000 prokaryotic genomes, representing around 19,000 species, only roughly 0.000002% of prokaryotic genomes have been sequenced. Based on these figures, the unclear origin of many bacterial genes is primarily attributable to insufficient sequencing data, rather than the genes being truly unique.

In addition, bacterial phages (viruses), known reservoirs for gene transfer [40], are vastly underrepresented in the genome databases. Even though phages likely outnumber bacteria by a factor of ten or more [41], fewer than 5,000 phage genomes are cataloged in NCBI (a 2023 report) [42]. Moreover, phage genomes contain an exceptionally high proportion of novel genes with uncharacterized functions, indicating that much genetic diversity remains unexplored [41]. It has been argued that phages possess some of the greatest genetic novelty in the biological world, with as many as 80% of their encoded genes lacking similarity to known proteins and being of unknown function [41]. Thus, the challenge in tracing a gene’s origin is often more indicative of limited sequencing efforts than true genetic uniqueness.

1. Limited Domain Coverage: Our study focused solely on bacteria. However, similar issues exist for archaea and eukaryotes. For example, the Genome Taxonomy Database has 584,382 bacterial genomes versus 12,477 archaeal genomes (a 2024 report) [43], suggesting that orphan designations in archaea are likely due to under-sampling. In eukaryotes, approximately 41,000 genomes (from about 2,300 species) have been assembled (a 2024 report) [44], yet an estimated 8.7 million eukaryotic species exist worldwide [45]. This indicates that our current data represents only about 0.026% of eukaryotic genomes in nature. Thus, the inability to trace a gene’s origins in archaea and eukaryotes is often more a reflection of limited sequencing rather than true uniqueness.
2. Extinctions Lead to Limited Knowledge of Gene History: It is estimated that approximately 99.9% of all species that have ever lived on Earth are now extinct [46]. The vast majority of species went extinct before the advent of DNA sequencing. Tracing the evolutionary history of genes from extinct organisms is an extreme challenge. This gap in our data particularly affects genes that arise via de novo birth or horizontal gene transfer, since the donor species may be extinct. Thus, the inability to trace a gene’s origins is often more a reflection of the lack of sequenced genomes from extinct species rather than a true indication of uniqueness.
3. Implications for Evolutionary Interpretations: Overestimating de novo gene birth rates can mislead our understanding of gene evolution and, beyond science, fuel misconceptions about the origin of species. The assumption that orphan genes must have evolved extremely rapidly has even been used to argue for irreducible complexity, an argument often cited by creationists. Such conclusions rest on two flawed assumptions: first, that orphan genes will remain orphan indefinitely, and second, that any gene without a currently identifiable homolog must be unique. The limited scope of our sequencing data undermines both assumptions. Our study demonstrates that many genes previously classified as orphan genes do not retain that status indefinitely. We also showed that many genes, previously considered unique due to their apparent lack of homologs in external lineages, have since been shown to share homology, indicating they are not truly lineage-specific.

We initially retrieved 385 protein sequences from TRGdb for 10 bacterial species, and after completing all four BLASTp steps, 79 sequences remained (S1 File). However, four predicted proteins from K. pneumoniae were excluded from further analysis because the genome for this bacterial species was not fully assembled, which limited the ability to validate the presence of candidate orphan genes through this approach. As a result, genes from K. pneumoniae were not studied further. Of the 385 initial protein sequences, only 75 remained for further evaluation, representing an 81% reduction in potential orphan genes in just two years.

Overall, adopting provisional descriptors, such as candidates for newly identified genes lacking known homologs, would more accurately convey the provisional and potentially temporary nature of their apparent novelty. This change highlights the uncertainty in gene origin due to incomplete genomic data rather than inherent uniqueness. By adopting a change in terminology, we can foster clearer communication about what is known and what remains to be discovered in the evolution of genes.

Prioritizing Candidate Orphan Genes for Experimental Studies. Our analyses identified 75 candidate orphan genes, but the only definitive way to confirm whether these sequences represent true genes is through experimental validation. This next step focuses on prioritizing the most promising candidates for follow-up studies. While we highlight the strongest candidates here, it is important to note that those excluded from this subset may still encode functional genes; they were simply ranked lower based on the current computational evidence.

In this analysis, we focused exclusively on protein-coding genes, as these are generally easier to assay experimentally than RNA-encoding genes (Fig 4a). AlphaFold [47,48] served as an initial filter to help identify candidate orphan genes whose protein sequences are capable of folding into stable structures. It provides an average confidence score referred to as pLDDT (predicted Local Distance Difference Test), which classifies predictions into four categories: very low (<50), low (50–70), high (70–90), and very high (>90) confidence. For our analysis, “very low” and “low” were grouped as low confidence, and “high” and “very high” as high confidence. The majority (56%) of candidate orphan protein sequences exhibited low structural confidence. Only 15% depicted high confidence (Fig 4b), while 29% of candidates did not produce any structural prediction (Fig 4b).

Download:

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

Fig 4. Prioritizing Candidate Orphan Genes for Experimental Validation.

a. Workflow for prioritizing candidate orphan genes based on predicted protein structure, promoter elements, mRNA expression, and ribosome-binding site (RBS) strength. b. AlphaFold structural predictions for 75 candidate orphan proteins. Shown are representative models from each confidence tier: one with a high average predicted Local Distance Difference Test (pLDDT) score (90.56) and one with a low average pLDDT score (30.22). 10 proteins from each category (high confidence, low confidence, and no prediction) were selected for downstream analyses. c. Promoter motif search using BPROM and BioCyc. Excluding Mycobacterium species, every candidate orphan protein across all confidence tiers contained at least one predicted promoter site. d. mRNA expression profiling based on RNA-seq datasets retrieved from the SRA, confirming transcript presence (but not expression level) for selected candidate orphan genes. e. Translation Initiation Rate (TIR) estimates. High-confidence candidate orphan genes exhibited the highest mean TIR, while the no-prediction group showed the lowest average TIR.

https://doi.org/10.1371/journal.pone.0338891.g004

One reason for the high number of candidate orphan proteins with low confidence in structure could be correlated with the GC content. Codon usage is highly influenced by GC content in candidate orphan genes, which in turn plays a vital role in shaping the properties of proteins [49]. High GC genomes are likely to encode proteins that are enriched with amino acids such as alanine, proline, and glycine [49]. These amino acids tend to promote disorder, contributing to structural flexibility [50,51]. Consequently, candidate orphan proteins from genomes with a high GC proportion have often been identified as Intrinsically Disordered Proteins (IDPs). Such proteins do not possess a fixed three-dimensional (3D) structure in solution and exist in dynamic, flexible conformations, particularly in the absence of other proteins [52–54]. It has been revealed that low pLDDT scores might indicate the intrinsic disorder nature of proteins. Large-scale structural annotation of the human proteome using AlphaFold demonstrated that regions labeled as IDPs overlap with low-confidence predictions [55]. AlphaFold employs a structure-based algorithm that relies on sequence homology to known proteins; therefore, IDPs, due to their structural plasticity, are poorly predicted by such algorithms [55–57]. Although low-confidence AlphaFold predictions do not preclude candidate orphan genes from encoding real proteins, high-confidence models enable us to prioritize those candidates most likely to succeed in in vitro or in vivo experiments.

Following AlphaFold predictions (Fig 4b), 10 proteins from each category (high confidence, low confidence, and no result) were selected randomly to further investigate their likelihood of being functional genes (Fig 4c-e). This analysis included checking for the presence of start and stop codons, promoter prediction, assessing mRNA expression, and evaluating translation initiation efficiency by analyzing ribosomal binding sites. All the sequences had predicted start and stop codons; however, due to the high likelihood that these codons might occur by chance, we further examined the upstream sequence of the start codon for potential promoters using BPROM.

BPROM [58] predicted promoter regions for all candidates except those from M. tuberculosis (Fig 4c). This software is based on the Sigma 70 promoter motif, and according to the BPROM website, it has ~ 80% accuracy in predicting E. coli promoters. In contrast, M. tuberculosis does not rely on the classical Sigma 70 promoter but instead utilizes alternative sigma factors such as SigA, SigB, and SigE, which in turn recognize different promoter motifs. These motifs often diverge from the canonical −10 and −35 sequences targeted by BPROM [59–61]. Consequently, BPROM is not well-suited for identifying promoters in M. tuberculosis, so we also provided data where it is excluded (Fig 4c). We attempted to predict promoters for M. tuberculosis using BioCyc, but were only able to predict promoters for some of the M. tuberculosis genes. Consequently, promoter screening appears to have limited utility as a primary tool for M. tuberculosis promoters, which contribute to the majority of orphan proteins (of the 10 species we probed) and represent a special case due to their unique genomic and regulatory features.

To evaluate transcriptional activity of our candidate proteins, we queried the Sequence Read Archive (SRA) [62], a comprehensive repository of public RNA-seq datasets spanning diverse organisms and conditions. We used tBLASTn to detect matching mRNA reads for each protein. Quantifying read abundance by reconstructing RNA-seq reads across hundreds of millions of data points from various experiments and research groups is computationally challenging; therefore, we adopted a more qualitative approach. We examine the expression of each gene from three randomly selected SRA datasets to determine whether each gene exhibits detectable transcriptional evidence. Remarkably, all proteins with high-confidence structural models (100%) were supported by SRA reads, whereas 80% of low-confidence proteins and 90% of those lacking structural predictions showed evidence of transcription (Fig 4d). These results demonstrate that high structural confidence is strongly predictive of mRNA expression and thus provides a powerful criterion for prioritizing candidates for experimental validation.

To assess which candidate orphan genes had potential for high translational rates (protein production rate at the ribosome), we estimated them using the Ribosomal Binding Site (RBS) Calculator [63]. On average, proteins with high structural confidence exhibited the highest predicted Translation Initiation Rates (TIRs) (Fig 4e), suggesting that these proteins are more likely to be efficiently translated. These findings further support their potential for experimental validation. In summary, our integrated analysis, which combines structural prediction, transcriptional evidence from the SRA database via tBLASTn, and translation initiation efficiency, underscores the importance of considering both structural confidence and expression data when prioritizing candidate orphan genes for experimental validation. Proteins with high structural confidence not only demonstrate reliable expression but are also predicted to be more efficiently translated, making them prime candidates for further investigation.

Based on our findings, AlphaFold and the RBS calculator are particularly effective for identifying promising experimental candidates. Both tools are relatively fast, reliable, and can be easily automated. In contrast, the presence of start and stop codons was not informative because all gene-predicting programs are based on these features. Promoter prediction was less reliable for non-model organisms, such as M. tuberculosis. In addition, the distance a promoter is from a gene varies among organisms, making predictions often less useful, especially for non-model organisms, where promoters have been less studied. While SRA data can provide valuable expression data, analyzing it is time-consuming, making it less practical for large-scale screening. Furthermore, the SRA data alone provide a binary response (yes or no in the presence of RNA), without considering the quality or level of the RNA, without in-depth analysis.

Through this work, we identified 10 high-confidence candidate orphan genes for experimental validation and assessed their predicted subcellular localizations. Using bioinformatic tools, we classified 1 candidate as a membrane protein, 4 as soluble proteins, 3 as secreted proteins, and 2 as signal-peptide-only fragments (S2 Table). The occurrence of signal-peptide-only sequences is particularly intriguing, as it suggests these peptides may represent remnants of larger proteins from which downstream domains were lost through genetic mutation. Further in silico and experimental searches for additional homologs could test this hypothesis by uncovering the missing protein segments. Collectively, these findings imply that the 2 signal-peptide-only sequences may not constitute genes with functions, but rather vestiges of proteins that are no longer functional in these bacteria.

Implications, Future Work, and Unanswered Questions. This study utilized the TRGdb database, which was not constructed with all available genomes for each species at the time. As a result, our analysis was limited to individual genomes rather than considering all strains within a species. Developing an updated database that integrates a broader set of genomes, combined with the comparative framework established in this study, could further refine the estimated number of candidate orphan genes. Such an expansion would likely increase the detection of homologs present in additional strains or closely related taxa, thereby providing a more comprehensive and accurate view of orphan gene diversity.

A major challenge we encountered in this study was the frequent absence of clearly reported gene names or sequences in the literature. While many studies report the percentage of orphan genes identified, they often fail to provide a list of the specific genes (by name or gene ID), even in supplementary materials. In some instances, links to supporting data were nonfunctional, displaying messages such as “Page Not Found,” “This site can’t be reached,” or similar errors [3,64–67]. For this analysis, we relied on TRGdb [13], which provides a well-organized and searchable resource that facilitates gene identification and retrieval for downstream analyses. Nevertheless, the lack of detailed, gene-level information in previous studies remains a major barrier to reproducibility and comparative research, limiting the ability to validate reported findings.

Similarly, while analyzing bacterial promoters, we found that no single software tool was sufficient for accurate predictions across all species, requiring the use of multiple tools. This highlights the need for a more comprehensive bioinformatics tool capable of predicting promoter locations using genome sequences across a wide range of bacterial species.

Another significant challenge was analyzing gene expression data from the SRA. Although large-scale RNA-seq datasets are available, identifying accession numbers that match the exact strain used in our study proved difficult. Due to feasibility constraints, we limited our analysis to a small subset of accessions at a time. Additionally, our analysis focused solely on determining whether the candidate orphan gene sequences were transcriptionally active without taking into account expression levels, experimental conditions, false negatives in the data, contamination issues, or potential regulatory influences.

A further issue arises from the way BLAST/NCBI handles result ordering when excluding a specific organism. For example, in our E. coli screening pipeline, Step 1 requested the top 100 hits while excluding E. coli, yielding no hits from any other species. In contrast, Step 2 used the top 5,000 hits with E. coli excluded and did return matches from other organisms. This discrepancy occurs because BLAST first compiles all potential hits (the top were E. coli sequences), and only afterward removes those belonging to the excluded taxon. In Step 1, the first 100 results were entirely E. coli, so once those were filtered out, nothing remained. By expanding the search to 5,000 hits in Step 2, BLAST was able to gather enough non–E. coli matches to populate the filtered set.

While this workaround served our immediate needs, it has inherent limitations. As genomic databases continue to grow at an exponential pace, it will soon be common to see more than 5,000 high‐scoring E. coli sequences for a given gene. At that point, even requesting 5,000 hits will still capture only E. coli, and any more distantly related homologues, though present, will be excluded because they never enter the initial pool of top matches. In principle, one could raise the cutoff to 50,000 or beyond to ensure coverage of distant homologs, but NCBI currently caps BLAST results at 5,000 entries. To make candidate orphan gene identification and other evolutionary analyses more robust and accessible to the broader research community, BLAST/NCBI will need to either raise this 5,000‐hit ceiling or implement alternative filtering strategies that avoid this queueing limitation. A change needs to be implemented soon because the NCBI genomic repository is growing at an extraordinary pace. Alternatively, researchers may be able to circumvent this issue by using tools such as DIAMOND [68] or MMseqs2 [69], which support larger datasets and are not limited by this result.

In selecting candidate orphan genes for experimental validation, we prioritized proteins with high-confidence structural predictions from AlphaFold. These candidates also exhibited strong transcriptional evidence, including the presence of mRNA, promoters, and higher average Translation Initiation Rates. These indicators suggest that these genes are likely to be expressed and translated into functional proteins, making them suitable starting points for further characterization. However, it is essential to note that the absence of a confident AlphaFold prediction (or the other analysis we performed) does not necessarily imply that a gene is non-coding. Our strategy focuses on identifying candidates that are most likely to yield experimental validation; we do not exclude low-confidence or no-result candidates as legitimate protein-coding genes. The limitations we encountered in candidate selection mirror broader challenges in microbial genomics, particularly the uneven taxonomic representation of sequenced species.

A recent study determined that among the studied bacterial species, a disproportionate amount of research focuses on just 10 species, with more than 90% of bacteriology articles examining less than 1% of all species [26]. This limited representation affects our understanding of candidate orphan genes. In other words, as we sequence more genomes, we are likely to uncover additional orphan genes as genomic datasets from poorly studied or underrepresented species often do not contain sufficient annotation to recognize homologous genes, leading to an overestimation of orphan genes in earlier studies; however, with continued sequencing, genes once considered unique to a specific lineage may no longer be viewed as distinct.

We currently lack comprehensive experimental evidence confirming that candidate orphan genes represent genuine protein-coding sequences; however, our analysis has identified several promising candidates. Functional characterization of these genes would substantially advance the field, as experimental validation is essential to determine whether these sequences encode true proteins or merely reflect misannotated gene fragments.