General architecture

The pipeline consists of seven independent modules written in Python using Jupyter notebooks Fig 1. The notebooks are chained together as a pipeline, but they can also be used as standalone applications, or integrated into other pipelines. First, the pipeline takes a list of genes and extracts exonic variants from Ensembl retaining only those variants that are predicted to affect protein sequences. Next, the program integrates the variants from ClinVar and maps them to Ensembl. Similarly, variants are extracted from the literature, here using the literature mining by Rafique et al. [15], and mapped to Ensembl. Subsequently, the harmonized collection of variants is consolidated in a database stored in the form of a text file that can easily be parsed and reused. Finally, the table of variants is mapped to all the transcripts linked by Ensembl to the genes of interest to obtain protein sequences of all the possible isoforms encoded by these genes. In this last step, the DNA sequences are translated to amino acid sequences and stored as protein FASTA files and here the user can also filter them to the isoforms of interest. To visually inspect the results, a separate module allows overlaying all the variants that can possibly affect a given gene onto the corresponding amino acid sequences.

Download:

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

Fig 1. General architecture of the pipeline.

Module 1: extract variants from Ensembl affecting genes linked to monogenic diabetes. Module 2: filter variant by consequence on the protein. Module 3: extract variants from ClinVar affecting genes linked to monogenic diabetes. Modules 4 and Module 5: extract the variants from the literature using the mining by Rafique et al. Module 6: consolidate the variants in a single table. Translation: produce the possible variant protein sequences. MODY—maturity-onset diabetes of the young, MD—monogenic diabetes, ND—neonatal diabetes, API—application programming interface, PMID—identifiers of scientific publications from the PubMed database, dbSNP identifiers—identifiers of the genomic variants from the dbSNP database, that start from “rs”.

https://doi.org/10.1371/journal.pone.0300350.g001

Module 1—Mining variants in Ensembl

In Module 1, a list of genes is mapped to Ensembl (here and further version 108) genes, transcripts, and exon identifiers using the Ensembl REST API [16]. Subsequently, the Ensembl REST API is queried using the exon identifiers to return all variants in Ensembl overlapping with the corresponding regions along with their annotation (identifiers, coordinates, consequences, etc.) For multi-allelic variants, we treat every alternative allele as a different entry, and add it as a new line to the table produced. The reference table is given in the Supplementary materials, S1 Table.

Module 2—Categorizing by consequence and pathogenicity

Module 2 takes as an input the table of exonic variants from Ensembl generated in Module 1. For each consequence type, the prevalence of variants with different levels of pathogenicity is computed and visualized as heat maps Fig 2A. The same is done for the data from ClinVar (version October 2022) with all the variants regardless of the associated phenotype Fig 2B.

Download:

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

Fig 2. Distribution of the consequence types within pathogenicity categories of the variants.

A: Variants from Ensembl, B: Variants from ClinVar. The heatmap shows the percentage of variants within pathogenicity categories for each consequence type. Arrows indicate the consequence types that were left in the Ensembl reference table created in Module 2 after filtering. “Stop lost” is marked with a blue arrow as it has conflicting evidence in two sources (see detailed in the text).

https://doi.org/10.1371/journal.pone.0300350.g002

By default, only variants whose pathogenicity is classified mostly as likely pathogenic or pathogenic are retained for further analyses: “missense_variant“, “protein_altering_variant“, “coding_sequence_variant“, “frameshift_variant“, “splice_donor_variant“, “splice_acceptor_variant“, “splice_donor_5th_base_variant“, “start_lost“, “stop_gained“, “stop_lost“, “inframe_deletion“, “inframe_insertion“.

Module 3. Mapping of the variants from ClinVar

Module 3 takes as an input three tables exported from ClinVar containing all the variants returned after querying three phenotypes: “MODY”, “monogenic diabetes”, and “neonatal diabetes”. ClinVar annotates variants with different levels of pathogenicity: “pathogenic”, “likely pathogenic”, “uncertain significance”, “likely benign” or “benign”. The variants presenting dbSNP (release 154) identifiers are passed to Module 2. The others are mapped to Ensembl using genomic coordinates and alleles and predicted consequences are obtained with the Ensembl Variant Effect Predictor (VEP) [17] called on all variants using the Ensembl REST API. The variants with predicted consequences according to Ensembl are filtered in the same way as in Module 2 and listed in S2 and S3 Tables in the Supplementary materials. The variants where no consequence type was returned were all either short deletions, short insertions, or other short fragment replacements. These variants are listed in S4 and S5 Tables in the Supplementary materials. This subset is directly passed to Module 6. Note that the full set of tables is stored on GitHub.

Module 4. Mapping variants from the literature (step 1 of 2)

Module 4 takes variants linked to monogenic diabetes according to the literature to cover the variants where the association with monogenic diabetes is not yet consolidated in ClinVar or Ensembl. As input we used the variants mined in the review by Rafique et al. [15] and provided as a supplementary table in their publication. The module uses the Vcfanno [18] library to annotate the variants with DNA coordinates and the dbSNP identifiers where possible. The Ensembl REST API is queried as in Module 1, and the results are passed to Module 6. The resulting variants in this step are given in Supplementary S6 Table.

Module 5. Mapping variants from the literature (step 2 of 2)

For the variants that could not be mapped automatically in Module 4, the title of the publication as obtained from Module 4 is queried against the Entrez Programming Utilities API (https://www.ncbi.nlm.nih.gov/books/NBK25500). to return the PubMed (https://pubmed.ncbi.nlm.nih.gov). identifiers (PMIDs) of these articles. Note that some of the PMIDs were mapped and had to be added manually. Next, these PMIDs are used to query the same API and return the “rs” identifiers of the variants mentioned in these publications. Finally, the variants are mapped to Ensembl using their identifiers and passed to Module 6. Note that not all these variants could be mapped automatically and those that did not map to Ensembl were formatted manually for input to Module 6. Particularly, the variants from ClinVar that did not map and as a result do not have the “rs” identifiers were added to the table with the identifiers from ClinVar. Besides the identifiers, the table contains the chromosome number, the position of the variant, and the reference and alternative alleles. In total, 8 variants were added to VCF manually. The resulting variants in this step are given in Supplementary S7 Table.

Module 6—Consolidation as table and VCF file

Module 6 combines all the tables produced by the previous modules and creates Venn diagrams showing the number of variants obtained from the different sources and their overlap. This table is then used to create a Variant Calling Format (VCF) file listing the site of all the variants.

We further categorized the variants into two levels of pathogenicity evidence. Level 1 variants include the variants from ClinVar obtained in Module 3 after querying three phenotypes: “MODY”, “monogenic diabetes”, or “neonatal diabetes” regardless of their reported clinical significance, complemented with the variants extracted from the literature in Module 4 and Module 5. Level 2 variants undergo stricter filtration criteria. For the variants obtained from ClinVar in Module 3, all variants labeled as benign, likely benign, or of uncertain significance were filtered out. For the variants obtained from the literature in Module 4 and Module 5, variants in BLK, KLF11, and PAX4 were removed as these genes were reported to lack pathogenicity in MODY in more recent literature [6]. The output tables of this module are given as Supplementary S8 and S9 Tables for the 1st and the 2nd level variants correspondingly.

Translation of the variant sequences into protein sequences

The translation step was performed using the ProVar tool (https://github.com/ProGenNo/ProHap). Shortly, the variants from the table obtained in Module 6 were mapped to canonical cDNA sequences from Ensembl to retrieve all the transcripts of the same gene with the annotation of start and stop codons. After mapping, all the sequences were translated to their amino acid sequences and written into a protein FASTA file. Both the table output and FASTA examples are given in Supplementary materials deposited on (https://doi.org/10.6084/m9.figshare.21444963.v2).

Sequence overlay

All the variants from the resulting database were overlaid with the reference protein sequences obtained from Ensembl for all the transcripts of all genes. The variants are represented using two separate rows corresponding to the two levels of pathogenicity confidence (see example in Fig 3). FASTA files are parsed using the Pyteomics library [19]. The protein sequences and variants are plotted using the Matplotlib library [20].

Download:

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

Fig 3.

A: All isoforms of the protein product of HNF1A with the amino acid variants from the database mapped on the sequence. B: Examples of the random fragments of the products of the canonical isoforms of HNF1A and ABCC8 showing the difference in density of variants from level 1 and level 2 databases. The top row of blue dots represents the positions of the variants from the level 2 database, whilst the bottom green row represents the positions of the variants from the level 1 database. Being the most well-studied gene, HNF1A has an almost equal number of dots in both rows as most of the reported variants are revised and confirmed as pathogenic. ABCC8 has a lot of variants in the level 1 database that are not confirmed as pathogenic and, thus, are not in the level 2 database.

https://doi.org/10.1371/journal.pone.0300350.g003

Mining variants in genes linked to monogenic diabetes

We mined variants based on a list of 109 genes linked to monogenic diabetes, aiming at being as comprehensive as possible. These 109 genes consist of: i. 14 MODY genes taken from OMIM [21] or other reviews on MODY such as [3]; ii. 10 genes linked to neonatal diabetes, lipodystrophy, and insulin signaling taken from [2]; iii. 77 genes having variants with any evidence of association with either MODY, neonatal diabetes, or just the condition referred to as “monogenic diabetes” by ClinVar.

Categorizing by consequence and pathogenicity

Since monogenic diabetes, being a Mendelian disease, is determined mostly by rare, highly penetrant coding variants [22], we focused on exonic variants when mining Ensembl. Nevertheless, some variants reported in ClinVar and the literature mapped to untranslated regions (UTR) and splice regions. In these cases we decided whether to keep these variants or not based on the pathogenicity of variants in each consequence type category. To produce protein sequences, we focused on predicted consequences: “missense variant”, “protein altering variant”, or “coding sequence variant”, and ignored variants on a transcript that were “non-coding transcript variant” or “synonymous variant”. For other types of consequences, we focused on those enriched with pathogenic and likely pathogenic variants: “missense variant”, “protein altering variant”, “coding sequence variant”, “frameshift variant”, “splice donor variant”, “splice acceptor variant”, “stop retained”, “stop gained”, “inframe insertion”, and “inframe deletion” (Fig 2). The “stop lost” consequences yielded different prevalence of pathogenic or likely pathogenic variants when considering the consequences reported by Ensembl vs. ClinVar (Fig 2), which can be explained by the differences between Ensembl and ClinVar. Here, the Ensembl dataset consists of the exonic variants, but is not linked to any pathological phenotype. The ClinVar dataset, though, was not narrowed down to any particular regions in the genes of interest, but all the variants in it are linked to clinical conditions. Therefore, the Ensembl dataset is enriched with protein coding variants, whereas the ClinVar dataset is enriched with pathogenic variants. The “stop lost” consequences were included in our analysis. Altogether, the resulting table variants contained 69,256 unique variants located in 109 genes.

Categorizing by consequence and pathogenicity

We mapped 2,701 of these variants to variants linked to monogenic diabetes according to ClinVar or the literature, termed level 1 variants thereafter: 2,220 (82%) mapped uniquely to ClinVar, 136 (5%) to publications reviewed by Rafique et al., and 345 (13%) to both (Fig 4). An effect on protein sequences was predicted for 2,624 (97%) of them, resulting in 12,643 different protein sequences when accounting for all isoforms. The reason for some variants not reaching the final translated sequences is that some transcripts in Ensembl do not have an associated canonical protein product. These are included when selecting exons in Module 1 but do not yield protein sequences. On the other hand, genetic variation can cause translation of UTRs and other normally untranslated regions [23], and genetic variation in the UTRs and splice regions might affect the translation of the proteins in an indirect way [24, 25], but the effects of these variants on amino acid sequences remain challenging to predict. We further filtered the variants to retain only the pathogenic and likely pathogenic variants, termed level 2 variants thereafter, yielding 876 variants, of which 641 (73%) mapped uniquely to ClinVar, 160 (18%) to publications reviewed by Rafique et al., and 75 (9%) to both (Fig 4). An effect on protein sequences was predicted for 714 (82%) of them, resulting in 3,776 distinct protein sequences.

Download:

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

Fig 4. Venn diagram representing the number of variants in two levels of the database and the number of variants taken from different sources.

Level 1 database (at the left) consists of i. ClinVar variants retrieved when querying three phenotypes, i.e. “MODY”, “monogenic diabetes”, and “neonatal diabetes” mapped to Ensembl; ii. All variants from Rafique et al. mapped to Ensembl. Level 2 database (at the right) consists of i. ClinVar “pathogenic” + “likely pathogenic” variants; ii. Variants from Rafique et al. excluded BLK, KLF11, and PAX4.

https://doi.org/10.1371/journal.pone.0300350.g004

In both cases, the overlap between variants from ClinVar and the literature is limited. This can be explained by the fact that the ClinVar dataset consisted of the variants linked to all kinds of monogenic diabetes (109 genes) and the Rafique et al. dataset consisted only of the MODY variants (14 genes). Thus, we re-ran the pipeline considering only MODY variants from ClinVar. The overlap of the ClinVar MODY dataset was then 18% and 10% for the level 1 and 2 variants respectively. While the overlap was improved, it is still limited. This illustrates the importance to consider different sources of information when studying monogenic diabetes, and rare diseases in general.

It should be noted here that, in their work, Rafique et al. also used ClinVar as a source of variants. The variants from ClinVar that were not found by the literature text mining algorithm are listed in a separate supplementary file in their publication. We decided not to include this list in our work as we have included all the monogenic diabetes-associated variants from ClinVar anyway. This is another reason why the overlap of the variants taken from Rafique et al. and ClinVar seems limited in our analysis.

Distribution of variants among the genes

We observed strong disparities in the number of variants linked to monogenic diabetes among the different genes. HNF1A, GCK, and HNF1B are the genes presenting the most variants in both level 1 and level 2 databases, indicating that most of the monogenic diabetes-associated variants have been reported in these genes, as well as most of the ones confirmed as pathogenic or likely pathogenic (Fig 3). Conversely, genes like ABCC8 and KCNJ11 feature many level 1 variants, while only a few of those are confirmed to be pathogenic or likely pathogenic (Fig 3). In fact, of all the genes observed, just about a third bear more than 20 variants, and only around 5% have more than 100 variants. Fig 5 represents the number of gene variants in both levels of the database, where the number of variants in the gene is more than 20. S10 Table of the Supplementary materials and figure “number of variants per gene” in GitHub represent the number of variants in each gene in both levels of the database. Besides the biological reasons that some genes have more association with monogenic diabetes than others, this effect can be due to study biases or heterogeneous levels of information on these variants. For example, the most common genes known to cause MODY (HNF1A, GCK, HNF1B, and HNF4A) [2] might have received more attention than others (e.g. ABCC8 and KCNJ11), yielding a lower confirmation rate for the pathogenicity of the variants. It is also worth noticing that for the time of this manuscript, only variants in HNF1A, GCK, and HNF4A are reviewed by the expert panel on ClinVar (https://clinicalgenome.org/affiliation/40016). At the same time, ABCC8 is reported to be also associated with other types of diabetes [26] and definitely deserves attention from the perspective of MODY associations as this may bring the researchers closer to the overall understanding of diabetes causal mechanisms and more precise treatment outcome predictions. The variant distribution in the studied genes can be observed in the figures visualizing the variants in the protein sequences (Fig 3 and others in the “figure” directory at (https://github.com/kuznetsovaks/MD_variants).

Download:

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

Fig 5. Number of variants in the genes with the largest number of variants.

Selected are the genes in which the number of variants in the level 1 database is more than 20. The full bars represent the number of variants in the level 1 database, and the brown part of the bars represents the number of variants in the level 2 database.

https://doi.org/10.1371/journal.pone.0300350.g005

Pipeline reproducibility

In order to check whether our pipeline is applicable to other rare diseases and show the reproducibility of the approach, we have run it on Hajdu-Cheney syndrome (HCS). Similar to monogenic diabetes, this is a rare monogenic condition that can be caused by a number of variants in protein-coding regions. All the intermediate files and the results of this analysis can be found in the “HCS” directory at at the GitHub. Finally, we collected all the variants reported to be linked to HCS and highlighted the ones that are classified as pathogenic by ClinVar. All the variant sequences have been translated and added to the human proteome fasta database and the positions of the variants in the protein sequence have been visualized.