Detect de novo expressed ORFs in transcriptomes with DESwoMAN

Abstract

De novo gene emergence refers to the process by which new genes arise from mutations in previously non-coding genomic regions. Investigating early stages of de novo gene emergence is essential for understanding the mechanisms that enable gene formation from scratch. Several workflows have been developed to detect and study the early stage of de novo gene emergence in specific species. However, no software currently exists that can automatically identify and characterize novel, unannotated open reading frames from a transcriptome, and analyse their mutations and fixation patterns within or across species. We introduce DESwoMAN (De novo  Emergence Study With Outgroup MutAtioNs), a software tool designed to (i) detect newly expressed non-canonical ORFs (neORFs) in transcriptomes, (ii) filter neORFs with no homology to outgroup genes, and (iii) search for syntenic sequences homologous to neORFs in outgroup genomes (and optionally transcriptomes) and analyse mutations in coding features between these sequences. We applied DESwoMAN with two different strategies to three setups. Our results highlight the tool’s capabilities and demonstrate its potential for elucidating the early stages of de novo gene emergence. DESwoMAN is available at https://github.com/AnnaGrBio/DESWOMAN. It is implemented in Python3 and comes with a docker image on DockerHub for easy installation and execution including all dependencies.

Introduction

De novo gene birth is the process by which a non-genic region acquires genic features by mutation [1–11]. According to the model proposed by Carvunis et al. [12], the emergence of genes from scratch follows two main steps: First, a genome acquires by mutation an open reading frame (ORF) and transcription. Assuming the newly expressed ORF (neORF) can be translated, it is qualified as a proto-gene. Such proto-genes are assumed to evolve first through neutrality, a stage that is likely very dynamic, as several studies have demonstrated a high turnover in gain and loss of recently gained new transcripts and ORFs within a species [13, 14]. However, proto-genes can become subject to selection pressure [15–18]. This can lead some proto-genes to become fixed and functional to a species. Such genes are then called de novo genes.

Young de novo genes can be detected in species or phylogenetically restricted groups [19–22]. However, validating a de novo gene emergence and understanding the underlying mechanisms remains a methodological challenge. To infer the de novo emergence status of annotated genes in a genome, several bioinformatic tools have been developed [19–21, 23–36]. All these pipelines start with annotated genes and aim to validate which of them has emerged de novo in the species of phylogenetically restricted groups.

However, the earliest stages of such genes are typically missed by traditional gene annotations. First, because most proto-genes are not fixed in a species and do not necessarily appear in reference annotations. Second, because genome annotation pipelines use gene homology or known genic features (e.g. conserved protein domains, splice sites, coding sequence length, and exon–intron structure) to annotate genes in new genomes [37]. However, youngest de novo genes neither have detectable homology to other genes, nor exhibit known genic features and are therefore missed by such an approach.

To detect early stages of genes and validate their de novo emergence status, several studies have used transcriptomes to search for de novo genes (defined as coding genes or not) [2, 3, 8, 10, 10, 13, 38–47]. Although existing pipelines are highly comprehensive and detailed, reproducing their results typically requires advanced computational skills and extensive decision-making at multiple stages of a long and complex process [11]. This is largely because these workflows rely on the integration of numerous software tools, custom intermediate scripts, and subjective analytical choices. Moreover, most current approaches do not explicitly examine the mutational events that drive the transition from a non-coding sequence to a newly emerged ORF (de novo ORF). Such analyses require the extraction and comparison of syntenic, non-genic homologs across closely related outgroup genomes, which require high computational skills. Finally, methodological strategies differ substantially between studies, further complicating systematic comparisons [48].

In this paper, we define a neORF as a transcribed ORF that does not overlap annotated canonical genes. An neORF is classified as de novo when comparative genomic analyses indicate that it originated through mutations in previously non-coding genomic regions.

Here, we present DESwoMAN (De novo  Emergence Study With Outgroup MutAtioNs), a standardized and fully automated pipeline designed to automatically detect neORFs based on transcriptomes, validate their de novo status, and extract syntenic homologous regions to neORFs from outgroup genomes. Based on the extracted syntenic homologous sequences, DESwoMAN identifies different mutations responsible for the coding or non-coding status of a sequence using inter-genome comparisons, within the same species or between closely related species. The algorithmic method is based on Grandchamp et al. [46], but includes a series of modifications and has been fully automated for end users.

We conceptualize de novo gene emergence as a continuum of molecular states, defined by the progressive acquisition of genic features by non-coding sequences. Within this framework, different stages can be distinguished based on the degree of genic features acquired and their evolutionary persistence. Proto-genes correspond to intermediate stages of this continuum: they are expressed ORFs that show evidence of persistence and may begin to experience selective pressures, but whose functional role remains limited [49] and which are not yet fixed within species. Here, we specifically focus on the earliest stage of this process, corresponding to genomic regions that acquire transcription (e.g. polyadenylated transcripts) and contain an ORF, but lack evidence of evolutionary constraint or functional integration. We refer to these elements as neORFs, which may be transient and are not necessarily fixed. Inter-genome comparisons facilitate the study of de novo gene emergence across evolutionary scales. These categories are defined operationally to facilitate empirical investigation. The specific criteria can be adjusted depending on the biological system and available data, and the pipeline implements a standardized procedure based on these definitions.

Materials and methods

The DESwoMAN software is an automated pipeline designed to detect and characterize the emergence of de novo neORFs within a single workflow (Fig. 1). It implements two complementary strategies, referred to as Strategy 1 and Strategy 2, which share a common conceptual framework but differ in their input requirements and analytical objectives. A detailed description of both strategies and all parameters is provided in the manual (Supplementary Data).

Figure 1.

Flowchart of the DESwoMAN methodology.

In both strategies, DESwoMAN operates through three main stages, although the third stage differs substantially. In the first stage, unannotated transcribed ORFs are identified. In the second stage, the absence of homology with known sequences is assessed. In the third stage, Strategy 1 identifies syntenic homologs (if present) in outgroup genomes and characterizes the mutations distinguishing coding from non-coding states. This strategy requires independently assembled genomes to ensure accurate genomic context and synteny inference. In contrast, Strategy 2 groups neORFs into orthogroups based on their presence across multiple transcriptomes mapped onto a common reference genome, allowing comparisons across individuals or conditions without requiring multiple genome assemblies.

DESwoMAN is designed to detect early stages of de novo gene emergence. Strategy 2 is particularly suited for identifying neORFs that are variably expressed across individuals or populations and are therefore not fixed, whereas Strategy 1 enables validation of de novo status through comparative genomics across species. These two strategies thus address complementary aspects of de novo gene emergence across different evolutionary contexts.

The definition and validation of de novo emergence remain subjects of debate [11]. While we strongly recommend running all three stages of DESwoMAN to robustly assess the de novo status of neORFs, any ORF detected in the first stage that lacks annotation in the query genome is initially classified as a neORF. The potential de novo status of these candidates is subsequently refined through the following stages.

Identification of de novo neORFs—Strategy 1

Running DESwoMAN has both mandatory and optional parameters. The first mandatory input is the annotated genome and corresponding transcriptome of the organism of interest (the query). The genome enables the detection of novel ORFs in the organism, while the transcriptome allows the determination of whether these ORFs are transcribed. The type of transcriptome used can vary depending on the user’s research objectives. The second mandatory input consists of one or, preferably, multiple annotated genomes from outgroup species or populations, referred to as target genomes. These target genomes are used to identify syntenic homologs of the neORFs detected in the query genome and to determine whether these homologs are coding or non-coding. If the homologs are non-coding, this suggests that the coding status was acquired in the query genome, confirming its de novo emergence and enabling the study of potential mutations responsible for this acquisition. Moreover, optionally, the user may also provide transcriptomes corresponding to the outgroup genomes, to study the transcriptional status of these homologs.

Optionally but highly recommended, the user can provide a dataset of protein, DNA, or RNA sequences from outgroup species. These datasets are used for homology searches. Only neORFs detected in the query genome that lack homology to known genes from other genomes (preferably the entire NR database) are retained as neORFs. This Stage remains optional, and the user may choose to perform the homology search independently at a later stage.

From the starting point, DESwoMAN automatically executes the following steps:

1. Stage1. Based on user-defined criteria, all transcripts from the query transcriptome exceeding the specified expression threshold are selected for downstream analysis (Fig. 1A).

2. The selected transcripts are then filtered to retain only those derived from user-specified genomic regions (intergenic, intronic, and antisense), chosen from several predefined options (Fig. 1B).

3. In the selected transcripts, all ORF(s) corresponding to a protein in a range of 30–3000 amino acids are detected according to various criteria defined by the user. Redundant ORFs originating from overlapping transcripts are detected to retain only a single representative ORF. Reverse ORFs are removed (Fig. 1C). By default, ORFs start with an ATG codon; however, users can specify an alternative start codon.

4. In case several ORFs are retained per transcripts, one or several of them can be selected according to user criteria (Fig. 1D).

5. DESwoMAN then extracts the genomic coordinates of the selected ORFs, along with the coordinates of any introns they may contain (Fig. 1E).

6. The resulting set of ORFs represents the neORFs. Their spliced and unspliced sequences are extracted. At the end of this first Stage, DESwoMAN generates an initial file storing a set of information about the neORFs. If a protein or DNA/RNA database has been provided for homology filtering, the software automatically proceeds to Stage 2 (Fig. 1F).

7. neORFs are filtered to retain only those with no detectable homology to known proteins or DNA/RNA sequences in the supplied database (Fig. 1G). Protein homology searches are performed using DIAMOND [50], while DNA/RNA homology searches are performed using BLASTn. The stringency of the homology searches is defined by the user, with default parameters set to the more sensitive mode in Diamond and a P-value threshold of 0.01 for BLASTn. The more comprehensive the dataset, the more effective this filtering Stage becomes. Following this Stage, the list of neORFs is reduced according to the homology results and options selected. If no database is provided, DESwoMAN skips Stage 2. Stage 3 is executed automatically. In this Stage, DESwoMAN gathers information to enable the user to assess whether the candidate neORFs have arisen de novo.

8. To do this, DESwoMAN searches for canonical genes surrounding each query neORF, or overlapping it depending on the genomic location of the ORF (Fig. 1H). The number and distance of the canonical genes depend on the user’s selection.

9. DESwoMAN then searches for homologous genes of these canonical genes in the outgroup genomes with reciprocal homology hits (Fig. 1I). DESwoMAN extracts proteins from annotated genes using GffRead [51] and validates homology through reciprocal BLASTp searches with DIAMOND using the default fast mode.

10. When canonical genes are detected in outgroup genomes and if synteny is conserved, DESwoMAN retrieves the corresponding genomic region in each outgroup and searches for sequences homologous to the neORF within these regions (Fig. 1J). Unspliced neORFs are used as queries for a BLASTn search against outgroup genes. The homologous sequences are detected with BLASTn (e-value 0.01, other parameters correspond to the BLASTn default settings.

11. When homology is detected, DESwoMAN realigns each homologous outgroup sequence with the neORF, removing any intronic regions from the alignment. DESwoMAN performs a pairwise sequence alignment with PairwiseAligner() function from Bio.Align package from biopython [52]. The PairwiseAligner is used in global mode, implementing a Needleman–Wunsch-type dynamic programming algorithm with affine gap penalties. Matches were assigned a score of +1 and mismatches a score of 0, while gap-opening and gap-extension penalties were set to −10 and −1, respectively. The software then evaluates whether the homologous region satisfies a set of criteria that indicate whether it also represents a transcribed ORF (Fig. 1K). The criteria include: The presence of a complete start codon aligned with the neORF start codon, the presence of a complete stop codon aligned with the neORF stop codon, the presence or not of premature stop codons, the existence of Frameshift mutations and their positions, the position and number of substitutions, evidence of transcription, and a score of conservation developed by Wacholder et al. [17]. This score basically measures the similarity between the neORF and its corresponding homolog, based on the amino acid similarity. Any frameshift, for instance, drastically decreases the similarity score between the two proteins encoded by the ORFs. At the end of Stage 3, DESwoMAN summarizes all this information in one results file.

Based on the selected criteria, the user can then identify which neORFs have likely arisen de novo. For instance, neORFs that were detected in Stage 1 as non-annotated, retained in Stage 2 because they show no homology to any known coding genes, and that have syntenic homologs in outgroup genomes which lack coding properties, can reasonably be considered de novo. The genomic coordinates and sequences of neORFs syntenic homologs are also provided, facilitating downstream analyses. When introns are present in the sequence (both in neORFs and their homologs), they are placed in lowercase in the sequence output. DESwoMAN provides additional scripts to apply a more stringent filter for de novo origin on the output neORFs. This allow assessing whether the most distant detected homolog in an outgroup species is noncoding based on the detected mutations.

Compare neORFs expression—Strategy 2

The second strategy is designed to identify neORFs across different transcriptomes, rather than between genomes, with all transcriptomes mapped onto a common reference genome. This approach also enables assessment of whether a neORF is preferentially expressed under specific conditions or in particular tissues. Stages 1 and 2 of DESwoMAN are largely similar to those in Strategy 1, with some notable differences.

First, instead of a single query transcriptome, multiple transcriptomes (at least two) are provided by the user. All query transcriptomes must have been assembled via mapping to one single supplied reference genome. Second, no outgroup genomes are required. DESwoMAN will generate groups of neORFs sharing the same genomic locus that are expressed across the various transcriptomes. Finally, during Stage 1, neORFs are identified within each query transcriptome, and the Stage 1 output file encodes the name of the transcriptome in which each neORF was detected. Stage 2, if executed, applies filtering across the entire set of identified neORFs.

The main modifications in Strategy 2 occur during Stage 3. Here, the objective is not to identify the mutations responsible for the emergence of newly transcribed ORFs, as all transcripts are assembled by mapping to a single reference genome, which precludes mutation analysis. Instead, the goal is to cluster transcribed neORFs into groups sharing the same genomic locus across transcriptomes. To achieve this, DESwoMAN groups identical ORFs (Fig. 1K’) detected across transcriptomes while verifying the consistency of their genomic origin. Although coordinate-based approaches (e.g. overlap of genomic intervals) could be used to identify candidate shared loci, such methods alone may group distinct ORFs with partially overlapping coordinates. By combining sequence identity and genomic localization, DESwoMAN ensures that neORFs are clustered according to shared genomic loci rather than reflecting duplicated or ambiguously mapped sequences. Upon completion of Stage 3, DESwoMAN produces an output file listing groups of neORFs co-expressed across multiple transcriptomes.

DESwoMAN user parameters

Throughout the three main Stages of the software, the user selects from a range of options provided by DESwoMAN. These choices are specified prior to execution, either via a graphical user interface or a configuration file.

DESwoMAN implementation

DESwoMAN is implemented in Python 3. It requires three non-Python dependencies—BLAST [53], gffread [51], and diamond [54]—for which reason we provide a docker image on DockerHub to run DESwoMAN directly with all dependencies installed. A more detailed documentation of both strategies and all parameters of DESwoMAN can be found in the manual (Supplementary Data) and in the readme of the github repository (https://github.com/AnnaGrBio/DESWOMAN).

Use cases: human and fruit fly neORFs

To illustrate the use of DESwoMAN, we collected/assembled 17 transcriptomes and applied it to biological datasets using three different setups.

First setup: HumanSetupS1: With Strategy 1, we use the human reference genomes (GRCh38.p14 from the NCBI RefSeq database [55]) to search for neORF emerged de novo in human. neORFs are identified in human RNA-seq from brain (ID hsa br F 1 SRR306838) [56] from the NCBI Sequence Read Archive [57], since brain RNA seq was available for all target species and has not yet been extensively studied compared to testis. Comparative analyses of mutations distinguishing coding from non-coding states were performed using the reference genomes from NCBI RefSeq assembly of Bonobo (GCF_029289425.2), Gorilla (GCF_029281585.2), Chimpanzee (GCF_028858775.2), Macaque (GCF_003339765.1), Orangutan (GCF_028885625.2), and mouse (GCF_000001635.27), along with their corresponding assembled brain transcriptomes (Supplementary Data S1.1) from [56]. For Stage 2 (filtering based on lack of homology to known sequences in outgroups), we constructed two separate datasets: one containing protein sequences and another containing non-coding RNA (ncRNA) sequences. Both datasets include sequences from several non-mammalian outgroup species (Supplementary Data S1.1 and Supplementary Table S1).

Second setup: DrosoSetupS1: Using Strategy 1, we focused on identifying youngest neORFs that are not yet fixed in a species. We searched for neORFs in a line of Drosophila melanogaster collected in Finland (FI) by the European Drosophila Population Genomics Consortium. The neORFs were identified from whole-body RNA-seq data and a line specific genome reported by Grandchamp et al. [46], with sequencing, assembly, and mapping details provided therein. To assess whether these neORFs emerged de novo specifically in Finnish population, we performed comparative analyses of mutations distinguishing coding from non-coding states using six outgroup D. melanogaster populations (DK: Denmark; ES: Spain; SE: Sweden; UA: Ukraine; TR: Turkey; ZI: Zambia), based on their corresponding genomes and whole-body transcriptomes also described by Grandchamp et al. [46]. For homology searches, we constructed two distinct datasets—one containing protein sequences and the other containing ncRNA sequences—each including data from several outgroup Drosophila species (Supplementary Data S1.1 and Supplementary Table S2).

Third setup, HumanSetupS2 : Using Strategy 2, we analysed five human RNA-seq libraries [56] from distinct tissues—brain, cerebellum, heart, kidney, and testis—which served as query transcriptomes (Supplementary Data S1.3). The same human reference genome as in HumanSetupS1 was used to assemble all five transcriptomes, with reads mapped separately to generate independent assemblies for each tissue. neORFs were identified within each transcriptome and subsequently grouped into groups of neORFs sharing the same genomic locus according to their tissue-specific expression. For the homology searches performed by DESwoMAN, we constructed two datasets: one containing protein sequences and another containing ncRNA sequences. The list of species included in these datasets is provided in (Supplementary Data S1.3 and Supplementary table S3).

The assembly and strategy applications for setups can be found in Supplementary Data S1.2. Computational performance was assessed using the GNU time utility. For DrosoSetup1: DESwoMAN required approximately 4.9 GB of peak memory and achieved partial parallelization, with an effective CPU usage of 4.8 cores. The analysis of the benchmark dataset completed in 19 min and 35 s on a standard Linux workstation. For HumanSetup1: DESwoMAN required approximately 10.5 GB of peak memory and achieved partial parallelization with an effective CPU usage of 7 cores. The analysis of the human benchmark dataset completed in 5 h and 41 min on a standard workstation.

Results

Identifying neORFs with DESwoMAN

We developed DESwoMAN, a software to detect neORFs and assess their de novo status. In order to ascertain the practical benefits of the software and to gain novel insights about de novo emergence mechanisms, we apply DESwoMAN in this study to three different setups with two different strategies.

Strategy 1 allows to detect neORFs in a genome with a corresponding transcriptome, and study the mutations that could have led to their emergence by searching syntenic homologues in outgroup genomes. The HumanSetupS1 serves as a dataset to study fixed neORFs in Homo sapiens in comparison to other mammalian species, while the DrosoSetupS1 serves as a dataset to study neORFs present in one or several lines of D. melanogaster but not yet fixed in the species.

HumanSetupS2 serves as a dataset for Strategy 2 that allows to detect neORFs in several transcriptomes from one single species, and study common expression.

Detecting mutations in neORFS with Strategy 1

We studied the conservation of coding feature of human neORFs in outgroup species (HumanSetupS1), as well as the coding feature of neORFs from a line of D.mel from Finland in other lines (DrosoSetupS1), with DESwoMAN’s Strategy 1. After the first Stage, DESwoMAN detected 905 intergenic neORFs in humans, which do not correspond to any canonical genes (Supplementary Data S4.1 and Supplementary Fig. S3). After the second filtering Stage, which removes neORFs homologous to known proteins or ncRNAs, 7 neORFs were eliminated, resulting in a total of 898 neORFs. In Finnish Drosophila, 1668 intergenic neORFs were identified, 1562 of which show no known homology in existing databases.

For 63.44% to 71.76% of the 1562 D. melanogaster’s neORFs, DESwoMAN detects syntenic homologous sequences in the genomes of the other lines of D. melanogaster (Table 2 and Supplementary Data S4.2, Supplementary Figs S3–S9). We find 68.37% to 92.2% sequences homologous and syntenic to human neORFs in the six different primates genomes. In mice, however, only 1.0% of the neORFs have syntenic homologous sequences.

Table 2.

Number of neORFs and their syntenic homologous sequences

Sample/species	# syntenic homologs	% syntenic homologs
Query : FI	1562	–
DK	991	63.44
ES	1121	71.76
SE	1060	67.86
UA	1069	68.43
TR	1038	66.45
ZI	1013	64.85
Query: human	898	–
Chimpanzee	827	92.09
Bonobo	828	92.20
Gorilla	819	91.20
Orangutan	757	84.29
Macaque	614	68.37
Mouse	9	1.00

As expected, coding features of syntenic homologous sequences exhibit reduced conservation with increased phylogenetic distance in the HumanSetupS1 (Fig. 2), while it’s quite homolog withing Drosophila lines. Start and stop codons are, on average, fairly well conserved within syntenic homologous sequences (https://zenodo.org/records/14936107) with an exception in the far-related mouse. Transcription is more little between D. melanogaster populations, ranging from 29.52% to 31.79% of syntenic homologous sequences being fully transcribed in the same orientation. The transcription status of syntenic sequences homologous to human neORFs is lower in mammalian species, ranging from 4.88% to 25.39%, excluding mouse. Conservation scores based on frameshift were high on average in all fruit fly (71.64%–76.48%) and range from high to low in outgroup species of humans with increasing phylogenetic distance (86.14%–52.12% excluding mice).

Figure 2.

Conservation of coding features in syntenic homologous sequences. The X-axis represents the target species or lineages in which homologous neORFs were identified, and the Y-axis represents the percentage of conservation of different features between query and target sequences. Each dot represents, among all neORFs with a syntenic homolog in the considered target species, the percentage of sequences exhibiting a given feature: dark blue, percentage of syntenic homologs with the presence of a conserved start codon; blue, percentage of syntenic homologs with the presence of a conserved stop codon; yellow, percentage of syntenic homologs with the presence of a premature stop codon specific to the target sequence; grey, percentage of syntenic homologs with the fully transcribed target sequence. The cyan dots indicate the conservation score, defined as the average frameshift conservation score between neORFs and their syntenic homologous sequences. The red dots represent the substitution rate, corresponding to the average percentage of substitutions in syntenic homologous sequences compared to their respective neORFs.

DESwoMAN provides mutation information so that users may interpret and validate the de novo status of neORFs according to their criteria. For instance, in Drosophila, 50 neORFs have syntenic homologs detected in all lineages, exhibiting a canonical ATG start codon, a stop codon, no frameshifts, no premature stop codons, and complete transcription across all lineages. Based on these criteria, these 50 neORFs should be considered not de novo, or at least fixed across all lineages, while all others are not fixed in the species and emerged likely through mutations. However, the number of neORFs shared by all lines increases to 667 when only the absence of frameshifts is considered (frameshift probability score >90%).

In humans, none of the neORFs were detected as complete across all other species. However, in a simple comparison with the chimpanzee, 107 neORFs share all coding attributes between the two species, indicating that these neORFs are not de novo, or that their emergence occurred in the common ancestor of the two species.

DESwoMAN applies a synteny window to detect syntenic homologous sequences of intergenic neORFs, which can be adjusted/configured by the user. To determine synteny, homologous conserved genes are used as anchors and the window size specifies how many of these anchors up- or downstream of the neORF are considered. Moreover, syntenic homologs can be detected by using simple reciprocal BLAST hits, or by using best reciprocal BLAST hits. Simple reciprocal BLAST hits refer to any two sequences that are found through a BLAST search, while best reciprocal hits are a stricter subset with only the top-scoring sequence of the respective BLAST search.

To illustrate the impact of having different options offered by the software, we investigate whether the size and flexibility of the synteny window for intergenic neORFs affects the number of syntenic homologous sequences that are detected (Fig. 3, https://zenodo.org/records/14936107). We test five different synteny window sizes on the two setups, as well as a no-synteny option. With simple reciprocal hits (Fig. 3a and c), we detect more syntenic homologous sequences of human neORFs than of fruit fly neORFs in outgroup genomes. In both setups we observe that the larger the synteny window is, the higher is the number of syntenic homologous sequences. In fruit fly we find on average 37% of neORFs in other fruit fly samples with a synteny window of 1, which increases to an average of 71% with a synteny window of 5 (Fig. 3a). In the human setup, neORFs have on average 55% of syntenic homologous sequences in outgroup species (excluding mice) with a synteny window of 1, and this increases to 87% on average with a synteny window of 5 (Fig. 3c, https://zenodo.org/records/14936107).

Figure 3.

Impact of synteny window size and homology validation. (a) DrosoSetupS1 with simple reciprocal hits, (b) DrosoSetupS1 with best reciprocal hits, (c) HumanSetupS1 with simple reciprocal hits, and (d) HumanSetupS1 with best reciprocal hits.

For both HumanSetupS1 and DrosoSetupS1, the absence of synteny criteria resulted in nearly 100% (96%–97% for fruit flies and 86%–98% for human outgroups) of detected syntenic homologous sequences, except in mice where 14% of neORFs are found as syntenic homologous sequences in the genome. In the HumanSetupS1 we observe that the greater the phylogenetic distance is, the fewer syntenic homolgous sequences were detected for all synteny windows (Fig. 3c, https://zenodo.org/records/14936107).

With best reciprocal hits (Fig. 3b and d), as with simple reciprocal hits, we observe that the larger the synteny window is, the more syntenic hits are detected. However, while the percentage of syntenic homologous hits remains almost unchanged for all fruit fly outgroups for any synteny window https://zenodo.org/records/14936107), a much more pronounced decrease was observed for human outgroups. For example, compared to simple reciprocal hits, the number of homologous hits observed for a window size of 2 with simple reciprocal hits shows an average of 83% between outgroup species, excluding mice, whereas this average percentage of syntenic homologous hits drops to 21% with best reciprocal hits. In the HumanSetupS1, the phylogenetic distance influences the amount of homologous hits in the same way.

Analysis of tissue-specific neORF expression with Strategy 2

We applied DESwoMAN’s Strategy 2 to HumanSetupS2 to detect neORFs and assess how their transcription differs across tissues. Three independent groups of neORFs are investigated: intergenic, intronic, and antisense neORFs.

Across all the studied tissues, the same pattern of neORF frequency has been identified: antisense neORFs are most numerous, followed by intergenic ones, and intronic ones are the least numerous (Table 3). We find most neORFs across all three categories in cerebellum and fewest neORFs in testis.

Table 3.

neORFS per transcriptomes

Transcriptome	# transcripts	genomic positions	# new transcripts	ORFs	neORFs
Human brain	32 851	Antisense	5920	11 677	3931
		Intergenic	1483	2080	905
		Intronic	488	635	299
Human cerebellum	48 810	Antisense	9774	16 216	6183
		Intergenic	4479	5940	2766
		Intronic	2272	3013	1326
Human heart	25 313	Antisense	4699	7658	2991
		Intergenic	539	615	311
		Intronic	442	432	231
Human kidney	32 664	Antisense	5932	8595	3629
		Intergenic	969	1170	533
		Intronic	448	461	241
Human testis	18 231	Antisense	2431	3817	1593
		Intergenic	288	326	173
		Intronic	90	82	46

All identified neORFs are classified into groups of neORFs sharing the same genomic locus by DESwoMAN (Fig. 4). Across all three neORF categories (antisense, intergenic, and intronic), the vast majority of neORFs is tissue-specific, with a consistent decline in the number of groups of neORFs sharing the same genomic locus shared by a higher number of tissues.

Figure 4.

Tissue-specificity of neORF sharing the same genomic locus in human.

Discussion

DESwoMAN: automated neORF detection and analysis

In this study we present DESwoMAN, a tool developed to automatise the detection of neORFs based on transcriptome data and identify their de novo status. Our tool offers a high flexibility regarding the addressable biological questions by providing two different strategies and user-adjustable parameters to cover a wide range of use cases. In the three different use cases presented in this study, we gain biological insights about the earliest stages of de novo gene emergence through inter-species comparisons, as well as intra-species comparisons at the population level and regarding tissue-specificity of neORFs.

DESwoMAN reports multiple downstream analyses results next to the identified neORFs, such as syntenic homologous sequences and properties of the neORFs and their syntenic homologous sequences in the form of mutations associated with the coding status. However, the biological interpretation of the reported results remains a crucial task to be carried out by the user. For example, if all syntenic homologous sequences of a neORF reported by DESwoMAN share the same coding status, there is a high likelihood that the neORF does not represent a de novo gene, but rather a conserved gene, which is not yet annotated. It is important to note in this context that the definition, detection and validation of a potential de novo status is already highly variable across studies [5, 7, 21, 35]. DESwoMAN validates neORFs by identifying syntenic non-coding homologous sequences, which serve as a baseline for confirming a de novo emergence in most of the studies in the field.

DESwoMAN provides a wide range of customizable parameters, as summarized in Table 1, allowing users to customize the analysis to specific datasets and research questions. Key parameters, such as the TPM expression threshold, synteny window size, and homology search settings, can significantly influence both the number and confidence of identified de novo gene candidates. For example, higher TPM thresholds increase stringency by excluding lowly expressed transcripts, which may reduce the total number of neORFs detected but improve overall confidence. Similarly, the choice of synteny window size affects the detection of homologous regions: smaller windows may miss distant homologs, whereas larger windows can increase the risk of false positives. Homology search parameters, including BLAST/DIAMOND sensitivity and identity thresholds, further influence the balance between sensitivity and specificity.

Table 1.

DESwoMAN optional parameters

Stage	Parameter	Options
Stage 1	Minimum TPM threshold	Users can define the expression threshold for transcripts to be retained by DESwoMAN
	Filter transcripts overlapping with TEs	When TEs have been masked in the query genome(s), transcripts overlapping TEs can be optionally excluded.
	Remove transcripts with unknown orientation	In certain transcriptome assemblies, some transcripts are annotated with “.” to indicate that their transcriptional orientation could not be determined by the assembly software. These transcripts can be optionally excluded by the user.
	Genomic position of candidate neORFs	Users can select to analyse transcripts located in intergenic, intronic, antisense, or genic transcripts, or any combination of these categories.
	Remove genic splice variants	When a novel transcript represents a splice variant of a known gene, it can be optionally excluded from the analysis.
	ORFs to retain in candidate transcripts	When multiple ORFs are identified on the sense strand of a transcript, users can choose to retain all ORFs or select based on the first start codon, highest Kozak score, or longest ORF.
	Minimum 5′UTR and 3′UTR length	Users can specify minimum UTR lengths, and ORFs with shorter 5′ or 3′UTRs will be excluded from selection.
Stage 2	Homology search	Depending on whether homology searches are conducted against a protein or DNA/RNA database, users can specify the BLAST and/or DIAMOND parameters to be applied.
Stage 3 Strat 1	Determine homology between annotated genes	During the search for orthologs of canonical genes flanking or overlapping a neORF in target genomes, users can opt for either reciprocal best hits or all hits.
	Synteny window for homology search	Users can define the size of the synteny window within which to search for neORF homologs in target genomes.
	Maximum sequence fraction for detecting premature stop codons	When analysing neORF homologs, users can set the maximum fraction of the homologous sequence to examine for premature stop codons.

Default parameter values were selected based on the most commonly used settings in de novo gene detection, with particular emphasis on transcript expression thresholds [11, 35]. Users are encouraged to adjust these parameters according to the species under study, evolutionary distance, and specific research objectives. Detailed guidance on parameter customization is provided in the Supplementary Manual.

Furthermore, many other different metrics can be used to assess coding status conservation, such as ancestral reconstructions [58] or phylogenetic tools combined with protein sequence homology detection [43]. DESwoMAN evaluates six features derived from sequence alignments to assess the coding potential of candidate neORFs. Other potentially relevant features, such as splice site conservation, translation evidence, or mutations affecting polyadenylation signals, are not currently considered by the pipeline. In particular, DESwoMAN assumes conservation of splice sites in homologous sequences and infers homologous structures based on aligned intronic regions. This assumption represents a limitation, as splice site conservation may not hold across all evolutionary contexts. To address this, DESwoMAN reports the genomic coordinates of syntenic homologs, enabling users to extract neORFs and their corresponding homologous sequences for downstream analyses using alternative conservation metrics or validation approaches.

Apart from varying definitions and validation of a potential de novo and coding status, e.g. [5, 7, 21, 35], several factors can influence the reliability of the results. If a low number of target genomes is used, for example, a missing coding status in all syntenic homologous sequences of an neORF could be misinterpreted, if a syntenic homologous coding ORF exists in other genomes, not included in the study. A higher number of well-chosen query and target genomes or transcriptomes can therefore lead to more robust conclusions.

Comparison to other methods

To our knowledge, no other software comparable to DESwoMAN currently exists. Nevertheless, several studies have already investigated de novo genes detected in transcripts. For example, [39] searched for fixed de novo genes in human using several transcriptomes. Transcriptomes from six human tissues also from [56] were merged, and intronic and intergenic neORFs were detected. Their de novo emergence was validated by the absence of homology to proteome of other species. A total of 1512 intergenic ORFs were discovered, compared to 1562 with DESwoMAN after the second Stage. The orders of magnitude between the two studies are comparable, though our third Stage demonstrate that most of the neORFs we detected are not fixed in the species. Stage 3 allowed to show that 107 of these neORFs have a syntenic homolog with all coding features in chimpanzee, and can no longer be considered as human specifics.

Transcript data have also been used to investigate the emergence of de novo genes at the population level. For example, in 2022, [59] searched for de novo genes in the accessory glands of D. melanogaster, using RNA-seq data from the accessory glands of six lines. They detected 49 intergenic and 84 intronic de novo genes, when we detected 50, based on whole body transcriptomes. However, the two methodologies differ in several respects; notably, [59] assembled their transcriptomes de novo, and homology was assessed for the entire set of transcripts, not only the ORFs. Using a workflow similar to Cridland’s, [60] identified 90 de novo genes by studying transcriptomes from 29 lines of D. melanogaster expressed in accessory-gland RNA-seq data.

Strategy 1: neORF properties across species and populations

In order to demonstrate the functionality of DESwoMAN, we applied Strategy 1 to the HumanSetupS1 for comparison across species and to the DrosoSetupS1 for comparison across populations within a species.

Syntenic homologous sequences of neORFs are on average well detected across all tested species or populations, and exhibit low substitution rates when compared to their corresponding neORFs. However, most of them (94.5%–100%) lack at least one coding feature in at least one of the target genomes, confirming with more power an emergence from scratch through mutation of the considered neORFs. The most common missing feature is transcription, with >70% of syntenic homologous sequences being not or not fully transcribed. This pattern is observed in both Strategy 1 setups and supports a high turnover in transcription gains and losses [13, 61]. Furthermore, this finding underscores the role of transcriptional activation in the early stages of gene emergence [47], even though some of these cases might be false negatives as de novo gene could have low expression levels [38].

Another general pattern observed is a lower number of conserved coding features in syntenic homologous sequences as phylogenetic distance from the query species increases, suggesting a low fixation rate of coding features over evolutionary time. Interestingly, start and stop codons are not conserved in 10%–20% of cases, and frame conservation scores are around 80%, while substitution rates remain very low in these syntenic homologs (approximately 1% of sequences). This apparent contrast—low overall sequence divergence but reduced conservation of coding features—may indicate that coding-related elements are more prone to disruption or are subject to distinct selective pressures. These observations are consistent with previous studies suggesting that newly emerged coding features may experience weak or transient selection during early stages of fixation [62–65].

Impact of synteny window size

We investigate the impact of synteny window size and BLAST hit detection regarding the accuracy at detecting syntenic homolog sequences of intergenic neORFs. Across all tested setups, larger synteny windows result in the detection of more syntenic homologous sequences. Several evolutionary mechanisms can make a larger window size necessary to find the surrounding, homologous genes, such as genome reshuffling, inversions, duplications, gene losses, horizontal gene transfer, and others [66, 67]. However, there is a trade-off between window size and accuracy. Increasing the window size can increase the number of identified sequences, while decreasing the accuracy at the same time by adding false positives.

This is best seen in the drastic increase in homologous sequences within the DrosoSetupS1 when no synteny criterion is applied. Without an applied synteny criterion, nearly all neORFs in both setups exhibit homologous sequences in almost all target species. Applying synteny windows of different sizes influences the number significantly, suggesting that several of the identified homologous sequences detected with large windows could be false positives or the result of a bigger genomic reshuffling. For example, neORFs can be very small [68–70] and sometimes be associated with transposable elements (TEs) [71, 72], in such a way that without synteny window, an ORF can be detected several times in a genome. We examined the overlap between Drosophila neORFs and TEs (Supplementary Data S3, Tables S4 and S5). We found that 718 of the 1562 neORFs overlapped at least one TE in D. melanogaster. Moreover, between 68% (UA) and 71% (DK, ZI) of the neORFs lacking any syntenic hit were TE-overlapping. The extensive mobility of TEs in D. melanogaster and their substantial overlap with neORFs [72] may therefore contribute to the low level of synteny conservation observed. However, after removing neORFs overlapping TEs in both the human and Drosophila datasets (Supplementary Data S3 and Supplementary Figs S1 and S2), frameshift conservation scores as well as start/stop codon conservation in homologous sequences were not substantially modified. These results suggest that TE insertions alone do not explain the lower conservation observed among Drosophila lines compared with mammalian genomes. Several non-mutually exclusive mechanisms may contribute to this pattern. In particular, the large effective population size of Drosophila species is associated with high levels of standing genetic variation and increased efficacy of selection, leading to elevated levels of nonsynonymous polymorphism within populations [73]. In addition, Drosophila genomes are characterized by abundant single-nucleotide polymorphisms (SNPs) and substantial insertion/deletion turnover [74, 75], which may further reduce frameshift conservation at the intra-species level.

Moreover, several studies have shown that de novo genes at early stages can undergo duplication [46], leading them to be detected at different genomic locations. Additionally, little is known about duplication in non-genic regions [76, 77], which may contribute to the high rate of detected homology across genomes.

Comparing the simple reciprocal BLAST hits with only the best reciprocal BLAST hits, also highlights the importance of a suited homology detection method. The percentage of syntenic homologous sequences of neORFs identified through simple reciprocal BLAST hits is higher in target species of the HumanSetupS1 than in populations of D. melanogaster in the DrosoSetupS1. This finding is in contrast to the assumption that there should be a higher similarity between populations of the same species than between different species. However, as stated earlier, the significant overlap with TEs of neORFs without synteny might play a huge role in this output. Moreover, comparing these numbers to the best reciprocal BLAST hits, we observe that the percentage of syntenic homologous sequences drops drastically in the HumanSetupS1, while the results remain almost unchanged in the DrosoSetupS1. This finding confirms that reciprocal BLAST hits recover members of larger gene families, while best reciprocal BLAST hits recover better the corresponding orthologous sequences [78, 79]. Since mammalian gene families are, on average, larger than Drosophila gene families [80, 81], this hypothesis could explain the observed patterns in the results between the HumanSetupS1 and DrosoSetupS1.

Our findings therefore emphasise the need for suited synteny methods to validate homologous sequences of neORFs, fitting the used input data and phylogenetic setup. Full genome alignments can provide a more comprehensive approach for synteny detection [17]. However, whole genome alignments are not applicable for every phylogenetic setup and costly or difficult to get (therefore, we go a middle way). DESwoMAN assesses synteny by using annotated genes as anchors for intergenic neORFS. Again, the quality of the results is highly dependent on the quality of annotations. However, alternative anchors become available for assessing synteny [82]. Nevertheless, our results show that our method identifies a large proportion of syntenic homologous sequences well.

Regardless of the applied synteny method or window size or usage of best reciprocal BLAST hits, our results confirm that the phylogenetic distance is one of the most important factors for the detection and validation of neORFs and their corresponding homologous sequences. In the mouse, as a far-related species, for example, we could detect homologous sequences for only 18% of human neORFs even without a synteny criterion applied.

Strategy 2: tissue-specificity of neORFs in human

Another use case of DESwoMAN is to investigate the tissue-specificity of neORFs, which is implemented in an automatised way through Strategy 2. As a biological use case we utilize a different Human Setup (HumanSetupS2) with transcriptomes from five different tissues. We observe a universal pattern across all neORF categories (antisense, intronic, and intergenic) identifying the vast majority of them to be tissue-specific with a decreasing number of neORFs shared between an increasing number of tissues. This result suggests that transcription of new ORFs is a more complex event and therefore rare compared with the emergence of other coding features in an ORF. This result aligns with previous findings by Wacholder et al. [17] or Grandchamp et al. [13], who show that de novo transcript gain and loss is a highly dynamic process, with fast gain and loss processes. Understanding the gain, loss, and fixation of new transcription events presents a crucial challenge for a deeper understanding of neORF gain and fixation. DESwoMAN helps investigating the birth of neORFs independently of their fixation status. Therefore, our tool can help to advance our understanding of the earliest stages of de novo gene emergence and, with its transcriptomic focus, facilitate insights into the role of transcription in this process. Additionally, our study detects fewer neORFs in testis compared to other tissues. To further investigate this observation, we normalized our analysis by considering the proportion of neORFs relative to (i) the total number of transcripts and (ii) the number of de novo transcripts (Supplementary Data S4.3 and Supplementary Table S6). These results confirm that testis exhibits a lower proportion of neORFs compared to other tissues. This appears inconsistent with previous studies reporting that testis is a hotspot for de novo gene emergence [41, 45, 83]. However, it is important to note that our analysis specifically focuses on neORFs, which likely represent early-stage protogenes, whereas most studies in the literature describe annotated or more evolutionarily established de novo genes. Several technical factors may also contribute to this observation. In particular, the high complexity of the testis transcriptome may affect transcript assembly and reduce the detection of short and lowly expressed ORFs. In addition, the filtering criteria applied in Strategy 2 may preferentially exclude such candidates. Finally, we note that in the context of Strategy 2, the de novo status of candidates is not validated through comparative genomics.

Taken together, our results do not contradict the established role of testis in de novo gene emergence, but rather suggest that different stages of this process may be captured differently depending on the methodological approach.

Limitations and further improvments

While DESwoMAN was designed to minimize conceptual biases at each step, several methodological limitations should be acknowledged.

First, the reliability of input data can be influenced by various factors. For example, genomes with high TE content are more challenging to analyse [84] as they may result in a higher incidence of incorrectly mapped or misassembled transcripts. Transcriptome assembly errors cannot be fully excluded, particularly for lowly expressed transcripts. This issue has been highlighted in other systems, including plants, where high TE content can interfere with accurate transcript reconstruction [85]. To mitigate this issue, default parameters include expression-based filtering to retain consistently detected transcripts across samples. In addition, ORF prediction is restricted to well-supported transcript structures, thereby reducing the likelihood of artifactual ORFs.

Second, the absence of homology is assessed exclusively using sequence-based approaches. As a result, structural homology is not explicitly evaluated, which may lead to an underestimation of distant homologs. More specifically, the reliance on reciprocal BLAST searches may introduce false negatives. For instance, de novo genes exhibiting weak or partial similarity to unrelated sequences in outgroup species may be incorrectly discarded. This limitation reflects a common trade-off in homology-based methods between sensitivity and specificity, where stringent criteria reduce false positives at the cost of potentially excluding true candidates.

Third, in Step 3 of Strategy 1, synteny is inferred using neighbouring genes rather than whole-genome alignments. Although genome-wide alignments would provide a more robust framework to identify homologous regions, such resources are not consistently available across species and can be computationally demanding. Future developments of DESwoMAN could integrate genome alignment-based approaches when suitable data are available, further improving the accuracy of homology inference.

Overall, these limitations should be considered when interpreting the results, particularly with respect to the balance between sensitivity and specificity in de novo gene detection.

Fourth, DESwoMAN currently relies on genome assemblies and associated annotations as input data, which may limit its direct applicability to standard population genomic datasets typically distributed as variant call format (VCF) files. While the use of genome assemblies enables access to full genomic context, including syntenic relationships and large-scale structural variations, such information is not readily available from VCF data alone. In particular, the identification of syntenic homologs in Strategy 1 requires precise genomic positioning and the ability to detect structural events such as inversions or TE insertions. As a result, DESwoMAN is best suited for datasets where high-quality genome assemblies are available. Future developments could extend the pipeline to incorporate VCF-based approaches, for example by reconstructing individual-specific sequences prior to analysis.

Fifth, DESwoMAN currently restricts ORF detection to canonical ATG start codons and imposes a conservative minimum ORF length of 30 amino acids. While these choices reduce false positives and ensure high-confidence identification of neORFs at a genome-wide scale, they also intentionally exclude potentially functional short ORFs and micropeptides [86, 87], as well as ORFs initiating at alternative start codons [88–90]. Such elements have been increasingly reported in the literature, including in de novo genes and small UTR-associated ORFs, but their detection often requires experimental evidence or specialized computational methods. Consequently, DESwoMAN may underestimate the full diversity of emerging ORFs, reflecting a deliberate trade-off between sensitivity and specificity. Future versions of the pipeline could provide options to explore shorter ORFs or alternative initiation codons, allowing users to investigate these additional classes of candidate neORFs.

Finally, DESwoMAN relies on pairwise sequence alignments between each neORF and its corresponding homologous sequence to infer substitutions and frame-shift events. While this approach allows direct comparison and maintains computational efficiency, it does not account for information from multiple sequences simultaneously, as would be achieved with multiple sequence alignment (MSA) or phylogeny-aware methods. In addition, the choice of alignment parameters, particularly gap opening and extension penalties, can influence the inferred gap patterns and, consequently, the detection of frame-shift events. Although the current implementation uses parameter settings designed to limit the introduction of multiple short gaps, alternative alignment strategies, such as codon-aware or phylogeny-aware approaches, could improve the accuracy of evolutionary inference, especially for sequences evolving under weak selective constraints. Future developments of DESwoMAN could explore these approaches to enhance robustness.

Conclusion

Understanding how de novo genes emerge—from initially non-coding sequences to functional genes—is a critical yet still poorly understood aspect of genome evolution.

These genes often encode proteins with entirely novel functions, making them particularly relevant not only for understanding evolutionary processes, but also for applications in biotechnology and protein engineering. A deeper understanding of these early molecular events could illuminate the evolutionary origins of gene families, how life on earth emerged and evolved and support the rational design of novel proteins in synthetic biology.

Until now, research on de novo gene evolution has largely focused on genes that have already been fixed within a species, typically exhibiting stable transcription and complete coding features. However, the earliest stages, where non-coding sequences first gain transcriptional activity and gradually acquire coding potential, remain largely unexplored.

To address this gap, we developed DESwoMAN, a tool designed to detect and validate neORFs from transcriptome data and analyse the mutations involved in their emergence by comparing them with identified syntenic homologous sequences. To our knowledge, DESwoMAN is the first fully automated pipeline capable of performing these analyses based on transcripts, enabling systematic investigations into the initial steps of de novo gene formation.

DESwoMAN is highly customisable, allowing users to define analysis strategies and parameters according to specific research questions and dataset characteristics. In this study, we applied DESwoMAN to three distinct use cases, demonstrating its utility for investigating neORF evolution at the species, population and tissue-specific level and highlight biologically relevant factors and the impact of specific parameters.

DESwoMAN is intended to serve the scientific community as a valuable resource for gaining deeper insights into the early molecular mechanisms underlying de novo gene emergence and for supporting future work in evolutionary genomics and synthetic biology.

Supplementary data

Supplementary data is available at NAR Genomics & Bioinformatics online.

Conflict of interest

The authors declare no conflict of interest.

Funding

A.G. and M.L. were supported by the Deutsche Forschungsgemeinschaft priority program “Genomic Basis of Evolutionary Innovations” (SPP2349; Project No. 503272152 awarded to John Parsch [PA 903/12-1] and Erich Bornberg-Bauer [BO 2544/20-1]). M.L. was also funded by DFG project no. 431626672 awarded to Erich Bornberg-Bauer. A.G. was supported by Aix-Marseille University, and by Human Frontier Science Program Research Grant RGP004/2023 (doi.org/10.52044/HFSP.RGP0042023.pc.gr.168590) awarded to Erich Bornberg-Bauer, Anne-Ruxandra Carvunis and Christine Brun. E.D. was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation)—503348080.

Data availability

DESwoMAN is available in https://github.com/AnnaGrBio/DESWOMAN and code source in https://zenodo.org/records/22146842. All result and output files, assembled mammalian transcriptomes, as well as Drosophila genomes and transcriptomes, can be found in https://zenodo.org/records/14936107 (version 3 is corrected for Drosophila, as it contains the same genomes, transcriptomes and results as for Versions 1 and 2, but with the correct IDs of Drosophila). A supplementary manual and document are in linked archived files.