A detailed protein-centric anatomy of variant enrichment across scales

We present a multifactorial analysis of missense variants observed in the general population (gnomAD database) [30], in comparison to somatic cancer-associated missense variants from the COSMIC database [31] and disease-associated missense variants from the ClinVar database [32]. Throughout this analysis, we further divide the gnomAD data by their minor allele frequencies (MAF) into common and rare variants, to investigate whether there are differences between these 2 subsets. We also considered all variants in a continuum of pathogenicity by ranking them using in silico variant impact predictions and evaluate features which are associated with such pathogenicity measures. A summary of the numbers of missense variants investigated is given in Table 1, and a more detailed breakdown is given in S1 Data.

Download:

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

Table 1. Numbers of missense variants which localise to different levels and regions of protein anatomy.

Data are listed for each of gnomAD, COSMIC, and ClinVar datasets. Here, “common” and “rare” variants are subset of gnomAD defined using the minor allele frequency (MAF) cutoff of 0.01, below which variants are classified as “rare.” This definition is used throughout this work except for the analysis on varying the “rarity” of variants (see main text). Note that the full-length domain-type statistics are omitted here, as by definition they will be identical to the “full-length domain” row. Fig 1B illustrates the definition of regions listed here in the first column.

https://doi.org/10.1371/journal.pbio.3001207.t001

Defining the protein anatomy

We compare pathogenic and population variants in terms of their associations with specific features across the molecular scale, in a framework we call “protein anatomy,” where we partition the human proteome in different ways. This includes the consideration of individual “proteins” (for example, investigating the enrichment of missense variants in the epidermal growth factor receptor [EGFR] protein), specific constituent “domains” of proteins (e.g., the EGFR tyrosine kinase domain), or generally for all instances of a “domain-type” (e.g., all tyrosine kinase domains) found in the human proteome. These are referred to as the “levels” of protein anatomy (Fig 1A). For each level, we analyse variant enrichment in full-length entities (i.e., protein/domain/domain-type, dependent on the level of interest) and also various constituent “regions” defined using different criteria (Fig 1B and 1C). These include protein structural information (partitioning into surface, core, and interacting interfaces), protein disorder predictions (segregating into ordered and disordered regions, within or outside of Pfam domains), and vicinity to functional sites such as phosphorylation and ubiquitination sites. We explore the interplay of “microscopic,” atomistic protein structural features, and large-scale, “macroscopic” features like functional pathways, as well as the various proteomics features collated (see Materials and methods and below), in terms of understanding the distributions of pathogenic and nonpathogenic missense variants over these features.

Download:

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

Fig 1. Variant enrichment crossing scales of protein anatomy.

(A) Levels of the protein anatomy. At the protein/domain level, the number of missense variants in a protein or domain is compared to the number of missense variants in the whole dataset which localise to defined proteins/domains. At the domain-type level, the number of missense variants in a particular Pfam defined domain-type is compared to the total number of missense variants which localise to any Pfam domain-type. These calculations are referred to as the “full-length” protein/domain/domain-type variant enrichment in this manuscript, in contrast with the calculations at regions of protein anatomy defined next. (B) Regions of the protein anatomy. We considered different levels of definition of protein regions, including (i) regions close to functional (phosphorylation/ubiquitination) sites; (ii) structural regions (core, surface [surf], and interface [interact]) of a protein; and (iii) regions predicted to be ordered or disordered which lie either within or outside of Pfam-defined domains. (C) Lists of regions considered at each level of the protein anatomy in this study. (D) The calculation of enrichment at the different levels is statistically assessed using the binomial distribution. The binomial cumulative distributive function constitutes a VES with value range 0 to 1, which quantifies enrichment. (E) Enrichment of COSMIC missense variants in protein core, surface, and interface regions, across a list of annotated oncogene (orange annotations next to the dendrogram) and TSG (blue) products. Size of points denote the level of statistical significance for the calculated VES. The genes were grouped into 2 clusters using hierarchical clustering over the VES statistics (clusters highlighted by grey rectangles; also see dendrogram); the number of oncogenes and TSGs in each cluster is noted. VES statistics can be readily browsed on the online ZoomVar web application. CDF, cumulative distributive function; EGFR, epidermal growth factor receptor; TSG, tumour suppressor gene; VES, Variant Enrichment Score.

https://doi.org/10.1371/journal.pbio.3001207.g001

Evaluating variant enrichment

To quantify missense variant enrichment, we employ a similar approach to that used in the prediction of cancer driver genes [17]: Variant enrichment has been modelled as the assignment of an observed number of variants into different levels and regions of the protein anatomy. Based on the size (in terms of the number of amino acids) of the entity of interest, we calculated the probability of observing the given number of variants localised to this entity. This quantified the likelihood of observing the given amount of variants over the null scenario where no bias towards localising to any entities were expected, while also taking the size of the region/protein/domain into account. Using the binomial distribution (Materials and methods, Eq 1), this probabilistic calculation yielded a Variant Enrichment Score (VES), ranging from 0 to 1 (Figs 1D and S1 and S2; also see Materials and methods), allowing for intuitive interpretation. For details, please read S1 Text. We also quantified the robustness of such quantification of variant enrichment: The significance of the enrichment/depletion of missense variants, in terms of their density, is assessed by comparison to simulated null distributions, in which the number of missense variants is kept identical to that observed in the data, but their positions within the protein are randomised. This goes beyond similar studies (e.g., [20,21,26]) and addresses biases which could result from the selective focus in molecular studies of disease-related proteins.

Validating VES calculation

We first applied the method of VES calculation and examined known cases of variant enrichment. By examining a curated list of oncogenes and tumour suppressor genes (TSGs) [33] in the COSMIC dataset, we found, in agreement with others [25,26,34], that these proteins could be classified into 2 groups by considering their patterns of variant distribution, one comprising proteins enriched in mutations mainly at interaction interfaces and surfaces, and another group in the core (Fig 1E). Some proteins in the latter group also show enrichment in interacting interfaces, but a clear depletion of mutations at the surface is evident. The segregation of these 2 groups in terms of cancer driver status has strong statistical support (Fisher exact test p-value = 0.0042): The first group of proteins is mainly (17 out of 24) composed of oncogenes, and the other mainly of TSGs (17 out of 25). These results are consistent with the hypotheses that activating mutations in oncogenes are likely to affect particular functions by perturbing specific interactions, while inactivating mutations in TSGs abrogate protein function [26,34]. This validates the ability of our VES formula to recapitulate known variant-enriched and depleted cases.

Robust statistical quantification confirms and extends variant enrichment patterns in large variant datasets

We reasoned that with the statistical framework we established, as well as the large variant datasets at hand, we could give robust quantification of the enrichment of variants in health and disease. While some trends of variant enrichment have been previously discussed [17–22], our method enables a unique probabilistic evaluation of these patterns in a uniform manner over ClinVar, COSMIC, and gnomAD datasets.

Population and disease-associated variants localise to different protein regions

We first analysed the trends in the enrichment of variants in different regions of the protein anatomy, defined using structural information, order and disorder, and the vicinity (distance ≤ 8 ) of the variant positions to PTMs (see above).

The following findings are highlighted:

Different structural localisation of pathogenic versus population missense variants.

We find ClinVar variants to be enriched in both protein cores and interfaces but depleted on protein surfaces (Figs 2A–2C and S3). This reflects the potential disruption, caused by such mutations, of structurally and functionally important sites [20,21,25]. The enrichment of ClinVar variants is further demonstrated by their tendency to affect residues which are highly connected when considering network representations of protein structures (see S3 Text). GnomAD variants (both common and rare) and somatic variants falling outside of cancer-related genes display the opposite trend, as variants tend to localise preferentially to protein surfaces, and are therefore less likely to impact on protein structure and function than either core or interface mutations. Somatic variants in cancer genes follow trends closer to ClinVar variants, with slight, but significant, depletion on the surface, but enrichment in the core. Protein interfaces are enriched in disease-associated variants but depleted of gnomAD rare variants. GnomAD common variants appear neither significantly enriched nor depleted; however, this may result from the relative sparsity and high dispersion of the data (fewer variants are shared between many individuals; see Table 1). Interestingly, COSMIC noncancer gene variants appear depleted in interacting interfaces. However, it becomes clear that they are actually significantly enriched when compared to simulated null distributions (see S3 Fig) and that this enrichment is due to a small subset of proteins which harbour a large number of variants at interface regions. Genes in which these variants reside may be putative driver genes (see S4 Data), as a number of known driver genes are enriched in variants in protein interface regions [17,26,28], and this phenomenon has been exploited by Porta-Pardo and colleagues [17] to identify cancer driver genes.

Download:

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

Fig 2. The localisation of variants to protein regions.

(A–C) The density of mutations in different protein regions, calculated using Eq 1. Density values (ω) were log-transformed such that negative values indicate a depletion of missense variants, while positive values indicate enrichment. Error bars depict 95% confidence intervals obtained from bootstrapping. Significance was calculated by comparison to simulated missense variant distributions (significance level indicated by: * q-value < 0.05, ** q-value < 0.001, *** q-value < 0.0001). Note here the COSMIC set is split into cancer genes (i.e., mutations mapping to proteins found in the COSMIC CGC) and noncancer gene subsets. Data are shown for protein surface (surf, panel A), interacting interface (interact, B), and core (C). (D–F) Density of mutations analogous to panels (A–C) but in regions defined by order and disorder. Data are shown here for intra-domain ordered (intra-ord, panel D), intra-domain disordered (intra-dis, E), and inter-domain disordered (inter-dis, F) regions. See S2 Data for the underlying data. CGC, Cancer Gene Census.

https://doi.org/10.1371/journal.pbio.3001207.g002

Towards a domain-centric landscape of variant enrichment

We now proceed from the protein level to examine variant enrichment across domain-types. Agglomerating missense variants at the domain-type level has the advantage of enhancing the statistical power to detect variant enrichment in terms of different protein structural features ([37] and references therein). Here, in contrast to previous studies which focus on variants clustered in sequence or structure space [22,37,38], we survey the landscape of variant enrichment across domain-types and compare the patterns of enrichment of variants from the different health and disease datasets which we have examined above. We focus our discussion on the most variant-enriched domain-types from each of the 4 variant sets. A comprehensive list comprising the union of the top 20 enriched domain-types for each dataset can be found in S4 Fig. Fig 3 shows some selected examples of this list; the missense variant enrichment at the full-length domain-type level (Fig 3A) and in each structural regions (Fig 3B) are depicted. Here, domain-types which are enriched in variants only in the COSMIC and ClinVar datasets can be seen, including known drug targets such as tyrosine kinase (Pkinase_Tyr) and ion channel (Ion_trans) domain-types. A handful of domain-types, which are only enriched in COSMIC variants, include Cadherin_tail and Laminin_G_2 (Fig 3A), both of which play an important role in cancer [39,40].

Some domain-types (e.g., Serpin, UDPGT, Collagen, and EGF_CA) contain variants from all 4 datasets. In such domains, it is likely that the precise structural localisation of a variant determines whether it plays a pathogenic role. Intriguingly, a few domain-types, such as NPIP (Nuclear pore complex interacting protein) and NUT (Nuclear Testis protein) appear only enriched in common variants (Fig 3A). This could suggest that these domains take part in functions for which it is desirable to maintain diversity within a population; however, little is known about either domain-type (Pfam accessions PF06409; PF12881). Therefore, this further highlights the bias in the number of studies targeting domains associated with disease, rather than those enriched in population variants.

Download:

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

Fig 3. A domain-centric landscape of variant enrichment.

Here, selected domain-types discussed in the main text are depicted. (A) VES at the full-length domain-type level. Bubble sizes are scaled by the adjusted p-value of the VES statistic. (B) VES calculated for each structural region (surface [surf], interacting interface [interact], and core) for the selected domain-types. (C) The number of drugs known to target proteins containing each domain-type is depicted as a bar graph. See S2 Data for the underlying data. NPIP, Nuclear pore complex interacting protein; NUT, Nuclear Testis protein; VES, Variant Enrichment Score.

https://doi.org/10.1371/journal.pbio.3001207.g003

It also becomes apparent that the global trends in variant localisation to the core, surface, and interface regions observed above are recapitulated here (Fig 3B) for those domain-types with structural coverage. The majority of domains are enriched in gnomAD (rare and common) variants at the surface but ClinVar variants at the core. For COSMIC, the patterns of localisation are more mixed, but it is clear that in comparison to the gnomAD sets, a larger proportion of domain-types are enriched in COSMIC variants at the core or interface. These include domain-types with known cancer driver associations, such as the P53 and VHL domains [41]. Case studies on CATH architectures [42] and DNA-binding domains further highlight our observed patterns of variant enrichment (see S3 Text).

We also explored how the targeting of domains by drugs and small molecules mapped to the observed landscape of variant enrichment. Using DrugBank [43] data, we observe that the targeting of domain-types by existing drugs is highly biased towards a small number of domain-types (Fig 3C), such as G protein-coupled receptors (GPCRs) and tyrosine kinase, as already extensively pointed out [44]. Indeed, we observe a large number of drugs targeting proteins containing 7tm (GPCR) domains. These domains are enriched in variants from the gnomAD and COSMIC database but are devoid of disease-associated ClinVar variants (Fig 3A). By analysing drug availability together with variant enrichment, this approach allows for more informed decisions in selecting new therapeutic targets. For example, there are domain-types which could be targeted by few or no drugs but are enriched in COSMIC and/or ClinVar variants. This could offer a starting point to prioritise drug discovery efforts for these domain-types. For domain-types already targetable by drugs, our analysis highlights domains to which multiple disease-associated variants localise, which could give scope for drug repurposing or redesign (see Discussion).

Put together, our statistical method gives probabilistic assessment of variant enrichment and yields robust quantification of these patterns across different protein and domain regions. This enables extending previously described trends of enrichment to large variant datasets and generating insights into variant enrichment in protein domains.

Functional and proteomics features distinguish variants in health and disease

We next asked whether our statistical framework could generate new insights into the relationship between missense variants and other biological features. While databases of functional annotation and measurements of transcript/protein abundance and stability are rapidly expanding, these features have been previously underexplored in the annotation of missense variants. Here, the VES framework provides metrics which could be readily explored in terms of their association with these features.

Disease-associated and population variants affect different functional pathways

We investigated whether variants from each dataset localise to proteins which are involved in distinct functional pathways. To do this, we performed Gene Set Enrichment Analysis (GSEA) [45] on lists of proteins ranked using their whole-protein VESs (Fig 4A) calculated for each dataset. The pathway enrichment scores were then subjected to clustering and Principal Component Analysis (PCA) (see Materials and methods). As shown in Fig 4B, variant enrichment segregates pathways into 3 clusters. Strikingly, each pathway cluster appears to have distinct characteristics (see Fig 4C–4E for the pathway terms belonging to each cluster). The cluster visualised in orange is primarily composed of terms associated with cancer, growth, and proliferation, whereas that coloured in pink contains pathways associated with splicing, transcription, translation, and metabolic terms. Pathways associated with sensory perception and the immune response are found in the “green” cluster. A handful of metabolic pathways also localise to this cluster; however, these appear to be more associated with environmental response and adaptation than those pathways found in the “pink” cluster; for example, pathways associated with the metabolism of drugs and xenobiotics are found here. For brevity, the “orange,” “pink,” and “green” clusters will be termed the “proliferation,” “nucleotide processing,” and “response” clusters, respectively, for the remainder of this text. A list of pathways assigned to each cluster is given in S7 Data.

Download:

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

Fig 4. Pathway clusters defined according to protein-wise variant enrichment.

(A) Schematic of GSEA. GSEA was performed over lists of proteins ranked by VES; here, enrichment over a pathway X for COSMIC variants were illustrated. The process was repeated over all KEGG pathways and the resulting NES matrix was subject to PCA and clustering analyses. (B) At the whole protein level, KEGG pathways form 3 clusters (k-means), here visualised as projected onto the first 2 principal components of the PCA. Pathway enrichment patterns are clearly distinct between COSMIC, ClinVar, and gnomAD (rare/common) data, as evidenced by the visualisation of factor loadings (arrows). See S5 Fig for projection onto 3 principal components. (C–E) Pathway terms visualised for the “proliferation” (C), “nucleotide processing” (D), and “response” (E) clusters, and sized by their cluster uniqueness score. The latter is defined as the average of the Euclidean distances to the two other cluster centres. For the top 5 unique pathway terms for each cluster, their pathway enrichment scores calculated with the 4 variant sets are also visualised in a heatmap. S7 Data contains the full list of KEGG terms mapped to these clusters. See S6 Data for the enrichment scores. GSEA, Gene Set Enrichment Analysis; KEGG, Kyoto Encyclopedia of Genes and Genomes; NES, normalised enrichment score; PCA, Principal Component Analysis; VES, Variant Enrichment Score.

https://doi.org/10.1371/journal.pbio.3001207.g004

Strikingly, this visualisation reveals that population variant datasets (gnomAD rare and common) are clearly separated from the disease-associated variants by the first principal component (PC1) (Fig 4B). Additionally, COSMIC variants are separated from ClinVar variants along the third principal component (PC3) (S5 Fig). Closer inspection of the pathway enrichment data for the top (most unique) pathways in each cluster reveals a distinction in terms of the functions that different variant datasets implicate (Fig 4C–4E). “Response” pathways appear to be enriched in variants in all 4 datasets, while “proliferation” pathways are consistently enriched in COSMIC variants alongside enrichment in ClinVar variants in a subset of these pathways. Enrichment in ClinVar variants is apparent for specific “nucleotide processing” functions, e.g., ribosomes.

Proteomics and transcriptomics features are associated with variant localisation

Proteins, of course, do not function in isolation but in the crowded environment of the cell [46]. Therefore, the properties of proteins in cells, including their quantities, turnover rates, and thermal stability, can crucially affect the fitness of a protein to perform its function. Here, we ask if variant enrichments are associated with these proteomics features. We have made use of large-scale proteomics data, including protein abundance data for various organs from PaxDb [10], proteomics surveys of protein half-lives and thermal stability [14,15], together with transcriptomics data (GTEx database [9]), to explore relationships between these features and variant localisation.

We first compared the thermal stability and abundance of proteins enriched in each class of variants. This comparison demonstrates that for proteins affected by ClinVar variants, their wild-types tend to be more stable and abundant in comparison to those proteins enriched with gnomAD variants (S6 Fig). Extending to the entire proteome, the protein-wise VES of disease-associated variants displays positive correlations with protein abundance, expression, half-life, and thermal stability, whereas population variants exhibit the opposite trend (Figs 5A and 5B and S7–S9). However, zooming into the enrichment of variants in the core of protein structures, we found that in comparison to all regions of proteins with resolved structure, proteins more enriched in ClinVar variants in the core tend to be less abundant and less stable, whereas the contrary is true for rare population variants (Fig 5). Thus, our results indicate 2 competing trends for disease-associated variants: (i) disease-associated variants tend to localise to more abundant and stable proteins, which may suggest that these proteins are more sensitive to perturbation by variants; (ii) disease-associated variants in protein cores tend to localise to less stable proteins, which is consistent with the idea that such proteins might be more easily destabilised to a degree at which function is deleteriously impacted (see Discussion). gnomAD common data also show negative correlations with protein stability, for variants occurring at the core; this could potentially support the argument presented by Mahlich and colleagues [47] that common variants could affect molecular function more than rare variants. However, we believe this is more likely to be due to the fact that very few common variants localise to protein cores, as shown in Fig 2B, resulting in sparse statistics (i.e., the correlation is calculated over VESs which are already very low). Analogous correlations for variant enrichments at protein surfaces display opposite trends to those observed at the protein cores (see S7 Fig). Due to the relative sparsity of variants which map to protein interfaces, we believe it is difficult to draw robust conclusions from any trends observed for correlations of proteomics data with variant enrichment at protein–protein interaction sites.

Download:

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

Fig 5. The protein-wise enrichment of missense variants in comparison to protein abundance, expression, and stability.

Spearman correlations for missense variant enrichment (quantified as VESs) at the full-length protein and the core with (A) protein stability in terms of melting temperature (Tm, °C) and (B) protein abundance (ppm) are depicted here. For (A), the Tm data was taken from epmid26379230, in which 2 measurements of Tm in the absence of any drug treatment were available; both measurements are considered and are denoted datasets “1” and “2” in the plot. For (B), only data from selected tissue types are listed. See S7 Fig for the complete list. S9 Data contains the underlying data. (C) Functional enrichment of proteins in KEGG pathways according to Tm. The NES for each pathway is shown on the vertical axis. KEGG pathways are listed on the horizontal axis and grouped to the 3 clusters as defined in Fig 4. See S11 Fig for a complete list of pathways depicted here. S10 Data contains the underlying data. KEGG, Kyoto Encyclopedia of Genes and Genomes; NES, Normalised Enrichment Score; Tm, melting temperature; VES, Variant Enrichment Score.

https://doi.org/10.1371/journal.pbio.3001207.g005

We highlight two more observations in terms of the interplay between proteomics features. First, the various proteomics features examined here are interdependent. Protein abundance and thermal stability are significantly correlated with one another (see S11 Data), in agreement with the work of Leuenberger and colleagues [48]. Moreover, thermal stability is correlated with the density of the protein core (see S3 Text and S10 Fig), albeit with a low correlation coefficient. The correlation values displayed in Fig 5 are also typically of a weak effect. Therefore, the interplay between variant enrichment and proteomics features appear multifaceted and complex. Secondly, in the analysis of protein abundance, the trends observed with variant enrichment at both full-length proteins and specifically the protein core are less pronounced for cell line data and break down for extracellular fluids (saliva and urine, Fig 5B). The correlation is most evident for tissues containing long-lived cell-types, such as the brain, ovary, and testis. Transcriptomics data (S8 Fig) again reinforce this picture, albeit with less contrast between datasets (particularly at the protein core). This brings finer granularity into assessing the impact of variants in different organs and contexts.

We finally ask whether correlations with these proteomic and transcriptomic features could be associated with the specific functional roles of the involved proteins. For the majority of proteomic and transcriptomic features, no clear associations with the functional clusters identified in Fig 4 can be detected (see S11–S14 Figs). An exception to this is protein thermal stability: Pathways which belong to the “proliferation” cluster are clearly enriched in proteins of lower stability than the other 2 clusters (Fig 5C). This suggests that proliferation-related proteins may be vulnerable to disruption by mutations which localise to their already unstable cores. Taken together, these analyses provide fine molecular details into defining both the resilience towards variants, and the sensitivity towards variants, for a given protein (see Discussion). Moreover, the association of variant enrichment with features such as abundance and stability is indicative of the condition (disease/health) associated with the variants.

Probing the spectrum of variant pathogenicity using protein features

The analyses above extract molecular features which distinguish between ClinVar, COSMIC, and gnomAD common and rare variants. While we successfully segregate common and disease-associated variants using pathway, structural, and proteomics features, variant impact is a continuum [49] ranging from benign to very damaging towards protein function. Instead of grouping variants by the databases from which they are collected and comparing these disparate variant subsets, we therefore ask whether the protein features discovered above could distinguish variant impact across a continuous spectrum, thereby validating the utility of these features in segregating missense variants.

Rare variants are similar to common variants

We first vary the criteria with which to define rarity of variants in the gnomAD set, to examine whether extremely rare variants would show characteristics akin to disease-associated variants. Fig 6 demonstrates that rare variants are more similar to common variants, both in terms of the functional pathways they affect and in terms of the protein regions they localise to (core, surface, and interface, order and disorder). If more stringent MAF thresholds are used to define rare variants, their properties move towards those of disease-associated variants but still remain closest to those of common variants (Figs 6 and S15). A visible separation between common and rare variants, especially in the pathway analysis, can only be seen if an extreme MAF cutoff (<0.00001) is used. This reinforces the boundary between population and disease-associated variants and supports the distinction in terms of molecular characteristics associated with rare population variants and disease-associated variants.

Download:

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

Fig 6. Rare variants are similar to common variants.

(A–D) As for Fig 4B, but only the first 2 PCs are depicted, and, in separated panels, increasingly stringent MAFs used to define rare variants. MAF cutoffs of 0.01 (panel A, data identical to Fig 4B), 0.001 (B), 0.0001 (C), and 0.00001 (D) are considered here. See S6 Data for the underlying data. (E–G) The localisation of rare variants to protein surface (E), interface (F), and core (G). Rare variants have been defined using different MAF cutoffs as shown on the x-axes. Here, the density metrics (ω) were log-transformed such that negative values indicate a depletion of missense variants, while positive values indicate enrichment. Results for the observed variants (red bars), as well as the background levels based on simulated null distributions (cyan bars), are shown. See S2 Data for the underlying data for the observed statistics. See S12 Data for statistics calculated for individual realisations of the simulated null distribution. MAF, minor allele frequency; PCs, principal components.

https://doi.org/10.1371/journal.pbio.3001207.g006

Variant impact predictors validate structural and proteomics features

The segregation of variants into disease (ClinVar, COSMIC) and healthy (gnomAD) subsets could not capture the continuous variation in their molecular impact. We therefore seek to examine the association of the protein features we have discussed above, with orthogonal measures which treat, as a continuous variable, the impact of variants pooled from all the datasets. If such association persists, this suggests that the features described above do indeed have segregating power to identify pathogenic variants from polymorphisms which carry little impact on biological function. We pooled all analysed variants together (Fig 7A) and rank them by 2 in silico variant impact predictors, REVEL [7] and CADD [16]. Based on the scores provided by these predictors, we proceeded to compare “tolerable” and “damaging” variants labelled by these predictions in terms of the features we discussed above. The majority of ClinVar variants were classified damaging (Fig 7B), whereas the majority of COSMIC and gnomAD variants were deemed tolerable. Under our chosen cutoff (score = 20), CADD classifies a greater proportion of variants as damaging (S16 Fig). We found that damaging variants are enriched in protein cores and interacting interfaces while depleted in protein surfaces (Fig 7C), similar to the ClinVar and COSMIC variants above (Fig 2). Tolerable variants show the opposite trends which resemble that of the gnomAD variants. Similarly, protein stability and abundance also segregate tolerable from damaging variants in the same way for ClinVar/COSMIC versus gnomAD variants (Figs 7D–7E and S17 and S18): Damaging variants tend to localise to more stable and more abundant proteins. However, such variants are more likely to be found in the core of proteins which are less stable and less abundant. These distinctions described here are also present if variants are binned using CADD scores (S17 and S18 Figs). These data indicate that the molecular features that we discovered via comparing ClinVar, COSMIC, and gnomAD variants are independent from such crude labelling based on the databases from which these variant data are collected. Rather, they are associated with variant impact, as assessed by orthogonal measures independent from our approach.

Download:

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

Fig 7. Orthogonal variant impact predictions validate structural and proteomics features.

(A) Schematic of pooling variants and annotating variant impact. (B) Breakdown of variant impact classified by REVEL in the variant datasets. For COSMIC variants in cancer genes, variants were segregated depending on whether they are driver mutations (curated in IntoGen [50], version 2016.5). (C) The enrichment of tolerable and damaging variants in different protein structural regions. Variants are annotated using REVEL. The bars represented the median density (ω, here taken logarithm such that negative values indicate depletion and positive values indicate enrichment) of 1,000 bootstrapped samples, each a subset of 50,000 variants. The error bars represented 95% confidence intervals from such bootstrapping. See S16 Fig for analogous results for CADD. (D–E) The correlation between VES calculated at the whole-protein level (“whole”) and the protein core, with protein stability (panel D, melting temperature or Tm) and abundance (E). Identical to Fig 5 but with the “tolerable” and “damaging” classification under REVEL score. See S17 and S18 Figs for data on CADD, and plots for all tissues represented in the abundance dataset. (F) The enrichment of surface, core, and interacting interface over variants ranked by REVEL score. The enrichment score from the GSEA procedure was plotted here. The absence of enrichment would result in a flat line at 0 (dashed black line). Curves represent data from 1 representative bootstrapped sample; the ribbons indicate 95% bootstrapped confidence intervals. (G) Pathway enrichment analysis for tolerable and damaging variants as defined by REVEL, i.e., underlying data identical to that of panels (B–D). Here, pathway enrichment scores were projected on 2 principal components analogous to Figs 4B and 6. Pathways were categorised using the scheme defined in Fig 4B. See S19 Fig for analogous results for CADD. S13 Data contains underlying data for all panels of this figure. GSEA, Gene Set Enrichment Analysis; Tm, melting temperature; VES, Variant Enrichment Score.

https://doi.org/10.1371/journal.pbio.3001207.g007

While these data serve as validation of the features we have discovered, the comparisons presented involve binning variants by their impact scores, which are based on arbitrarily determined cutoffs. To address this, we treated these impact predictions as continuous variables, ranked all variants accordingly, and repeated this analysis. By means of a GSEA approach which was also utilised above, we found that protein core and interface variants are indeed enriched in damaging variants, significantly more than the protein surface (Fig 7F). This further demonstrates the validity of structural localisation as a reliable feature to distinguish between tolerable and damaging variants. Finally, tolerable and damaging variants are also enriched indistinct functional pathways (Fig 7G). Some “response” pathways tend to be associated with tolerable variants, while “proliferation” and “nucleotide processing” pathways are more associated with damaging variants, analogous to the analysis on disparate variant sets as we see previously in Fig 4. Taken together, using orthogonal variant impact predictors, we validate that structural and proteomics features do indeed hold segregating power for tolerable and damaging variants. This suggests the use of the features discovered in our analyses above as a promising avenue to construct next-generation variant impact predictors (see Discussion).

Data sources

In this study, we have used variant data from ClinVar (dbSNP BUILD ID 149) [32] (for germline disease variants), COSMIC coding mutations (v80) [31] (for somatic cancer variants), and gnomAD exome data [30] (for population variants). Genomic positions were mapped against protein sequence data from UniProt [64] and Ensembl [65], protein structural data from the Protein Data Bank (biounit database, downloaded 28/04/2017), and protein interaction data from a large nonredundant protein–protein interaction network (UniPPIN) [66], which incorporates various interaction databases [67–71] and recent large-scale experimental studies [72,73,74]. Protein thermal stability and half-life data were obtained from separate large-scale studies [14,15]. Transcriptomic data were taken from GTEx [9], while protein abundance data (protein per million [ppm]) for each tissue/sample type were obtained from PaxDb [10]. Both PaxDb and GTEx normalised data were taken directly for use without additional filtering steps.

ZoomVar database

ZoomVar was constructed by mapping human protein sequences to resolved structures/homologues from the PDB using BLAST [75]. Protein domains were defined by scanning UniProt sequences against the PFAM seed library [76] using HMMER [77]. Per-residue mappings were performed by the alignment software T-COFFEE [78] or Stretcher [79] (which was used to map UniProt and Ensembl sequences which were not of the same length and were too long to align using T-COFFEE). These generated correspondences between PDB structures and those proteins/domains with structural coverage. Interaction complexes were inferred from homologues (defined using BLAST). As an example, if protein A and B are annotated as interacting in UniPPIN, and their structure homologues A’ and B’ are located in a resolved structural complex (and at least 1 residue from each protein is involved in a shared interface), residues from A and B are mapped onto A’ and B’ to infer their interaction interface. The partner-specific regression formula from HomPPI [80] was used to assign a score and confidence level to way. Only heterocomplexes are considered.

Definition of regions

Classification of variant pathogenicity

Variants in each dataset were annotated according to protein region localisation using the ZoomVar database. Table 1 listed the total number of missense variants in each dataset which have been mapped to each region considered in this study.

In the exploration of variant enrichment in different structural regions, the COSMIC data was divided into “cancer genes” and “noncancer genes” subsets, taking “cancer genes” as variants in the genes which comprise both tier 1 and tier 2 of the Cancer Gene Census (CGC) (COSMIC v84). The noncancer gene subset contains all other variants. In addition, to address the effect of labelling variants by the databases from which these data are collected, variants are also pooled and annotated using 2 existing in silico variant impact predictors CADD [16] and REVEL [7] predictions, using the Ensembl Variant Effect Predictor (VEP) [86]. Variants were labelled “damaging” and “tolerable” using cutoffs discussed in the publications [7,16] of these predictors (CADD: > 20 as “damaging”; REVEL: >0.5 as “damaging”; “tolerable” otherwise). The association between such variant pathogenicity measures and the molecular features extracted when comparing the disparate variant sets were subsequently investigated. Since these predictors are independent (i.e., not constructed directly with features investigated here) from our study, this mediates the impact of arbitrary grouping variants into ClinVar, COSMIC, and gnomAD common and rare subsets, and serves to validate the segregating power of our molecular features on variant pathogenicity.

The protein anatomy

In this study, missense variant enrichment was quantified across the “protein anatomy,” in which we partition the human proteome in different ways (see Fig 1A and 1B). We first define a list of “levels” of the anatomy, namely: (i) individual “proteins”; (ii) specific “domains” of proteins; and (iii) instances of a “domain-type” across the human proteome. See Results for detailed examples. Here, missense variant enrichment quantification was considered in both of the following scenarios: (i) for a given full-length instance of a “level,” relative to all other instances at the given “level” (e.g., for epidermal growth factor receptor [EGFR] relative to all other proteins in the human proteome), or (ii) for a given “region,” relative to all regions defined under a given criteria at a relevant level (e.g., for the protein core relative to all protein structural regions). For the sake of clarity, in this Materials and methods section, the instance of interest is hereafter referred to as the “entity” of interest; its use is clarified in the next paragraph and illustrated in Fig 9.

Download:

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

Fig 9. Illustration of missense variant enrichment calculation.

(A) The number of missense variants is modelled as a binomial variable. The cumulative distributive function of this binomial variable is taken as a VES for the level examined. (B) Illustration of the choice of parameter in defining the binomial variable used in calculating the VES. (C) Examples of defining the parameter p in quantifying variant enrichment for different cases. VES, Variant Enrichment Score.

https://doi.org/10.1371/journal.pbio.3001207.g009

Calculation of missense variant enrichment

To calculate the enrichment of missense variant, we first sought for a probabilistic calculation of the likelihood of observing the given number of missense variants inside an entity. We modelled this problem as the assignment of n variants into various entities and calculated the probability that k of those localise to the entity of interest. Assuming no bias towards any entity, the expected probability to assign 1 variant to an entity would be the ratio between the size (in terms of number of amino acids) of that entity to the total size of all entities.

This can be formulated as a classical binomial distribution problem. The binomial cumulative distributive function (Eq 1, also illustrated in Fig 9A) was used to assess the missense variant enrichment of a given entity, and the two-tailed binomial test was used to assess the significance of enrichment/depletion. (1) where k is the number of observed missense variants which localise to the given entity, n is the total number of missense variants which localise to all relevant entities, and p is the expected probability to assign 1 variant to the entity. p is given by the following expression: As an example, when evaluating enrichment in the core of a given protein with complete structural coverage (i.e., all residues within the protein can be assigned to any one of surface, core, or interacting interface): The definition of n, k, and p is illustrated in Fig 9B and 9C.

In Eq 1, P(X_entity) ≤ k) is the cumulative probability of observing k missense variants in the chosen entity. We took the approach used in [17] to identify protein interfaces enriched in cancer variants and generalised this to look not only at interfaces but other regions and levels defined in our protein anatomy (see above). Note that this formulation of the binomial model tests for the concentration (or avoidance) of variants to a specific entity, instead of modelling each “trial” of the binomial process being the generation of a variant. A detailed comparison and illustration of methods to calculate variant enrichment can be found in S1 Text.

Hereafter, we refer to the binomial cumulative distributive function of this binomial distribution as the Variant Enrichment Score (VES). Fig 9C contains specific examples of applying this scheme of VES calculation to different levels and regions of the protein anatomy.

For each analysis at the protein or domain level, the background proteome is defined as the summation of all UniProt proteins/domains containing missense variants in any of the datasets analysed. Proteins belonging to immunoglobulin and T cell receptor gene family products were filtered from all analyses (HGNC definition [87]), to avoid the inclusion of variants which could have arisen from the process of affinity maturation. For all calculations of enrichment and simulations involving protein or domain “regions” (e.g., core, surface, and interface), cases where the region is of size 0, or where the protein/domain contains no missense variants, were omitted in this analysis. Note that this framework of variant enrichment quantification, in contrary to others [22,37,38], is not designed to detect mutational “hotspots” clustered in sequence or structure space. Instead, it quantifies the extent to which missense variants are populated in the entity concerned, evaluating whether the number of such variants are more or less than expected. S1 Fig shows the distributions of whole-protein VES and their associated p-values for all the variant sets examined in this work.

The overall missense variant enrichment for each dataset was also calculated using a density-based metric ω (see Eq 2). (2) where S refers to the size in terms of the number of residues, and X refers to the number of variants.

This expression therefore examines the ratio between the number of variants in the entity of interest versus that of other entities, normalised by their sizes. As expected, the density shows good correlation with VES (S2 Fig); for simplicity and consistency, we use the VES metric for variant enrichment/depletion within a given entity (e.g., core of protein Z), whereas the density metric is used to describe a collection of entities at a given level/region (e.g., cores of every human protein).

Here, 95% confidence intervals were estimated via bootstrapping (10,000 iterations). The 2-tailed significance of enrichment/depletion was estimated by simulation of the null background. A total of 10,000 simulations were carried out for each dataset, in which the number of variants which localise to a given entity was kept constant, but their location within the entity randomised. The density of variants was calculated for each simulation and compared to the actual value in order to derive a p-value. Simulations were performed in this way, keeping the observed number of missense variants fixed, in order to overcome bias which stems from the assumption that variants are uniformly distributed throughout the proteome.

Gene set enrichment

Gene enrichment analyses were performed using Gene Set Enrichment Analysis (GSEA), using the implementation provided by the R fgsea package [88]. Given an enrichment statistic for each query gene, the GSEA algorithm outputs a score per gene set, which quantifies the enrichment of query genes in the sets examined. This is then normalised by the size of the gene set, to give a normalised enrichment score (NES) [45]. We utilised the centred VES enrichment statistic, i.e., subtracting 0.5 from the VES, as input into the GSEA algorithm. Thus, proteins with fewer missense variants than expected would have a negative score.

Definition of pathway clusters

The pathway normalised enrichment scores (NESs), calculated at the whole protein level for each dataset, were used to perform K-means clustering of KEGG pathways [89]. The R package NbClust [90] was used to determine the optimum number of clusters.