RNA-seq datasets

To explore the effects of intra-condition count variation on the detection of differentially expressed transcripts using real data we used both intra and inter-study datasets. At an intra-study level, we combined publicly available RNA-seq data from brain tissue (prefrontal cortex) of 12 tame and 12 aggressive red foxes (S1 Table) generated within the same study [36]. For the inter-study case, we combined multiple publicly available RNA-seq datasets from several dogs and six wolves [32–35], also derived from brain tissue (frontal cortex, cerebral cortex, prefrontal cortex and frontal lobe) (S1 Table). In relation to the latter, dogs 1 to 6 and wolves 1 to 6 were derived from Albert et al., (2012), dog 6 from Roy et al., 2013, dog 7 (two replicates) from Fushan et al., 2015 and dogs 8 and 9 (three replicates each) from Hoeppner et al., (2014) as described within the table. All the samples were downloaded from the National Center for Biotechnology Information (NCBI) and the European Bioinformatics Institute (EMBL-EBI), covering a wide range of ages, both sexes as well as multiple replicates and sequencing strategies (S1 Table). We selected these specific case studies because, firstly we were interested in evaluating the effects of intra-condition count variation both at intra and inter-study levels and the domestic dog, being a model organism, has an available high-quality reference transcriptome as well as several high-quality RNA-seq datasets generated across different studies; while secondly, we sought to perform a brief exploratory inter-study scan to investigate if the domestication of both dogs and foxes has resulted in the co-expression of a set of common brain genes, relative to their “wild/aggressive” type, since behavioural modifications are considered to have been the first target in domestication [42].

Reads from all samples were mapped to the dog reference transcriptome [32], which contained 26,107 annotated transcripts (Ensembl CanFam3.1, release 92) [43], using Bowtie v.2.3.4.1 [44]. We did not use a splice-site aware mapper, such as Tophat2 [45] or HISAT2 [46], since introns were not expected to be present within the reference transcriptome. Reads were not mapped to the dog genome as our aim was to explore the effects of intra-condition variation on a predefined set of reference transcripts, and not infer novel transcripts from these previously published datasets. We did not use the fox reference transcriptome (available on Ensembl) for the mapping the fox datasets within our overall analysis as we were interested in directly comparing differentially expressed transcripts from the same underlying reference set identified using the fox, dog and wolf datasets. We did however map all of the fox datasets to the fox reference transcriptome in order to confirm that the proportion of reads mapped was similar to that when using the dog reference transcriptome. The pileup.sh script from the BBmap package [47] was used to obtain per transcript abundance estimates (measured by the number of mapped reads to the corresponding transcript) in each sample. Read counts from technical replicates of “Dog_8” and “Dog_9” were averaged and merged into one file (S1 Table), while read counts from the two biological replicates of “Dog_7” were treated separately. Finally, to confirm that the final read count numbers were reliable relative to the dog reference transcriptome, all read datasets from foxes, dogs and wolves were re-mapped using the pseudo-mapper kallisto v0.46.1 that implements a rapid and accurate kmer search based strategy for estimating transcript abundance counts [48]. For each dataset an r² correlation value was calculated describing the linear correlation between the per-transcripts counts obtained following Bowtie2 mapping and the corresponding abundance counts obtained using kallisto.

Software

TVscript requires as input: (1) multiple files containing per transcript read counts (one per sample), (2) a file containing the lengths of the transcripts that the reads were mapped to, (3) a percentile threshold value for intra-condition variation and (4) a configuration file that indicates the locations all files as well as the condition allocations of the count files in (1). An example configuration file along with further details is available from the SourceForge project page. The steps that TVscript implements to identify transcripts associated with high levels of disjointed read counts are (see Fig 1 for a workflow): i) each input dataset, containing count values from a particular sample, is allocated to either condition A or B, as indicated within the configuration file; ii) counts are normalized by dividing them by the length of the corresponding reference transcript and by the sum of all counts for that sample; (iii) for each reference transcript (t), the absolute pairwise differences between normalized read counts across all samples within condition A are calculated; (iv) the corresponding variances are calculated; (v) steps (iii) and (iv) are repeated for condition B; (vi) variance scores from each condition are placed in ascending order and associated with corresponding percentiles; (vii) reference transcripts are removed (or filtered) if their variance score is above that associated with the user specified percentile threshold; (viii) raw read counts associated with the remaining transcripts are outputted into separate files that correspond to each input dataset. As to avoid overwriting the original count files, the names of updated count files are specified within the configuration file. These updated count files can be used as input for differential expression analysis software such as DESeq2. Note: in relation to step (vi) the final list of variance scores obtained are a representation of those derived from both conditions, and during step (vii) it is the variance score that is associated with the user-defined percentile that is used. The latter means that for separate comparisons (e.g. wolves vs. dogs and aggressive vs. tame foxes) at a given filtering threshold the number of filtered transcripts may not be identical (despite the same reference set being used); as the underlying variance score distribution calculated for each can be different.

Download:

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

Fig 1. Diagramatic overview of how reference based transcripts are removed by TVscript.

Steps (i) to (viii) indicate actions taken. Reference transcripts (green circles) are showen for diagramtic purposes in order to highlight how read counts (grey circles) across replicates are treated for each transcript independently. Read counts are grouped into individual files (gray rectangles) in accordance to replicate. These files are grouped into one of two conditions (blue and brown boxes). Remaining keys are indicated at the bottom.

https://doi.org/10.1371/journal.pone.0274591.g001

TVscript is implemented in Java programming language and runs on all operating systems with installed Java Runtime Environment v.8.0 or higher. It is open source and available under the GNU General Public License v3.0. Source code, usage instructions and sample data can be found on the SourceForge project page: https://sourceforge.net/projects/tvscript/. Although TVscript is implemented in Java the steps involved can be readily implemented within any language (e.g. R or python), using the detailed description provided above as well as the Java source code that is fully available. There are no dependent packages where code is unavailable. At the time of development we choose Java mainly due to its platform independence, which can be an advantage within setting up analysis pipelines involving many different tools. That said we are aware that many differential expression analysis tools are R based and future demand may warrant a supported R version.

Controlled intra-condition variation within simulated data

We tested TVscript through a series of iterative differential expression analysis experiments involving highly controlled simulated count data. During each iteration the counts associated with each transcript, within each replicate dataset created, represented simulated transcript expression from the dog reference transcriptome and were obtained using CSReadGen [49]. Using the latter, the level of random background count variation away from that required for a normalized even coverage across all transcripts could be specified, as well as a subset of transcripts to be over represented within replicates of a condition. Background count variation refers to varying count levels associated with individual transcripts that are not maintained across replicates of a condition, thus effectively reflecting intra-condition noise, whilst specifying a subset of transcripts to be over represented across replicates of a condition reflects identifiable over expressed transcripts. A similar experiment to those described here, but in relation to the effects of chimerism on the results of differential expression, has been described in Linheiro and Archer (2021) [50].

As a preliminary, and to demonstrate the general reliability of DESeq2 in the absence of random intra-condition count variation, count data was simulated for 22,580 transcripts ranging in length from 300 to 5000 bp from within the dog reference transcriptome. For a single iteration ten replicates of count dataset, each representing three million read pairs (≈20X coverage), were generated in accordance with two conditions, A and B (five in each), where within condition B one hundred transcripts were selected for count over representation by a factor of two across replicates. For all other transcripts, counts were generated to represent an even distribution of reads (length 150 bp, insert size 300 bp). These count files were used as input to perform differential expression analysis between condition A and B in DESeq2 v.1.32.0, considering transcripts with p-adj < 0.05 (corrected by the Benjamini and Hochberg method) to be differentially expressed. The number of transcripts that were detected as being over-expressed was recorded. This was repeated one hundred times in two different ways: (i) the transcripts initially flagged for count over representation were kept constant throughout and (ii) during each iteration a new set of random transcripts for over representation was selected. A brief overview of the R-script we used for differential expression analysis within individual iterations is available on the Zenodo repository [51].

Next, a similar experiment was performed but where the level of random count variation introduced into the count datasets generated ranged from 1% to 10% in steps of one. Introduced variation was not coupled between replicates, thus reflected intra-condition variation. At each level of variation one hundred iterations of the following steps were performed. (i) Ten replicates were generated and allocated into two conditions A and B (five in each), where within B one hundred selected transcripts had counts over represented. (ii) At the percent level of variation associated with the iteration, that percent of transcripts from each of the ten replicates were randomly selected for count over representation. (iii) DESeq2 was used in a similar manner to before on the count files within conditions A and B to obtain a list of over expressed transcripts. (iv) TVscript was run using a 95^th percentile variance threshold to generate ten corresponding modified count files also separated into two conditions (A’ and B’). (v) DESeq2 was again used on these to obtain a list of over-expressed transcripts. (vi) The lists of over-expressed transcripts obtained in (iv) and (v) were cross compared. Once again this was repeated in two different ways: (i) the one hundred transcripts initially flagged for over representation were kept constant throughout all levels of variation and for each of the associated iterations and (ii) during each level of variation and for each iteration a new set of one hundred transcripts were randomly selected.

Exploring the removal of transcripts associated with high intra-condition variation within real data

For each case study (wolves vs. dogs and aggressive vs. tame fox) we ran TVscript using the count datasets described under the section “RNA-seq datasets” and by applying variance filtering thresholds corresponding to variance values associated with the 70^th up to the 90^th percentiles (in steps of five), and to the 91^st up to the 99^th (in steps of one). Steps of one were used in the latter as to allow for transcripts associated with the highest levels of intra-condition variation to be explored in more detail. For each threshold value, only transcripts with variance below that value were maintained. During each run, we recorded the number and IDs of all transcripts that were removed so that they could be cross-compared. Following each run, we used the updated count files produced to perform differential expression analysis using DESeq2 v.1.22.2. Transcripts with p-adj < 0.05 (corrected by the Benjamini and Hochberg method) were considered to be differentially expressed. DESeq2 identifies differentially expressed transcripts by estimating gene-wise dispersions and applying shrinking methods to model counts and thus effectively normalize for individual outliers [31]. Distributions of gene-wise dispersions following normalization are conveniently accessible and provide a good metric to visualize the effects of removing transcripts associated differing filter levels. Differential expression analysis using the original non-filtered datasets was also performed. For the aggressive vs. tame fox case study batch effects were not considered as all data came from the same study, tissue and sequencing run, additionally no further information about sample preparation was available. For the wolves vs. dogs case study we tested for effects based on tissue, primarily for quality control of the final transcripts we drew biological-related conclusions about, and compared results obtained to those in the absence of batch information. In our analysis we used differential expression results based solely on the latter, as firstly, effects associated with tissue at an inter-study level are unpredictable as there are many factors involved, such as precision of dissection, time of dissection, time to dissect, state of individual tissue samples as well as individual who prepared sample, and other than publication or information mentioned for the fox case study, no further information on batches was obtainable. Secondly, although DESeq2 provides an internalized method for accommodating batch effects that we applied (~batch + condition), the results obtained at an intra-study level, with well defined batches, between alternative methods of testing are variable [52]. Lastly, we were primarily exploring the effects of removing hyper variable transcripts on the mechanics of detecting differentially expressed transcripts and our simulations and case studies were a means to an end in achieving this. As long as input counts for a given filtering threshold within a given case study or a iteration were consistent with those of the initial input data, the effects of removing hyper variable transcripts could be observed, independent of other factors affecting the data prior to analysis. To visualize the overall effects of covariates broadly affecting the relationships between datasets within each case study, we performed a principal component analysis (PCA) using the plotPCA function from DESeq2 with non-filtered normalized count data.

To evaluate TVscript we used three metrics that when combined quantify the overall impact of intra-condition variation on downstream differential expression analysis. The metrics were: i) number of ambiguous positives within transcripts identified as being differentially expressed in the non-filtered datasets; ii) distributions of dispersion estimates and outliers in differential expression analysis for non-filtered and all filtered datasets; and iii) discordance in the list of differentially expressed transcripts between non-filtered and filtered datasets (selected percentiles: 97^th, 95^th, 90^th).

Gene annotation and gene family analysis

For each case study, differentially expressed transcripts obtained using the non-filtered and filtered datasets were matched to the correspondent gene ID. This was done with the R package BioMart [54] using the Ensembl Gene database (version 94). To begin to identify gene families that displayed similar regulation in both dogs and tame foxes, i.e. relative to the “wild/aggressive” type, we grouped up and down regulated genes into gene families. Genes within these families were then classified according to whether they were unique to dogs or tame foxes or shared between the two. Within each case study, a gene family was only considered if all the associated genes agreed in relation to their direction of expression (up or down regulation).

Mapping success of RNA-seq data

Mapping of the 44 datasets corresponding to dogs and wolves against the dog reference transcriptome revealed an average success of 60% and 58% respectively, in terms of the number of mapped reads (S1 Fig). Similar values between wolves and dogs were expected, given their recent divergence of ~23,000 years ago [55]. Comparable proportions of reads failing to map (~40%) have been previously reported for dog brain samples [33] and are most likely associated with i) novel genes; ii) regions that are not translated despite being transcribed; iii) contamination with genomic DNA; and iv) uncharacterized chimeras and other artefacts within reference sets resulting from library preparation during sequencing [56] and various assembly errors [57]. When a different mapping approach was used for each RNA-seq dataset (i.e. kallisto) transcript abundance counts remained consistent with those obtained following Bowtie2 mapping, as indicated by high r² correlation values (S2 Fig). R² values ranged between 0.8546 and 0.9944. All per-transcript mapped read counts, obtained following each mapping approach, have been made available on the Zenodo repository [58]. For the fox datasets, an average of 50% of reads mapped to the dog reference transcriptome using Bowtie2 (S1 Fig); confirmed by the kallisto estimated abundance counts (S2 Fig). This lower percentage of mapped reads, relative to the dog and wolf datasets, could be expected due to an increased genetic divergence from dogs (~10 mya) [59] together with the other aforementioned factors. However, when these fox datasets were mapped to the Ensembl available fox reference transcriptome using both Bowtie2 an improvement in the overall mapped read counts was not observed (S3 Fig).

Controlled intra-condition variation within simulated data

When not faced with increasing levels of random intra-condition count variation DESeq2 performed exceptionally well. For 86 of the one hundred iterations performed DESeq2 recovered all transcripts that were selected for count over representation (S4a Fig). Of the other 14 iterations the lowest number recovered was 72. Similarly, for the one hundred iterations where the random transcripts selected for count over representation were re-selected during each, in 87 cases all over represented transcripts were identified as being over expressed, whilst in the remaining 13 the minimum number identified was 72 (S4b Fig). However, as levels of introduced intra-condition variation increased, the number of transcripts identified by DESeq2 fell, both before and after filtering the input counts with TVscript (Fig 2). At all levels of intra-condition variation, the post-filtered data had an increase in the number of differentially expressed transcripts identified. It should be emphasized that the reduction in the number of transcripts identified as being over expressed is not a negative reflection on the performance of DESeq2, but instead it is a consequence of purposefully increasing the level of randomness within the count data. The same pattern is true when the one hundred selected transcripts for count over representation are re-selected within each iteration (S5 Fig).

Download:

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

Fig 2. Over expressed transcripts pre- and post-filtering using simulated data.

The number of transcripts identified by DESeq2 as being over expressed both prior to (light gray) and post (dark gray) filtering of count datasets within each of the one hundred iterations performed at each level of introduced random intra-condition count variation. Each iteration involved initially simulating ten count datasets divided into conditions A and B following which DESeq2 was run to attempt to identify the one hundred transcripts selected for over representation as described in the methods. Following this the ten simulated datasets were filtered using TVscript with a 95^th percentile threshold to generate corresponding filtered datasets (divided into corresponding conditions A’ and B’) on which DESeq2 was re-run.

https://doi.org/10.1371/journal.pone.0274591.g002

For iterations associated with each increment in random intra-condition variation, the number of transcripts commonly identified as being over expressed both prior-to and post filtering are presented in S2 Table (over represented transcripts kept constant across iterations) and S3 Table (over represented transcripts re-selected for each iteration). The proportion that these numbers make up relative to the maximum number of transcripts identified as being differentially expressed, pre- and post filtering, are presented in S4 and S5 Tables (constant) (re-selected). In all cases, below a 5% level of random variation these numbers are high (constant—1 to 4% averages: 0.96, 0.98, 0.94 and 0.81; re-selected—1 to 4% averages: 0.96, 0.97, 0.93 and 0.82), indicating that on top of additional transcripts identified post-filtering with TVscript, transcripts identified pre-filtering are still found. Consequently, this suggests that additional transcripts identified as being over expressed post-filtering are not at the expense of previously detected transcripts pre-filtering. Above the 5% level of intra-condition variation the ability to successfully identify the one hundred transcripts selected for over representation within condition B diminishes within iterations (Fig 2; S5 Fig). This could be indicative of a tentative estimate on the limit of at what level of random intra-condition count variation becomes inhibitory within differential expression analysis studies.

Exploring the removal of transcripts associated with high intra-condition variation within real data

No significant difference existed between the overall distributions of the per-transcript intra-condition variation values for wolf and dog samples (Wilcoxon-test, p-value < 0.198, Fig 3a). The PCA based on the entire set of normalized non-filtered counts, revealed that the wolf samples were more aggregated than dog samples (Fig 3c). For aggressive and tame fox samples, we observed a significant difference (Wilcoxon-test, p-value < 2.2e^-16, Fig 3b) between the distributions of the per-transcript intra-condition variation values, most likely resulting from an increased intra-condition variability within tame fox samples. In particular, we found five samples that were differentiated from the remaining seven in the PCA (Fig 3d), with 80% variance being explained by this clustering in PC1.

Download:

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

Fig 3. Characterization of intra-condition variation.

Percentile range of intra-condition variation scores (x-axis) observed prior to filtering, across both case studies, a) wolves (orange) and dogs (red); b) tame (dark blue) and aggressive (light blue) foxes. PCA plots based on normalized non-filtered count data of the individual datasets comparing c) wolf and dog, and d) tame and aggressive fox. In the latter only individual samples that were positioned within a distant cluster are labelled with the sample ID.

https://doi.org/10.1371/journal.pone.0274591.g003

Prior to differential expression analysis, for each case study (wolves vs. dogs and aggressive vs. tame foxes), TVscript was used to remove transcripts in accordance with a series of intra-condition variance thresholds (Fig 4a and 4b; S6 Table). Initially, for wolves and dogs 184 transcripts (out of the 26,107) associated with high intra-condition variation (99^th percentile and above) were removed, while for the aggressive and tame fox samples, 235 transcripts were removed. The number of transcripts removed was higher for the fox samples than for those of wolf and dog, reflecting the higher intra-condition variability present. Combined across the top ten levels of intra-condition variation, 12% (n = 3134) and 14.89% (n = 3888) of the reference transcripts were removed in wolf/dog datasets and aggressive/tame fox datasets respectively (S6 Table).

Download:

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

Fig 4. Removal of transcripts above specified levels of intra-condition variation.

Percentile range of combined intra-condition variation scores (x-axis) present in each case study, a) wolves and dogs; b) tame and aggressive foxes. The number of transcripts removed in the top five percentiles (from the 95^th to the 99^th) are presented in each panel.

https://doi.org/10.1371/journal.pone.0274591.g004

Differential expression analysis

Using non-filtered datasets as input to DESeq2, 430 differentially expressed transcripts were identified between wolves and dogs (Fig 5a; S7 Table). Of those, 259 were up regulated, while 171 were down regulated in dogs. Between aggressive and tame foxes, 651 differentially expressed transcripts were identified (Fig 5a; S8 Table), of which, 532 and 119 were up and down regulated, respectively, in tame foxes. Post filtering, within the first ten steps of size one from the 99^th down to the 90^th percentiles, the number of differentially expressed transcripts identified peaks at the 97^th (n = 430; up = 255, down = 175) and the 95^th percentiles (n = 730; up = 607, down = 123) in dogs and tame foxes (Fig 5a), respectively. This indicates that for these data the removal of the 3% (n = 854) and 5% (n = 1940) of transcripts associated with the highest levels of intra-condition variation maximized the detection of differentially expressed transcripts.

Download:

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

Fig 5. Effects of removing transcripts above specified levels of intra-condition variation on differential expression analysis.

a) Number of differentially expressed transcripts (DETs) identified using non-filtered (NF) and filtered datasets, based on the top 10 percentiles (99^th to the 90^th), for both case studies. Up and down regulated transcripts are represented by red and blue dots respectively. Gray arrows identify the selected thresholds for which the results of subsequent corresponding differential expression analysis were used for the identification of candidate transcript associated with tameness within each case study. b) Number of transcripts identified as differentially expressed within the non-filtered datasets that were associated the highest levels of intra-condition variation (99^th to 90^th) within both case studies, wolves and dogs (orange dots), and tame and aggressive foxes (blue dots).

https://doi.org/10.1371/journal.pone.0274591.g005

Evaluation metrics

(ii) Distributions of dispersion estimates.

Within both case studies the regression analysis indicates that removing transcripts associated with high levels of intra-condition variation improved correlation coefficients in relation to those from the non-filtered datasets (Fig 6a and 6b; S9 Table). Associated with the elevated levels of variation observed within the fox datasets, there was a better fit within the wolves vs. dogs comparison (r² > 0.7) than that of the aggressive vs. tame fox one (r² > 0.5). For the latter, there was visible elevation in the number of dispersed points around the line of best fit. With the removal of transcripts associated with the highest levels of intra-condition variation a reduction in the number of outliers within both case studies was also observed (Fig 6a and 6b; S9 Table).

Download:

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

Fig 6. Distribution of dispersion estimates.

Plots of final dispersion estimates for both case studies, a) wolves and dogs; b) tame and aggressive foxes, calculated using DESeq2 for the non-filtered (NF; orange) and 10% filtered datasets (90^th; blue). Each black dot represents a single transcript, and red dots represent outliers. The number of outliers and correlation index (r²) are displayed in the top right corner of each panel. Both x and y-axis are transformed into a logarithm scale. The line in each graph corresponds to the regression analysis between the mean of normalized counts and dispersion estimates.

https://doi.org/10.1371/journal.pone.0274591.g006

(iii) Discordance lists of differentially expressed transcripts between applied filter levels.

Within the wolves vs. dogs case study, from the 430 differentially expressed transcripts identified when using non-filtered data as input to DESeq2, 346 were maintained when using input data filtered at the 97^th, 95^th, and 90^th percentile threshold values (Fig 7 and S7 Table). 26 transcripts were added as differentially expressed following filtering. For this case study, the overall discordance between the differentially expressed transcripts identified using filtered and non-filter input data was 25.58% (Fig 7—inset table). For the second case study, aggressive vs. tame foxes, 504 out of the 651 differentially expressed transcripts identified using the non-filtered inputs were maintained when using filtered input data at the 97^th, 95^th, and 90^th percentile threshold values, with up to 114 being added following filtering (Fig 7 and S8 Table). This time the overall level of discordance was 40.09% (Fig 7—inset table). Importantly, in both case studies the added transcripts were consistently maintained across the three filter levels. This reflects the general tendency observed within iterative testing using simulated data where the differentially expressed transcripts identified using lower filter levels are maintained at higher levels of filtering in addition to any newly identified transcripts (S4 and S5 Tables).

Download:

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

Fig 7. Differentially expressed transcripts overlapping between non-filtered and filtered datasets.

Venn diagrams representing the number of overlapping differential expressed transcripts found following differential expression analysis using non-filtered datasets and filtered datasets (97^th, 95^th and 90^th percentiles), within each case studies. The inset table provides information about the number of differential expressed transcripts lost/added following each filter step in relation to the non-filtered dataset as well as the the percentage of total discordance.

https://doi.org/10.1371/journal.pone.0274591.g007

Candidate genes and gene families

By performing annotation using the filtered datasets where the number of differentially expressed transcripts was maximized (3% and 5%, in dogs and tame foxes, respectively), we found 21 gene families in common among the up-regulated genes in dogs and tame foxes. These 21 gene families contained 50 genes (Table 1), of which 19 were exclusive to dogs while 24 were exclusive to tame foxes. The remaining seven genes (RGR, CHRNA5, SQLE, ARHGAP25, ITGA7, MYO7A and TRIB2), were found to be commonly up regulated in both dogs and tame foxes. When batch effects based on tissue were considered RGR, CHRNA5, MYO7A and TRIB2 were maintained as being commonly up regulated (S6 Fig). Note: although in relation to the latter SQLE, ARHGAP25 and ITGA7 were lost, batch effects based on tissue across multiple studies where little other batch information is obtainable could be considered as a very conservative exclusion.

Download:

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

Table 1. Shared genes and gene families (Up regulation).

List of the gene families, and shared genes, that were commonly up regulated in dogs and tame foxes. The number, and name, of the genes within each gene family are provided, with the corresponding log2fold-change values in brackets for each species. Within each family, single genes were charecterized as shared between dogs and tame foxes (bold), or as exclusively to each of the two groups. When more than one transcript for a specific gene was present, all the log2FC values are reported.

https://doi.org/10.1371/journal.pone.0274591.t001

In addition, we also found three gene families, containing four genes, simultaneously down regulated in both groups (Table 2). Two of these genes (STMND1 and OASL) were shared between dogs and tame foxes, while the other two were unique to each group. When batch effects based on tissue were taken into account, STMND1 and OASL were maintained as being commonly down regulated (S6 Fig). The same analysis performed using the non-filtered datasets revealed similar results (S10 Table), although the RGR gene family which included a shared gene between dogs and tame foxes, was lost. This gene was not differentially expressed in the non-filtered fox dataset, representing an example of genes added as differentially expressed after filtering.

Download:

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

Table 2. Shared genes and gene families (Down regulation).

List of the gene families, and shared genes, that were commonly down regulated in dogs and tame foxes. The number, and name, of the genes within each gene family are provided, with the corresponding log2fold-change values in brackets for each species. Within each family, single genes were characterized as shared between dogs and tame foxes, or as exclusively to each of the two groups. When more than one transcript for a specific gene was present, all the log2FC values are reported.

https://doi.org/10.1371/journal.pone.0274591.t002