Structure-adaptive canonical correlation analysis for microbiome multi-omics data Authors: Linsui Deng, Yanlin Tang, Xianyang Zhang, Jun Chen Published: 2024 ABSTRACT Sparse canonical correlation analysis (sCCA) has been a useful approach for integrating different high-dimensional datasets by finding a subset of correlated features that explain the most correlation in the data. In the context of microbiome studies, investigators are always interested in knowing how the microbiome interacts with the host at different molecular levels such as genome, methylol, transcriptome, metabolome and proteome. sCCA provides a simple approach for exploiting the correlation structure among multiple omics data and finding a set of correlated omics features, which could contribute to understanding the host-microbiome interaction. However, existing sCCA methods do not address compositionality, and its application to microbiome data is thus not optimal. This paper proposes a new sCCA framework for integrating microbiome data with other high-dimensional omics data, accounting for the compositional nature of microbiome sequencing data. It also allows integrating prior structure information such as the grouping structure among bacterial taxa by imposing a “soft” constraint on the coefficients through varying penalization strength. As a result, the method provides significant improvement when the structure is informative while maintaining robustness against a misspecified structure. Through extensive simulation studies and real data analysis, we demonstrate the superiority of the proposed framework over the state-of-the-art approaches. 1 Introduction The human microbiome is the collection of microorganisms and their genetic makeup associated with the human body. It plays a critical role in human health and disease ranging from gastrointestinal diseases to various cancers (Sepich-Poore et al., 2021). To gain more mechanistic insights, multi-omics approaches have been increasingly employed in microbiome studies to elucidate the intricate interplay between the environment, the human microbiome and the host at different molecular levels (Hasin et al., 2017; Lloyd-Price et al., 2019). Although many multi-omics datasets have been generated in the past few years, it is unclear how to integrate them efficiently. One useful tool for multi-omics data integration is to perform canonical correlation analysis (CCA). CCA, due to Hotelling (1936), connects two sets of variables by finding a linear combination of variables that maximally correlate. However, the standard CCA fails when the sample size is strictly less than the number of variables as one can find meaningless solutions with correlations equal to one. Also, it does not perform variable selection and hence lacks interpretability. To circumvent these problems, sparse CCA (sCCA) has been proposed, aiming to find pairs of sparse canonical directions by imposing sparsity penalty. The first sCCA algorithm was presented by Parkhomenko et al. (2007), which, however, lacks exact criterion and biconvexity. Witten et al. (2009) applied the penalized matrix decomposition to cross-product matrix and yielded a straightforward formulation for sCCA. Some closely related methods include Parkhomenko et al., 2009; Lê Cao et al., 2009. Hardoon and Shawe-Taylor (2011) expressed the sCCA model as a primal-dual Rayleigh quotient, which takes the primal representation and kernel representation as the first view and second view, respectively. Chu et al. (2013) reformed CCA into a trace maximization problem and computed the sparse solution by the linearized Bregman method. To exploit the potential structural information among features, various forms of structure-adaptive sCCA have been proposed (Lin et al., 2013; Chen et al., 2012; Mohammadi-Nejad et al., 2017). In particular, Chen et al. (2013) proposed the structure-constrained sCCA (ssCCA) to exploit the phylogenetic structure in microbiome data. Advances in next-generation sequencing technologies have enabled the direct sequencing of microbial DNA to determine microbiome composition, using either targeted or shotgun approaches (Wensel et al., 2022). The resulting microbiome data is typically in the form of a count table that records the frequencies of detected taxa in specific samples. However, due to the complexities inherent in the sequencing process, the total count for a sample reflects the sequencing effort rather than the actual microbial load at the sampling site. Consequently, microbiome data are inherently compositional, meaning that we only have information about the relative abundances of taxa. This compositionality presents significant challenges in the statistical analysis of microbiome data. A change in the (absolute) abundance of one taxon can lead to apparent changes in the relative abundances of all other taxa, complicating the identification of the actual causal taxa (Yang and Chen, 2022). The compositional nature also renders many standard multivariate statistical models inappropriate or inapplicable (Aitchison, 1982). Many efforts have been made to address the compositionality in different contexts of microbiome data analysis. For example, Friedman and Alm (2012) developed an iterative procedure named SparCC that allows inference of correlations for compositional data by assuming that the number of taxa is large and the true correlation network is sparse. Lin et al. (2014) dealt with the variable selection in regression with compositional covariates. Jiang et al. (2019) addressed zero inflation and detected pairs of associated compositional and non-compositional covariates using a Bayesian zero-inflated negative binomial regression model. However, existing CCA methods including ssCCA could not address the compositional effects, potentially reducing its precision in recovering relevant taxa. We propose a new sCCA framework for integrating microbiome data with other high-dimensional omics data. The framework specifically addresses the compositional nature of the microbiome data. It also allows integrating prior structure information by imposing a “soft” constraint on the coefficients through varying penalization strength. As a result, the method provides significant improvement when the structure is informative while maintaining robustness against a misspecified structure. The developed tool aims to be an important resource for investigators to understand the interplay between the microbiome and host, decipher the molecular mechanisms underlying microbiome-disease association, and identify potential microbial targets for intervention. This paper is organized as follows. Section 2 introduces the new sCCA framework for integrating microbiome compositional data with (non-)compositional high-dimensional data. Section 3 extends the new framework to incorporate additional prior structural information. In Section 4, we conduct numerical simulations to demonstrate the effectiveness of our proposed methods. Section 5 applies the proposed methods in a real microbiome study to investigate the association between gut bacteria and its metabolic output. We conclude with a discussion in Section 6. 2 Compositional sCCA 2.1 Formulation Let us consider two random vectors X=(X1,…,Xp)⊤ and Y=(Y1,…,Yq)⊤ , where X contains the composition of p taxa and Y is a q -dimensional vector of non-compositional covariates. The nature of the composition makes X lie in a (p−1) -dimensional positive simplex. To address the compositionality, Aitchison and Bacon-Shone (1984) proposed applying the log-ratio transformation to compositional covariates resulting in Z/p=(log(X1/Xp),…,log(Xp−1/Xp)) , where Xp is chosen as the reference component. CCA for compositional data can be formulated to find canonical coefficients a=(a1,…,aq)⊤ and b−p=(b1,…,bp−1)⊤ so that the correlation between a⊤Y and b−p⊤Z/p is maximized. Note that b−p⊤Z/p=∑j=1p−1bj⁡log(Xj/Xp)=∑j=1pbj⁡log(Xj) with bp=−∑j=1p−1bj . To avoid the choice of a reference component, we can write the term b−p⊤Z/p in a symmetric form by noticing that b−p⊤Z/p=b⊤Z , where Z=(log(X1),…,log(Xp)) and b=(b1,…,bp) with b∈Bp≔c=c1,…,cp⊤:∑j=1pcj=0. Therefore, the compositional CCA aims to find a~ and b~ such that a~,b~=argmaxa∈Rq,b∈BpCorra⊤Y,b⊤Z=argmaxa∈Rq,b∈BpCova⊤Y,b⊤ZVara⊤Yvarb⊤Z. When the dimensions p and q are high (as compared to the sample size), regularization is required to encourage sparsity and to obtain a unique solution to the optimization problem. We let ΣYZ=Cov(Y,Z) , ΣY=Cov(Y,Y) and ΣZ=Cov(Z,Z) . Define the weighted l1 norm based on a vector of non-negative weights w=(w1,…,wp) for a vector b as ‖b‖1,w=∑j=1pwj|bj| . The compositional sCCA problem can then be formulated as maxa∈Rq,b∈Rpa⊤ΣYZbs.t.a⊤ΣYa≤1, ‖a‖1≤Ca, b⊤ΣZb≤1, ‖b‖1,w≤Cb, b∈Bp. (1) Here Ca,Cb>0 are some positive tuning parameters that control the global shrinkage level. The weight wj allows different penalization strengths according to the data or prior structure information. We will elaborate it in Section 2.3. It has been shown that in high dimensions, treating the covariance matrix as diagonal can yield good results. Following the same strategy adopted by many of the existing high-dimensional CCA algorithms (e.g., Witten et al., 2009), we substitute in the identity matrix for ΣY and ΣZ in the CCA formulation (Equation 1). Moreover, we write the (weighted) l1 constraints on a and b in the Lagrangian form. Given a set of n samples {Xi,Yi}i=1n , let Σ^YZ be the sample cross-covariance between Y and Z . We formulate the feasible CCA problem as mina∈Rq,b∈Rp−a⊤Σ^YZb+λa‖a‖1+λb‖b‖1,ws.t.‖a‖2≤1, ‖b‖2≤1, b∈Bp, which can be solved by iteratively optimizing the objective function with respect to one parameter while fixing the other parameter. Specifically, we have the following two updating steps.1. Update b : Fix a(t) and update b through bt←argminb∈Rp−at⊤Σ^YZb+λb‖b‖1,w s.t. ‖b‖2≤1, b∈Bp. (2) 2. Update a : Fix b(t) and update a through at+1←argmina∈Rq−a⊤Σ^YZbt+λa‖a‖1 s.t. ‖a‖2≤1. (3) 1. Update b : Fix a(t) and update b through 1. Update b : Fix a(t) and update b through 2. Update a : Fix b(t) and update a through 2. Update a : Fix b(t) and update a through 2.2 Algorithm In this section, we discuss the updates in Equations 2, 3. Define the operator gh,λ,w=argminb∈Bp12‖h−b‖22+λ‖b‖1,w. By exploring the Karush-Kuhn-Tuchker conditions, we obtain the following result. Proposition 2.1. Set b˘=argminb∈Bp−(a(t))⊤Σ^YZb+λb‖b‖1,w . The solution to (2) is given by bt=b˘,if ‖b˘‖2≤1,gΣ^YZ⊤at,λb,w‖gΣ^YZ⊤at,λb,w‖2,if ‖b˘‖2>1. Remark 2.1. We employ the augmented Lagrangian method (ALM) to solve the optimization problem in g(h,λ,w) . Specifically, the ALM involves the following two steps of iterations bt+1←argminb12‖h−b‖22+λb‖b‖1,w+μ121⊤b+dt2,dt+1←dt+μ21⊤bt+1, where μ1,μ2>0 are the step sizes in dual gradient ascent, which are set to be 1 in our numerical studies. The optimization problem in the first step can be solved using coordinate descent in an inner loop by iterating across the following p components bjt+1,r+1=11+μ1Shj−μ1∑ijbit+1,r+dt,λbwj, where S(a,λ)=sign(a)(|a|−λ)+ denotes the soft-thresholding operator and hj is the j th component of h . Using similar arguments, we can show that the solution to (3) is given by at+1=a˘,if ‖a˘‖2≤1,SΣ^YZbt,λaSΣ^YZbt,λa2,if ‖a˘‖2>1, where a˘=argmina∈Rq−a⊤Σ^YZb(t)+λa‖a‖1 and S(a,λ)=(S(a1,λ),…,S(ap,λ))⊤ . Set v=(v1,…,vq)⊤=Σ^YZb(t) . The objective function in the definition of a˘ becomes ∑j=1q(−ajvj+λa|aj|) . We see that aj(t+1)=0 if λa≥vj , and |aj(t+1)|=∞ if λa0 . We can also extend our approach to learn sub-spaces from multiple views, i.e., when we have multiple groups of measurements, X(i)∈Rpi , for i=1,…,m on matching samples. This situation naturally arises in multi-omics studies, where methylomic, transcriptomic, metabolomic, and microbiome data are collected from a single group of individuals. A number of approaches for generalizing CCA to multiple views have been proposed in the literature, and some of these extensions are summarized in Kettenring (1971). We could adapt these multi-view methods to incorporate the compositional constraint. There are several limitations to our framework. First, we did not simultaneously address the zero inflation commonly observed in microbiome data. We used a simple zero replacement strategy before running the C-sCCA. Although this strategy has been commonly used in microbiome data analysis at log scale (Zhou et al., 2022; Lin and Peddada, 2020), better methods can be developed such as replacing the log scale transformation by modified centered log-ratio transform designed for addressing zero inflation (Yoon et al., 2019), imposing another multinomial layer to account for sampling variability associated with the sequencing process (Chen and Li, 2013), or using more informative imputation methods such as mbDenoise (Zeng et al., 2022) and mbImpute (Jiang et al., 2021). Second, although our framework can select subsets of features that explain the largest correlation between the datasets, their detailed relationships can not be learned simultaneously. Developing methods that combine feature-level selection with the construction of feature-feature correlation networks is a promising area for future research. Third, we can enhance robustness to outliers through several strategies, including outlier detection and removal during data preprocessing, replacing empirical covariance estimators with robust estimators (Luo et al., 2024), and investigating the optimal choice of penalty functions (Chalise and Fridley, 2012), such as Huber loss, Tukey loss, or L0 penalty (Lindenbaum et al., 2022). Fourth, while our work focuses on linear associations, future extensions of our composite sCCA framework could capture nonlinear relationships by integrating with kernel CCA (Akaho, 2001; Fukumizu et al., 2007), deep CCA (Andrew et al., 2013), or nonparametric CCA (Lancaster, 1958; Michaeli et al., 2016). Lastly, our framework assumes that the association is mediated through the ratios of the compositional components since only ratios are meaningful for compositional data. However, when the association is at the level of absolute abundance - where the total microbial load also matters - our method may not work well. This limitation is more inherent to the constraints of current sequencing technologies than to the method itself.