Reporting phenotypes in mouse models when considering body size as a potential confounder
© Oellrich et al. 2016
Received: 21 October 2015
Accepted: 2 February 2016
Published: 9 February 2016
Genotype-phenotype studies aim to identify causative relationships between genes and phenotypes. The International Mouse Phenotyping Consortium is a high throughput phenotyping program whose goal is to collect phenotype data for a knockout mouse strain of every protein coding gene. The scale of the project requires an automatic analysis pipeline to detect abnormal phenotypes, and disseminate the resulting gene-phenotype annotation data into public resources. A body weight phenotype is a common result of knockout studies. As body weight correlates with many other biological traits, this challenges the interpretation of related gene-phenotype associations. Co-correlation can lead to gene-phenotype associations that are potentially misleading. Here we use statistical modelling to account for body weight as a potential confounder to assess the impact. We find that there is a considerable impact on previously established gene-phenotype associations due to an increase in sensitivity as well as the confounding effect. We investigated the existing ontologies to represent this phenotypic information and we explored ways to ontologically represent the results of the influence of confounders on gene-phenotype associations. With the scale of data being disseminated within the high throughput programs and the range of downstream studies that utilise these data, it is critical to consider how we improve the quality of the disseminated data and provide a robust ontological representation.
During the statistical comparison of control and gene-altered data, confounding variables associated with both the genotype change and the phenotype of interest can lead to an association that is true but potentially biologically misleading. The presence of the confounding relationship can lead to errors in the estimates of the relationship between the treatment of interest (here the genotype change) and the variable of interest (here the phenotype). Good experimental design can manage many potential confounders using standardisation e.g. with the potential confounder of age, the study would only test animals of the same age. An alternative strategy is randomisation, in which animals of multiple ages are tested in both control and the experimental knockout group. Yet another strategy is grouping (blocking) according to a confounding variable (e.g. pup or adult). Depending on the strategy applied, the final annotation could be specific to one particular age. To minimise the potential impact of confounders within IMPC, the community identified critical sources of variation in screens and used this to develop a standardised operating procedure which, where possible, minimises variation and captures potential sources of variation as metadata with each dataset. Metadata parameters (e.g. X-ray equipment) are included in the IMPReSS protocols and submitted metadata is used to determine comparison groups as part of the statistical analysis pipeline.
In many research studies, it is not possible to manage confounding variables during the design. For example, in many gene knockout studies, the knockout animals show an abnormal body weight change. Therefore, any other phenotypic traits (e.g. abnormal body fat mass MP:0012320) that correlate with body weight will also be impacted. As the experimenter cannot control this potential confounder through the design, it is necessary to consider statistical methods for non-equivalent groups . These include regression methods where the confounder is treated as a covariate, meaning the statistical test will assess the effect of the genotype on the phenotype after adjusting for the confounder’s relationship. This requires a dataset to be processed twice, first without and then with the confounder in the statistical analysis; giving two sets of results for the test of genotype. This granularity has a high potential value to improve our interpretation of the relationship between a gene and associated phenotypes. However, the vast majority of MP terms represent absolute phenotype changes in a variable of interest. The Mouse Genome Informatics database (MGI)  developed MP to manually curate the scientific literature. However, only in rare, clear cause and effect cases are confounding variables represented as part of the ontology. For example, the term “progressive muscle weakness” (MP:0000748) is defined as a muscle weakness that increases with time. Time or age are clearly contributing to the severity of the phenotype and thus represent knowledge that should be represented in the ontology . However, in many studies a confounding variable is noted by authors’ to contribute to a phenotype, but a clear cause and effect relationship is not established. The current mechanism employed by MGI is to manage confounders at the level of annotation by utilising free text qualifiers. For example, the curator will note if an author states body weight was a confounder when associating a phenotype to a genotype. With the scale of IMPC data and the automated aspect of statistical analysis and subsequent annotation, we have the potential to manage these issues in a consistent way and through standardisation better support downstream informatic analysis. The interest in including body weight as a covariate, in both high throughput phenotyping studies and small scale studies, is growing [8, 11–13]. This manuscript aims to raise awareness of the issues and demonstrate the potential value of addressing the problems. We then identify adaptations to the existing mechanisms utilised by the community that could address this new aspect where we wish to disseminate the outcome of an analysis that considers body weight as a confounder.
Data and scripts used to investigate and demonstrate issues presented within this manuscript are available at Zenodo .
Body weight as a confounder
Body weight is a highly heritable trait and is estimated to be a potential latent variable in a third of experiments studying knockout mice . It has been shown that body weight correlates with many variables, ranging from body composition to clinical chemistry . Including body weight in the computational analysis allows the phenotype to be assessed after adjusting for weight differences (see Additional file 1: Supplementary Methods).
Possible outcomes of a dual analysis process
A2: + weight
Absolute phenotype ∆
No longer significant–confounded by BW
Relative phenotype ∆
Adding weight increases sensitivity to detect ∆
Absolute phenotype ∆ and relative phenotype ∆
1. BW difference: still there is a significant ∆ as ∆ larger/smaller than expect for BW difference
2. BW same: a significant ∆. Weight explains variation but does not lead to phenotype difference.
Even in cases where it is clear that body weight is truly acting as a confounding variable and is not just explaining data variance (Table 1, row 1), causality is not determined. For example, we cannot assess whether the lean mass is lower in the Dlg4 line because the body weight is fundamentally lower or because there is less lean mass leading to a lower body weight. The refinement is therefore to consider the data and assess for both relative and absolute changes and disseminate this richness.
Magnitude of impact and complexity
Challenges applying existing solutions
As demonstrated with the provided analysis, taking confounding variables such as body weight into account may lead to a more comprehensive dataset and should be further investigated (see Table 1). The dissemination of the resulting annotation data is achieved through a collaboration between different communities. IMPC currently uses MP to annotate genes with phenotypes. MP is a pre-composed phenotype ontology in which every concept semantically describes one particular phenotype, e.g. decreased lean body mass (MP:0003961). While this paper generalises to gene-phenotype annotations, MGI distinguishes further the additional data such as the genetic background or the sex if there is a difference between male and female mice. While the majority of the annotations contained in MGI do not take confounders into consideration, sex in the presence of sexual dimorphism could be regarded as such and is captured at times in MGI. For example, the gene Dmxl2  exhibits sexual dimorphism such that the phenotype was only found to be significant in the females of heterozygous mice and this is recorded as a curator note.
Body weight is not the only variable that could be used to adjust for the size of the animal; alternatives include body length or width. Adjustment for body size as a confounder has unique challenges (see section ‘Body weight as a confounder’) and particular issues with determining causality. Thus, we investigated solutions for the standardised reporting of phenotypes after considering body weight as a confounder as a relative phenotype change within existing semantic frameworks and report our findings here. Potential solutions were limited to those we believed could be implemented as they had the lowest modification requirements on the existing dissemination pipelines, such as those maintained by MGI. We note that the discussed solutions only focus on future dissemination but do not include strategies on how to deal with legacy data.
Use of pre-composed ontologies
As mentioned before, the vast majority of phenotypes represented in the current version of MP constitute absolute changes that cannot readily be applied to confounder-adjusted phenotypes. In order to represent the results of a confounder-sensitive analysis, additional MP concepts would be needed that would allow a user to report relative phenotype changes (see column 2, Table 1, rows 3 and 4). For example, to represent the changes in the absolute and relative changes in mouse line Dlg4, we would need the additional concept “relative increase in lean body mass after body weight adjustment”. However, pre-composing concepts for relative phenotype changes would mean that for each phenotype that is influenced by one or multiple confounders (e.g. body size or length), multiple concepts for each unique phenotype-confounder relationship would need to be added (abnormal/increased/decreased). This would lead to a vast increase in the number of terms (i.e. term explosion) that need to be added and maintained within MP, which would be untenable. This may also be confusing for the user community of curators and annotators as the number and complexity of terms exposed for search and/or annotation grows.
Tagging pre-composed terms
An adaption to the pre-composed term is to associate an attribute to the annotation by addition of free text tags. This is equivalent to the current implementation used in literature curation at MGI. For example, a gene could possess an annotation “increased lean body mass”, with an annotation or ‘tag’ on this annotation detailing if any/which confounder has been used for adjustment, e.g. “after adjusting for body weight”. However, as the tags are not standardised this may result in non-comparable annotations of genes and an increase in curatorial workload. Furthermore, informatics tools are not capable of interpreting tags of gene annotations and may lead to erroneous presumptions (in the case of a relative change after confounder adjustment that would not be reported with absolute changes only; row 3, Table 1).
In order to disseminate relative phenotype changes to the broader community using tagged pre-composed phenotype ontology annotations, existing gene-annotation databases need to be able to store this additional data and expose this for query. This may require not only changes to the database itself, but also to web interfaces as well as services for data download, in addition to strategies for handling legacy data.
Standardised qualifiers of pre-composed term
A refinement to the preceding method, is to add standardised qualifiers to the genotype-phenotype annotation. One ontology that can be used to represent these standardised qualifiers is the Phenotype And Trait Ontology (PATO) [18, 19]. The difference between this solution and the previous is that the free text tag is replaced with an ontology term. This suggestion is similar to how sexually dimorphic associations are currently treated. For example, Kcne2 knockout mice have a number of abnormalities that are specific to the male mice and this is captured as a MP term with associated sex classification tag . The advantage of this solution is that the variability that may occur with free-text tags is reduced to a defined set of ontology concepts. However, following this solution would need an agreed set of ontologies used for the annotation of relative changes and possibly extension to these to account for all possible confounders.
Similar to the latter approach, third parties such as MGI can then choose to add these additional annotations to their data storage to hold the information for relative phenotype changes. This may mean that database schemes as well as provision and distribution methods need to be adapted to handle the additional data and be able to distinguish between absolute and relative phenotype changes. If these changes were to be integrated in existing databases, ways of handling legacy data need to be taken into consideration.
Entity 1: lean body mass
Quality: relative to
Entity 2: body weight
Applying a post-composed representation to confounder-adjusted phenotypes may lead to multiple sets of annotations to the same set of data as it still needs to be created for each confounder. Where required (e.g. Table 1, row 4), the absolute phenotype change could then be added as it has been done so far with MP annotations or if desired, uniformly with post-composed phenotype annotations.
Representation of confounder association with RDF triple representation
The ‘Standardised qualifiers of pre-composed terms’ approach could be formally represented with the Resource Description Framework (RDF) triple model . In an RDF triple, the annotation conforms to the format of < subject, predicate, object>. In our scenarios this would be an MP term as the subject which would be related to the confounder body weight (the object) via the relationship specified as the “relative to” (the predicate). The triple representation is only needed in the annotation arising from including the potential confounders as covariates in the analysis and is a natural extension of the preceding approach ‘Post composed phenotypes’.
There are multiple advantages of using RDF models. The first advantage arises from the graphical nature of ontologies in which the inter-relationships of multiple tiers are captured with a graph schema. In an ontology, a class can have multiple parents leading to the inheritance of qualities from different parents, which can be well and efficiently defined within RDF models. The alternative of storing this information is to use a Relational Database Management Systems (RDBMS). In RDBMS, a table scheme is used which faces the computational challenges of multiple joins when querying across many tables and is therefore less scalable. The second advantage is that RDF is a well-established community standard recommended by the World Wide Web Consortium (W3C)  and is readily extendible. For example, the same MP term can be associated to other confounders (e.g. body length) using the same predicate. This common structure will lead to a robust data model which will improve efficiency when searching for information. The Ontology for Biomedical AssociatioN (OBAN) is an example of an RDF implementation and has been successfully exploited to represent disease-phenotype associations  (Extended version will be published within this special issue) .
RDF triples can be stored within relational as well as graphical databases and data queries are performed with the SPARQL query language . In consequence, one would need to understand the technology and the query language to work with the data effectively, through provision of a (non-SPARQL) Application Program Interface (API) would address this for accessing the data.
Conclusions and future perspectives
In gene-phenotype studies, we have identified challenges with reporting phenotypes after adjusting for body weight using currently available semantic data representation frameworks. Weight is a complex confounder, as it cannot be controlled within the experiment and causality cannot be determined. However, analysing the data with and without body weight returns a richer understanding of the phenotypic abnormality. With interest growing in the impact of body weight on phenotypes and the scale of projects being conducted by high throughput phenotyping consortiums, being able to disseminate annotated phenotype data has become an important issue. We have demonstrated that the impact of including weight as a confounder in the analysis has significant impact on the annotations returned. While this example focuses on the description of mouse phenotypes, we perceive that this is a general problem with accessing phenotypes in all mammals including humans. The current solution implemented with mouse data has arisen from adapting the mechanisms developed for curating literature to a high throughput scenario and use of the ontology for analyses.
We coordinated our efforts with Medical Research Council (MRC) Harwell and MGI in discussions on refining annotation in high throughput phenotyping studies, where MRC Harwell focused on aging studies and how to manage time course studies . The issues were determined to be distinct, as the interpretation is more complex when considering body weight as a confounder. The complexity arises as we cannot determine causality, rather we are annotating the outcome of the statistical analyses.
In the process of this study, we were able to identify several possible solutions (see ‘Challenges applying existing solutions’) that could help with applying confounder-relevant information to gene-phenotype associations. These options have been limited to what we believe have the lowest modification requirements on existing dissemination pipelines, such as those maintained by MGI. However, each of these outlined options have to be assessed now in the broader community to arrive at a conclusion what is the best to pursue.
In future work, we aim to not only communicate with the broader community to find the most suitable solution, but also to assess the impact for other potential confounders not just body weight. These additional confounders will then be verified with what has been determined as the best solution to see that it can scale with the demands of the different confounders.
While we have assessed in this study the impact of confounders of gene-phenotype associations in mouse, this is highly likely to be equally relevant in other mammalian model organisms (e.g. rat). However, we identified practical solutions based on the mouse annotation-dissemination pathways and these might not be the optimal for other model organisms. The discussions within this manuscript are a good starting point for managing confounder in their community.
analysis pipeline 1
analysis pipeline 2
application program interface
international mouse phenotyping consortium
international mouse phenotyping resource of standardised screens
mouse genome informatics database
mouse genetics project
mammalian phenotype ontology
medical research council
ontology for biomedical association
phenotype and trait ontology
relational database management systems
resource description framework
wellcome trust sanger institute
We thank staff from the Sanger Institute’s Research Support Facility, Mouse Genetics Project and Mouse Informatics Group for their excellent support.
NAK, AO, TFM were funded by the National Institutes of Health Common Fund grant: (NIH) [1 U54 HG006370-01], HP is funded by EMBL-EBI Core funds. SS is funded by the Centre for Therapeutic Target Validation. In addition, the individuals’ affiliated to the WTSI institute were also funded by the Wellcome Trust grant: WT098051.
We thank Cynthia Smith, Susan Bello and Janan Eppig from the Jackson Laboratory, James Malone of FactBio Ltd, and Ann-Marie Mallon and Michelle Simon from MRC Harwell for helpful discussion on these issues.
The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.
- Brown SD, Moore MW. Towards an encyclopaedia of mammalian gene function: the International Mouse Phenotyping Consortium. Dis Model Mech. 2012;5(3):289–92. doi:10.1242/dmm.009878.View ArticleGoogle Scholar
- IMPReSS: International Mouse Phenotyping Resource of Standardised Screens. http://www.mousephenotype.org/impress.
- Kurbatova N, Mason JC, Morgan H, Meehan TF, Karp NA. PhenStat: A Tool Kit for Standardized Analysis of High Throughput Phenotypic Data. PLoS One. 2015;10(7), e0131274. doi:10.1371/journal.pone.0131274.View ArticleGoogle Scholar
- Smith CL, Goldsmith C-AW, Eppig JT. The Mammalian Phenotype Ontology as a tool for annotating, analyzing and comparing phenotypic information. Genome Biol. 2004;6(1):R7. doi:10.1186/gb-2004-6-1-r7.View ArticleGoogle Scholar
- Hoehndorf R, Hardy NW, Osumi-Sutherland D, Tweedie S, Schofield PN, Gkoutos GV. Systematic Analysis of Experimental Phenotype Data Reveals Gene Functions. PLoS One. 2013;8 (4). doi: 10.1371/journal.pone.0060847
- Washington NL, Haendel MA, Mungall CJ, Ashburner M, Westerfield M, Lewis SE. Linking human diseases to animal models using ontology-based phenotype annotation. PLoS Biol. 2009;7(11):2662. doi:10.1371/journal.pbio.1000247.View ArticleGoogle Scholar
- Smedley D, Oellrich A, Kohler S, Ruef B, Westerfield M, Robinson P et al. PhenoDigm: analyzing curated annotations to associate animal models with human diseases. Database. 2013 Jan 1; 2013:bat025 doi: 10.1093/database/bat025.
- Karp NA, Segonds-Pichon A, Gerdin AK, Ramirez-Solis R, White JK. The fallacy of ratio correction to address confounding factors. Lab Anim. 2012;46(3):245–52. doi:10.1258/la.2012.012003.View ArticleGoogle Scholar
- Eppig JT, Blake JA, Bult CJ, Kadin JA, Richardson JE, Mouse Genome Database G. The Mouse Genome Database (MGD): comprehensive resource for genetics and genomics of the laboratory mouse. Nucleic Acids Research. 2012; Jan1;40 (D1):D881-D886. doi:10.1093/nar/gkr974.
- Greenaway S, Blake A, Retha A, O’Leary J, Goldsworthy M, Bowl MR, et al. Automatically annotating temporal data from a phenotype-driven mutagenesis screen. In: Proceedings of Phenotyping Day at ISMB, Dublin, Ireland. 2015.Google Scholar
- Reed DR, Lawler MP, Tordoff MG. Reduced body weight is a common effect of gene knockout in mice. BMC Genet. 2008;9(1):4. doi:10.1186/1471-2156-9-4.View ArticleGoogle Scholar
- Karp NA, Melvin D, Mott RF, Project SMG. Robust and Sensitive Analysis of Mouse Knockout Phenotypes. PLoS One. 2012;7 (12). doi:10.1371/journal.pone.0052410.
- Tschop MH, Speakman JR, Arch JR, Auwerx J, Bruning JC, Chan L, et al. A guide to analysis of mouse energy metabolism. Nat Methods. 2011;9(1):57–63. doi:10.1038/nmeth.1806.View ArticleGoogle Scholar
- Anika O, Meehan T, Parkinson H, Sarntivijai S, White JK, N.A. K. Supporting data: Reporting phenotypes in model organisms when considering body size as a potential confounder. 2015. https://zenodo.org/record/32082?ln=en#.ViYKnE2FOiM.
- Karp NA, Speak AO, White JK, Adams DJ, Hrabe de Angelis M, Herault Y et al. Impact of temporal variation on design and analysis of mouse knockout phenotyping studies. PLoS One. 2014;9 (10):e111239. doi:10.1371/journal.pone.0111239.
- White JK, Gerdin AK, Karp NA, Ryder E, Buljan M, Bussell JN, et al. Genome-wide Generation and Systematic Phenotyping of Knockout Mice Reveals New Roles for Many Genes. Cell. 2013;154(2):452–64. doi:10.1016/j.cell.2013.06.022.View ArticleGoogle Scholar
- MGI. Dmxl2tm1a (EUCOMM) Wtsi Phenotyping Page. http://www.informatics.jax.org/allele/MGI:4432280.
- Gkoutos GV, Mungall C, Dolken S, Ashburner M, Lewis S, Hancock J et al. Entity/Quality-Based Logical Definitions for the Human Skeletal Phenome using PATO. In Engineering In Medicine and Biology Society. 2009:7069–72. doi:10.1109/Iembs.2009.5333362.
- Mungall CJ, Gkoutos GV, Smith CL, Haendel MA, Lewis SE, Ashburner M. Integrating phenotype ontologies across multiple species. Genome Biology. 2010;11 (1). doi:10.1186/Gb-2010-11-1-R2.
- MGI. Kcne2tm1a (EUCOMM) Wtsi Phenotyping Page. http://www.informatics.jax.org/allele/MGI:4431909.
- Haendel MA, Balhoff JP, Bastian FB, Blackburn DC, Blake JA, Bradford Y et al. Unification of multi-species vertebrate anatomy ontologies for comparative biology in Uberon. Journal of Biomedical Semantics. 2014; 5 (21). doi:10.1186/2041-1480-5-21.
- Lassila O, Swick RR. Resource Description Framework (RDF) model and syntax specification. 1999.Google Scholar
- Sarntivijai S, Vasant D, Saunders G, Bento P, Gonzalez D, Betts J, et al. Linking rare and common disease: mapping clinical disease-phenotypes to ontologies in therapeutic target validation. In: Proceedings of Phenotyping Day at ISMB, Dublin, Ireland. 2015.Google Scholar
- Brush MH, Mungall C, Washington NL, Haendel M. An Ontological Characterization for Integration of Genetic Variation Data. In: What’s in a Genotype? editor. Proceedings of lnternational Conference on Biomedical Ontology. 2013. p. 105–8.Google Scholar
- Feigenbaum L, Williams GT, Clark KG, Torres E. SPARQL 1.1 protocol. Recommendation, W3C. 2013.Google Scholar
- Koscielny G, Yaikhom G, Iyer V, Meehan TF, Morgan H, Atienza-Herrero J, et al. The International Mouse Phenotyping Consortium Web Portal, a unified point of access for knockout mice and related phenotyping data. Nucleic Acids Res. 2013;42(D1):D802–9. doi:10.1093/nar/gkt977.View ArticleGoogle Scholar