key: cord-0931444-szmkv9hk
authors: Kaur, Sandeep; Sikta, Neblina; Schafferhans, Andrea; Bordin, Nicola; Cowley, Mark J.; Thomas, David M.; Ballinger, Mandy L.; O’Donoghue, Seán I.
title: Streamlined use of protein structures in variant analysis
date: 2021-09-11
journal: bioRxiv
DOI: 10.1101/2021.09.10.459756
sha: e96d33c72911289d2feba867a1e7732c1758cdeb
doc_id: 931444
cord_uid: szmkv9hk

Motivation Variant analysis is a core task in bioinformatics that requires integrating data from many sources. This process can be helped by using 3D structures of proteins, which can provide a spatial context that can provide insight into how variants affect function. Many available tools can help with mapping variants onto structures; but each has specific restrictions, with the result that many researchers fail to benefit from valuable insights that could be gained from structural data. Results To address this, we have created a streamlined system for incorporating 3D structures into variant analysis. Variants can be easily specified via URLs that are easily readable and writable, and use the notation recommended by the Human Genome Variation Society (HGVS). For example, ‘https://aquaria.app/SARS-CoV-2/S/?N501Y’ specifies the N501Y variant of SARS-CoV-2 S protein. In addition to mapping variants onto structures, our system provides summary information from multiple external resources, including COSMIC, CATH-FunVar, and PredictProtein. Furthermore, our system identifies and summarizes structures containing the variant, as well as the variant-position. Our system supports essentially any mutation for any well-studied protein, and uses all available structural data — including models inferred via very remote homology — integrated into a system that is fast and simple to use. By giving researchers easy, streamlined access to a wealth of structural information during variant analysis, our system will help in revealing novel insights into the molecular mechanisms underlying protein function in health and disease. Availability Our resource is freely available at the project home page (https://aquaria.app). After peer review, the code will be openly available via a GPL version 2 license at https://github.com/ODonoghueLab/Aquaria. PSSH2, the database of sequence-to-structure alignments, is also freely available for download at https://zenodo.org/record/4279164. Contact sean@odonoghuelab.org Supplementary information None.

For many clinicians and researchers, one of their core tasks involves understanding the impact of a variant on a protein's function. Gaining this understanding can be a complex process, requiring integration of data from clinical reports, in vitro experiments, and in silico predictions. In many cases, this integration can be helped by using 3D structures of proteins, which can provide detailed insight into functional impacts of variants -especially how interactions with other molecules may be affected (Gress 2020) . One of the main methods for using these structures is simply as a location to map variant data; this then allows a researcher to consider spatial context when assessing functional impact. Many tools are available to help researchers derive insights from 3D structure (Glusman et al. 2017) . Some provide a streamlined workflow, but are mostly restricted to human proteins (e.g., MuPIT Interactive from Niknafs et al. 2013 ; VarSite from Laskowski et al. 2020) ; others are restricted to fixed sets of variants (e.g., MISCAST from Iqbal et al. 2020 ); many tools (e.g., dSysMap from Mosca et al. 2015 ; Mechismo from Betts et al. 2015) are mostly restricted to using only experimentally derived 3D structures (i.e., the PDB; Berman et al. 2000) . Other tools provide rich, generic capabilities with few restrictions, but can be complex and time-consuming to use (e.g., ChimeraX from Pettersen et al. 2020) . Unfortunately, as a result of these limitations, structural data are often not included in variant analysis, and often not seen as essential in variant analysis (e.g., Pabinger et al. 2014) . This is especially true in clinical genomics, where researchers are often under time pressure to analyze large numbers of variants. Thus, unfortunately, many researchers fail to benefit from the potentially valuable insights that could be gained from structural data. Addressing this issue was our core goal in this work, which arose from a collaboration between bioinformaticians and clinical genomics researchers. As a first step, we realized the need to devise a very simple entry point for including structural data in variant analysis; this entry point needs to be easily added to existing genomics workflows -much of which involves spreadsheets, as a major format for management and communication of mutational data. In addition, we wanted to overcome limitations with existing tools, as mentioned above; thus, we aimed to create a system that supports essentially any mutation for any protein, and lets researchers use all available structural data -including models inferred via very remote homology. Finally, our system should be easy to learn, and be fast and simple to use. To address these goals, we built upon Aquaria, a web-based, molecular graphics resource (O'Donoghue et al. 2015 ) that is freely available and open-source. Aquaria provides structural information for >500,000 protein sequences in SwissProt (The UniProt Consortium 2019), which includes proteins from many model organisms used in research. For each protein sequence, Aquaria uses HHblits (Steinegger et al. 2019 ) to pre-calculate alignments onto all available 3D structures. This generates >120 million sequence-to-structure alignments (Schafferhans and O'Donoghue 2020) , which gives an average of ~200 structural models for each protein sequence. In this work, we describe extensions to Aquaria that largely address the above limitations with existing structure-based variant analysis tools. The extensions include summarizing of structures containing the variant, incorporating new variant data from COSMIC (Tate et al. 2019) and FunVar (Sillitoe et al. 2021) , as well as improving how data are integrated and presented to the user -including data from UniProt (The UniProt Consortium 2019), SNAP2 (Hecht, Bromberg, and Rost 2015) , and PredictProtein (Yachdav et al. 2014) . We demonstrate these capabilities by analysing variants in two case studies: the human oncoprotein NRAS and the S protein from SARS-CoV-2.

For specifying variants, our system uses the HGVS notation (den Dunnen et al. 2016) , which is widely used in clinical genomics. Our system supports essentially all variants that can be expressed using the notation (Table 1) ; the only exceptions are cases where position is not specified -for example 'p.0', which indicates that no protein is produced, or '[?]', which indicates an allele with unspecified consequence -or cases where the mutation is silent (e.g. 'p.Cys370='). HGVS notation was developed for annotating genomics data; to increase convenience with protein sequences, we added two optional rules as minor extensions: (1) the 'p.' prefix optional, and (2) amino acids can be abbreviated using either three letter or single letter coding. Thus, in our system, 'C370T' is equivalent to 'p.Cys370Thr'. Currently, our system does not support annotation of DNA or RNA 3D structures; however, we are plan to support this later, in which case, mutations on such sequences will require the prefixes "c.", "g.", "m.", "n.", "o.", or "r." to specify numbering for coding DNA, genomes, mitochondrial DNA, non-coding DNA, circular genomes, or for RNA, respectively. is built mainly using JavaScript; the server uses Node.js, while the client uses primarily Vue.js and Jolecule (https://jolecule.com). Underlying Aquaria is PSSH2, a Sequel database of sequence-to-structure alignments (Schafferhans and O'Donoghue 2020). We have recently completed a major refactoring of the Aquaria user interface, which we plan to describe in a forthcoming publication. Below, we describe the new variant analysis capabilities. When a variant is specified by the user in the URL, Aquaria first identifies the consequence according to the patterns described in Table 1 . In parallel, Aquaria fetches feature sets from several external sources (below and Figure 1) ; once data from a source is received, it is extracted and converted to Aquaria's feature format (https://bit.ly/aquaria-features), and the queried variants are then checked to see if any of the features pertain to it. The features containing the variant are added to a new 'Added features' set, which is updated every time data from an external source is received. Upon initial page load, the 'Added features' are displayed on the 3D structure; however, the user can manually switch to any other feature set. It is worth noting that the mutation is specified via the URL's query string (i.e., following the '?' character), which gets sent to the Aquaria server; however, all feature classification, requests to external servers, and processing are performed on the user's browser (client-side). On the server side, the query is used to identify the best matching structure for initial display in Jolecule (the 3D viewer). Additionally, structure counts pertaining to the specified variant are determined (described in the next section).

On the server, the query is used to identify the best matching structure most relevant to the specified variants, from the list of all matching structures for a sequence (which were identified via homology modelling (Schafferhans and O'Donoghue 2020)) for initial display in Jolecule, as follows:

1. If a user specifies a PDB identifier in the URL, then the specified PDB is chosen for display. 2. Otherwise, when multiple variants are specified, a subset of structures is identified which covers the most number of variants, starting from the left-most user-specified (i.e. in the order specified by the user in the URL). 3. If any specified variants are 'missense' variants, then the subset is filtered further to identify structures containing the new amino-acid resulting from the missense variant. 4. If multiple structures result from the previous two filtering steps, then the structure with the highest rank, based on alignment identity score, coverage score, resolution and experimental method, is chosen from the filtered list. 5. If no structure results from (2) and (3), then the default best matching structure is added (i.e. the highest ranked matching structure for the sequence). For each specified missense variant the number of structures containing the new amino acid resulting from the mutation are identified. Additionally, the number of structures for a variant's sequence position are also identified and returned to the client. These are then added to the 'Added features' feature set, and are available for visualization and navigation via the popup.

The UniProt Knowledgebase (The UniProt Consortium 2019), consists of, perhaps the largest collection of protein sequences and their published curated annotations that describe regions or sites of interest in the protein sequence such as secondary structure regions or PTM sites. Aquaria fetches and displays the available curated feature sets via UniProt's API (Figure 1) . If the information for the user-specified variant in the URL is available in the following UniProt feature sets, namely 'Modified residue' (modified residues excluding lipids, glycans and protein cross-links), 'Mutagenesis site' (site which has been experimentally altered by mutagenesis), 'Sequence conflict' (description of sequence discrepancies of unknown origin), 'Sequence variant' (description of a natural variant of the protein), 'Cross-link' (residues participating in covalent linkage(s) between proteins), 'Site' (interesting single amino acid site on the sequence) or 'Metal ion-binding site' (binding site for a metal ion); these data are then collected and added to the 'Added features' feature set.

The Catalogue of Somatic Mutations in Cancer (COSMIC) (Tate et al. 2019 ) is a curated resource of somatic human cancers. The data are collected from literature curation of genes and genome-wide screens. To access these data, Aquaria fetches a list of all variants for a given protein available in COSMIC v68 (1,024,494 total variants) from MyVariant (Figure 1) . In the returned data, the variants are described at the genome level only; thus, Aquaria then uses data from SnpEff (Cingolani et al. 2012) to extract the corresponding HGVS protein description and variation effect, which determine the sequence highlighting for these variants, e.g. for 'start lost' mutations, the entire sequence is highlighted indicating that the entire protein is affected. If a COSMIC variant exists at the same position as the user-specified variant in the URL, then these data are collected and added to the 'Added features' feature set.

SNAP2 (Hecht, Bromberg, and Rost 2015) is an in silico prediction method that uses neural networks to predict effects of single amino acid substitutions on protein functions. For each residue in a protein sequence, SNAP2 provides two types of scores. The first, mutational sensitivity, consists of a list of 20 scores that indicate the predicted functional consequences of the position being occupied by each of the 20 standard amino acids. Large, positive scores (up to 100) indicate substitutions likely to have deleterious changes, while negative scores (down to -100) indicate no likely functional change. A single summary score is then calculated based on the total fraction of substitutions predicted to have deleterious effect, taken to be those with a score > 40. The second, mutational score, is based on the same 20 scores above, but calculates a single summary score for each residue as the average of the individual scores for each of the 20 standard amino acids. Aquaria fetches and displays these scores via the SNAP2 API ( Figure 1) . For a user-specified variant, Aquaria adds the variant-specific SNAP2 prediction as well as the average mutational sensitivity for that position to the 'Added features' feature set.

The CATH resource (Sillitoe et al. 2021) identifies and classifies structural and functional domains in proteins using a hierarchical classification that consists of class (secondary structure content), architecture (orientation of secondary structure), topology (shape and connectivity of secondary structures), homologous superfamily and functional family (functionally coherent groups). Aquaria fetches CATH annotations for a protein using multiple API-endpoints (https://github.com/UCLOrengoGroup/cath-api-docs) and makes them available as two feature sets namely, superfamilies and functional families, both of which contain the class, architecture and topology classification, along with gene ontology, enzyme commission terms and species distributions of the homologous superfamily. The functional family feature set additionally contains the functional family domain. Since the CATH annotations for SARS-COV-2 are not yet available via the API, we host and utilize a pre-release version of these annotations (Figure 1) . For a user-specified variant, if the variant position forms part of any domain, then they are added to the 'Added features' feature set.

FunVar is a new resource derived from CATH and other inputs (Sillitoe et al. 2021 ) that presents the predicted impact of variants (FunVar), on protein structures and functions, derived by identifying whether the variants lie within or in close proximity to the interface regions or the functional sites, which are highly conserved residues in CATH functional family multiple sequence alignments. FunVar is currently available for proteins implicated in human cancers (data from (The Cancer Genome Atlas Research Network et al. 2013) ) and, SARS-COV-2 viral and its human host interactor proteins (data from (Gordon et al. 2020) ). Aquaria fetches and displays the FunVars using the FunVar API (Figure 1) . If any FunVar variant exists at the same position as the user-specified variant in the URL, then these data are collected and added to the 'Added features' feature set.

The PredictProtein resource (Yachdav et al. 2014 ) provides comprehensive computational prediction regarding multiple aspects of protein structure and function. PredictProtein features include secondary structure, solvent accessibility, conservation and, protein and DNA binding. Aquaria fetches and displays these features using the PredictProtein API (Figure 1) . For a variant specified by the user in the URL, PredictProtein's conservation prediction and score are added to the 'Added features' feature set.

Our system allows variants to be specified via the following URL schema: https://aquaria.app/Human/NRAS?Gly12Asp

Our system supports most variants that can be described using HGVS notation (den Dunnen et al. 2016) , with minor exceptions as noted in Table 1 . The URL schema can also be used to specify a description for each variant, as shown below. The description can include HTML tags, thus allowing inclusion of images or interactive graphics.

https://aquaria.app/Human/NRAS?Gly12Asp."C5: Pathogenic" Description Finally, the schema can also be used to specify multiple variants for one protein:

https://aquaria.app/Human/NRAS?Gly12Asp."C5: Pathogenic"&Gln61Lys

The systematic pattern of the URL enables it to be easily used within spreadsheets -a common tool in clinical genomics -containing mutational data from patients. The spreadsheet formula below illustrates how this can be done by referencing cells that contain the gene name and the variant (in this example, cells A2 and B2, respectively):

=hyperlink("aquaria.app/Human/"&A2&"?"&B2,"Open in Aquaria")

Clicking on a link that follows the above URL schema opens a web page on the Aquaria resource that gives access to all 3D structures related to the specified protein, and shows the specified protein mapped both, onto a representation of the protein sequence, as well as the 'best' matching structure. See next section for details. which is critical for NRAS to function normally (Kapoor and Travesset 2015) . This structure-based insight is confirmed to be correct via information shown in the popup, thus providing a validation that the information presented can be effective in assessing variants. Aquaria also shows that the Gln61Arg mutation occurs in a region that is missing in the PDB entry shown by default; however, by using the Aquaria interface, a researcher can find other related structures where this region is present, demonstrating the value of making all related structures easily accessible. An example is provided in the URL below:

https://aquaria.app/Human/NRAS/4q21?Gly12Asp&Gln61Arg

Case study 2: SARS-CoV-2 S protein Here, we showcase how Aquaria's variant analysis capabilities can help in accessing three SARS-CoV-2 mutations of particular concern in the current COVID-19 pandemic (Leung et al. 2021; Tang et al. 2021; Greaney et al. 2021) ; the mutations can be accessed via the URL:

https://aquaria.app/SARS-CoV-2/S?K417N&E484K&N501Y

Only one of these mutations (E484K) occurs very close to a bound protein called 'Tyl', a nanobody known to block binding to ACE2, present in the structure shown (Figure 3A & B) . Thus, we infer that E484K -but not K417N or N501Y -may block binding of this nanobody. 

The URL schema for variant specification developed in this work is both human-readable and human-writeable; in addition to its utility in spreadsheets -enabled via standardized Aquaria URL schema and HGVS nomenclature -this improved readability compared with other related resources may help more researchers access and use structural information more routinely. To our knowledge, no other system supports essentially any mutation for any well-studied protein, uses all available structural data -including models inferred via very remote homology, summarizes structures available for a variant -integrated into a system that is fast and simple to use. By giving researchers easy, streamlined access to a wealth of structural information during variant analysis, our system will help in revealing novel insights into the molecular mechanisms underlying protein function in health and disease.

The Protein Data Bank

Mechismo: Predicting the Mechanistic Impact of Mutations and Modifications on Molecular Interactions

A Program for Annotating and Predicting the Effects of Single Nucleotide Polymorphisms, SnpEff: SNPs in the Genome of

HGVS Recommendations for the Description of Sequence Variants: 2016 Update

Mapping Genetic Variations to Three-Dimensional Protein Structures to Enhance Variant Interpretation: A Proposed Framework

A SARS-CoV-2 Protein Interaction Map Reveals Targets for Drug Repurposing

Comprehensive Mapping of Mutations to the SARS-CoV-2 Receptor-Binding Domain That Affect Recognition by Polyclonal Human Serum Antibodies

Integration of Protein Three-Dimensional Structure into the Workflow of Interpretation of Genetic Variants

An Alpaca Nanobody Neutralizes SARS-CoV-2 by Blocking Receptor Interaction

Better Prediction of Functional Effects for Sequence Variants

MISCAST: MIssense Variant to Protein StruCture Analysis Web SuiTe

Differential Dynamics of RAS Isoforms in GDPand GTP-Bound States: Differential Dynamics of RAS Isoforms

VarSite: Disease Variants and Protein Structure

Early Transmissibility Assessment of the N501Y Mutant Strains of SARS-CoV-2 in the United Kingdom

DSysMap: Exploring the Edgetic Role of Disease Mutations

Crystal Structure of the Human NRAS GTPase Bound with GDP

MuPIT Interactive: Webserver for Mapping Variant Positions to Annotated, Interactive 3D Structures

Aquaria: Simplifying Discovery and Insight from Protein Structures

A Survey of Tools for Variant Analysis of Next-Generation Genome Sequencing Data

CSF ChimeraX: Structure Visualization for Researchers

PSSH2 -Database of Protein Sequence-to-Structure Homologies

CATH: Increased Structural Coverage of Functional Space

HH-Suite3 for Fast Remote Homology Detection and Deep Protein Annotation

Introduction of the South African SARS-CoV-2 Variant 501Y.V2 into the UK

COSMIC: The Catalogue Of Somatic Mutations In Cancer

The Cancer Genome Atlas Pan-Cancer Analysis Project

UniProt: A Worldwide Hub of Protein Knowledge

PredictProtein-an Open Resource for Online Prediction of Protein Structural and Functional Features

We gratefully acknowledge the help of Paul Ashford and other members of the CATH team in advising us on accessing and interpreting FunFam and FunVar information. Also, we thank Bosco Ho for developing and maintaining Jolecule.

SOID and NS gratefully acknowledge support from the Garvan Research Foundation, Sony Foundation Australia, and Tour de Cure Australia. NB acknowledges funding from the BBSRC (BB/R009597/1).

 Tables   Table 1. Notation for variants currently supported by Aquaria's URL schema