Skip to main page content
U.S. flag

An official website of the United States government

Dot gov

The .gov means it’s official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2021 May 4;22(1):230.
doi: 10.1186/s12859-021-04041-7.

A new method for exploring gene-gene and gene-environment interactions in GWAS with tree ensemble methods and SHAP values

Pål V Johnsen  1   2 , Signe Riemer-Sørensen  3 , Andrew Thomas DeWan  4   5 , Megan E Cahill  4 , Mette Langaas  6
Affiliations

A new method for exploring gene-gene and gene-environment interactions in GWAS with tree ensemble methods and SHAP values

Pål V Johnsen et al. BMC Bioinformatics. .

Abstract

Background: The identification of gene-gene and gene-environment interactions in genome-wide association studies is challenging due to the unknown nature of the interactions and the overwhelmingly large number of possible combinations. Parametric regression models are suitable to look for prespecified interactions. Nonparametric models such as tree ensemble models, with the ability to detect any unspecified interaction, have previously been difficult to interpret. However, with the development of methods for model explainability, it is now possible to interpret tree ensemble models efficiently and with a strong theoretical basis.

Results: We propose a tree ensemble- and SHAP-based method for identifying as well as interpreting potential gene-gene and gene-environment interactions on large-scale biobank data. A set of independent cross-validation runs are used to implicitly investigate the whole genome. We apply and evaluate the method using data from the UK Biobank with obesity as the phenotype. The results are in line with previous research on obesity as we identify top SNPs previously associated with obesity. We further demonstrate how to interpret and visualize interaction candidates.

Conclusions: The new method identifies interaction candidates otherwise not detected with parametric regression models. However, further research is needed to evaluate the uncertainties of these candidates. The method can be applied to large-scale biobanks with high-dimensional data.

Keywords: GWAS; Gene–gene and gene–environment interactions; Model explainability; SHAP; Tree ensemble models; XGBoost.

PubMed Disclaimer

Conflict of interest statement

SINTEF DIGITAL is an institute at SINTEF. SINTEF is a non-profit research organization that provided the PhD Scholarship for first author via The Research Council of Norway. Neither The Research Council of Norway nor SINTEF DIGITAL had any role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. The authors declare that they have no competing interests.

Figures

Fig. 1
Fig. 1
An example with three constructed regression trees with six features xi,1 to xi,6 used as splitting points at each branch, and leaf node values. Also shown is the computation of f(xi) given an example of feature values xi. The structure of the trees opens the possibility to explore interactions since a path from a root node to a leaf node denotes a combination of feature values
Fig. 2
Fig. 2
The ranking, model fitting and explanation phases. In the ranking phase, the SNPs and environmental features are ordered by their relative importance. The ranking is achieved with XGBoost and SHAP values as explained in Fig. 4. In the model fitting process, the top ranked features are combined and modelled with XGBoost as described in Fig. 6. Finally, the explanations and interactions are obtained from the SHAP values. This is visualized in Figs. 10, and 11
Fig. 3
Fig. 3
All data available is divided into three subsets: Ranking data, fitting data and evaluation data. The ranking data is used to rank features by importance in order to remove noise. The fitting data is used to fit models by using the ranking derived from the ranking data. The evaluation data is finally used to explain what is considered important with respect to the predictions from the models trained on the fitting data
Fig. 4
Fig. 4
In the ranking process, multiple independent subsets are created and used in a cross-validation procedure with tree ensemble models. The trained models will be used to rank the importance of the features
Fig. 5
Fig. 5
The cross-validation phase when training data consists of F-2 specific merged folds. Training of the model will proceed as long as the performance on the validation data improves within a certain number of iterations as given by the early_stopping_rounds hyperparameter
Fig. 6
Fig. 6
Given a table of ranked feature importances, XGBoost models based on the top K features are trained in a new cross-validation procedure based on an independent set of individuals, namely the fitting data. We search for the XGBoost models that on average performs the best for a given set of hyperparameters (including the value of K) based on test data
Fig. 7
Fig. 7
Bland–Altman plot for the trained models used for ranking. No clear signs of overfitting, but one set of hyperparameters shows one cluster of poorer predictions than the others
Fig. 8
Fig. 8
The reason some models with hyperparameter set 1 in the ranking process underperforms is early stopping of the training. Notice also that a larger number of trees need to be constructed to get the same performance as for models with other hyperparameter sets
Fig. 9
Fig. 9
The model fitting process based on top K features from both the ranking process (for different sets of XGBoost-hyperparameters indicated by the different colours and the legend) and from BOLT-LMM, for different values of K. In a hyperparameter sets 1–8 (all with max_depth=2) from Table 4 in the model fitting process are used. In b hyperparameter sets 9–16 (max_depth=3) are used. Both figures show that the use of the ranking process gives in general better model performance than for the BOLT-LMM ranking. There is also some gain in performance by increasing the hyperparameter max_depth from two to three
Fig. 10
Fig. 10
A sina plot visualise the importance of each feature from a fitted model. Here we show the sina plot of the marginal effects for one of the four models constructed during the model fitting process when applied to the evaluation data from UK Biobank
Fig. 11
Fig. 11
Partial dependence plots for the pairs a rs180743 and rs171329, b rs17817449 and genetic sex, c rs17817449 and saturated fat intake, and d rs12123815 and rs12123815. In all panels we see how the SHAP values (vertical axis) depends on the feature value of the SNP (horizontal axis) and on the value of the second feature (color)
Fig. 12
Fig. 12
Stratified analysis of the top four interactions based on all unrelated individuals to illustrate how the log odds ratio, with 95% confidence intervals, of one feature changes depending on the value of another feature
See this image and copyright information in PMC

Similar articles

See all similar articles

Cited by

See all "Cited by" articles

References

    1. Loh PR, Tucker G, Bulik-Sullivan BK, Vilhjálmsson BJ, Finucane HK, Salem RM, et al. Efficient Bayesian mixed model analysis increases association power in large cohorts. Nature Genet. 2015;47(3):284–90. doi: 10.1038/ng.3190. - DOI - PMC - PubMed
    1. Visscher PM, Wray NR, Zhang Q, Sklar P, McCarthy MI, Brown MA, et al. 10 years of GWAS discovery: biology, function, and translation. Am J Human Genet. 2017;101(1):5–22. doi: 10.1016/j.ajhg.2017.06.005. - DOI - PMC - PubMed
    1. Yu J, Pressoir G, Briggs WH, Vroh Bi I, Yamasaki M, Doebley JF, et al. A unified mixed-model method for association mapping that accounts for multiple levels of relatedness. Nature Genet. 2006;38(2):203–8. doi: 10.1038/ng1702. - DOI - PubMed
    1. Maher B, Maher B, editor. Personal genomes: the case of the missing heritability [News]. Nature. 2008. - PubMed
    1. Burton PR, Hansell AL, Fortier I, Manolio TA, Khoury MJ, Little J, et al. Size matters: just how big is BIG? Quantifying realistic sample size requirements for human genome epidemiology. Int J Epidemiol. 2009. - PMC - PubMed

LinkOut - more resources