Integrating biological knowledge into variable selection: an empirical Bayes approach with an application in cancer biology
© Hill et al.; licensee BioMed Central Ltd. 2012
Received: 13 June 2011
Accepted: 19 April 2012
Published: 11 May 2012
An important question in the analysis of biochemical data is that of identifying subsets of molecular variables that may jointly influence a biological response. Statistical variable selection methods have been widely used for this purpose. In many settings, it may be important to incorporate ancillary biological information concerning the variables of interest. Pathway and network maps are one example of a source of such information. However, although ancillary information is increasingly available, it is not always clear how it should be used nor how it should be weighted in relation to primary data.
We put forward an approach in which biological knowledge is incorporated using informative prior distributions over variable subsets, with prior information selected and weighted in an automated, objective manner using an empirical Bayes formulation. We employ continuous, linear models with interaction terms and exploit biochemically-motivated sparsity constraints to permit exact inference. We show an example of priors for pathway- and network-based information and illustrate our proposed method on both synthetic response data and by an application to cancer drug response data. Comparisons are also made to alternative Bayesian and frequentist penalised-likelihood methods for incorporating network-based information.
The empirical Bayes method proposed here can aid prior elicitation for Bayesian variable selection studies and help to guard against mis-specification of priors. Empirical Bayes, together with the proposed pathway-based priors, results in an approach with a competitive variable selection performance. In addition, the overall procedure is fast, deterministic, and has very few user-set parameters, yet is capable of capturing interplay between molecular players. The approach presented is general and readily applicable in any setting with multiple sources of biological prior knowledge.
Ongoing advancements and cost reductions in biochemical technology are enabling acquisition of ever richer datasets. In many settings, in both basic biology and medical studies, it may be important to model the relationship between assayed molecular entities, such as genes, proteins or metabolites, and a biological response of interest.
Molecular players may act in concert to influence biological response: this has motivated a need for multivariate methods capable of modelling such joint activity. When sample sizes are small-to-moderate, as is often the case in molecular studies, robust modelling of joint influences becomes especially challenging. However, often it is likely that only a small number of players are critical in influencing the response of interest. Then, the challenge is to identify appropriate variable subsets.
Statistical variable selection methods have been widely used in the bioinformatics domain to discover subsets of influential molecular predictors. Both penalised likelihood and Bayesian approaches have been used in a diverse range of applications[1–6].
Bayesian approaches can facilitate the integration of ancillary information regarding variables under study through prior probability distributions. Ongoing development of online tools and databases have meant that such information is widely available, and depending on context, may include networks and pathway maps, public gene expression datasets, molecular interaction databases, ontologies and so on. However, while the idea of incorporating such information into variable selection has a clear appeal, it is not always obvious what information should be included nor how it should be weighted. Indeed, many existing Bayesian variable selection approaches do not attempt integrative analyses exploiting such information and instead employ standard priors that do not specify preferences for particular variables, but may, for example, encode a preference for sparse models[4, 7]. Several Bayesian variable selection studies have put forward simple approaches for incorporating prior knowledge by independently assigning each variable a prior probability of being included in the model[1, 6, 8, 9]. However, subjectively setting such hyperparameters for each variable may be difficult. Furthermore, prior independence may be a questionable assumption, since molecular variables are unlikely to influence a response independently of one another.
We develop a variable selection procedure in which an empirical Bayes approach is used to objectively select between a choice of informative priors incorporating ancillary information (‘biologically informative priors’) and also to objectively weight the contribution of the prior to the overall analysis. The work presented here is motivated by questions concerning the relationship between signalling proteins and drug response in human cancers. In the protein signalling setting (as also in gene regulation) there is now much information available, both in the literature and in diverse online resources, concerning relevant pathways and networks. We therefore develop pathway- and network-based informative priors for this setting, applying the methods proposed to automatically select and weight the prior and thence carry out variable selection.
The relationship between response and predictors is modelled using a continuous, linear model with interaction terms. In this way we avoid data discretization (which can lose information) yet retain the ability to capture combinatorial interplay. We take advantage of biochemically-motivated sparsity constraints to permit exact inference, thereby avoiding the need for approximate approaches such as Markov chain Monte Carlo (MCMC). This enables the calculation of exact probability scores over which variables are likely to be influential. The overall procedure is computationally fast: empirical Bayes analysis and subsequent calculation of posterior (inclusion) probabilities for 52 predictors via full model averaging required only 10 minutes (in MATLAB R2010a on a standard single-core personal computer; code freely available, together with simulation scripts, athttp://go.warwick.ac.uk/stevenhill/IBKVS). Moreover, the overall procedure we put forward is simple from the user perspective, requiring very few user-set parameters or MCMC convergence diagnostics.
The remainder of the paper is organised as follows. We begin below by defining notation and reviewing Bayesian variable selection. We then describe methods, including empirical Bayes analysis to objectively select and weight biologically informative prior information, pathway-based informative priors and exact inference. We illustrate our method on published single cell proteomic data and on proteomic data and drug response from ongoing work in breast cancer. We also compare the proposed approach to alternative methods. We conclude with a discussion of our results, the merits and shortcomings of our work, and highlight directions for further work.
Let be a vector of response values and for be corresponding p-dimensional candidate predictors. X i forms row i of the n × p predictor matrix X.
Let be a binary vector and be the number of non-zeros in γ. Then X γ is the matrix obtained by removing from X those columns j for which γ j = 0. Similarly, for a vector, a γ is obtained from a by removing components a j for which γ j = 0.
Bayesian variable selection
Bayesian linear model
where p(Y | γ,X γ ) is the marginal likelihood. P(γ) is a prior over models (the ‘model prior’), and is the main focus of the present paper. The marginal likelihood is obtained by integrating out the regression parameters β γ and variance parameter σ2 and thereby automatically penalises complex models with many parameters. This penalisation occurs because a more complex model has a larger parameter space. This means that for more complex models, the integral that defines the marginal likelihood is over a greater number of dimensions, with prior mass spread over a larger space. This in turn results in a lower marginal likelihood score.
Model selection and model averaging
These inclusion probabilities are a measure of the importance of each individual predictor in determining the response.
Evaluating the summation in Equation 3 requires enumerating the entire posterior over models P(γ | Y X). The model space Γ can be vast () even for moderate values of p . Thus, Markov chain Monte Carlo is often used to sample from the posterior over models thereby providing asymptotically valid estimates of the inclusion probabilities. As outlined in Methods below, we instead calculate exact inclusion probabilities; an approach rendered computationally viable through restricting the size of model space Γ. Justifications for such a restriction and advantages and disadvantages of an exact approach are provided in the Discussion below.
Calculating the posterior distribution over models (2) requires specifying a prior over Γ, P(γ). A common choice of prior assumes that the a priori inclusion probabilities are independent and have Bernoulli distributed marginal distributions P(γ j ) with success parameter Π. This hyperparameter may be a user-defined constant or may itself have a Beta prior[9, 14]. In the former case, small values are often chosen to promote parsimonious models.
These priors provide no information regarding specific predictors and do not utilise domain knowledge. Employing predictor dependent hyperparameters Π j enables incorporation of prior knowledge that some predictors are more important than others. However, utilising such a prior may be difficult in practice due to the many hyperparameters that must be subjectively specified. We note also in this formulation, prior inclusion probabilities are still independent.
We now describe the Bayesian variable selection method used in the present work. We describe in turn, an extended linear model including interactions between predictors, exact computation of posterior inclusion probabilities, biologically informative model priors and empirical Bayes learning of associated hyperparameters.
Bayesian linear model with interaction terms
We extend the classical linear model in Equation 1 above to enable combinatorial relationships between predictors and response to be captured. Given model γ, response Y i depends in a non-linear fashion on the included predictors X iγ whilst remaining linear in the regression parameters. In particular, the mean for Y i is a linear combination of included predictors and all possible products of included predictors. For example, if with γ3 = γ5 = 1, we have Y i = X iγ β γ + α Xi 3Xi 5 + ε i . We extend the predictor matrix X γ and regression coefficient vector β γ to include the interaction terms and coefficients respectively, and we denote the extended versions by and. All columns in are standardised to have zero mean and unit variance.
Exact posterior inclusion probabilities
We enforce a restriction on the number of predictors that are allowed to be included in the model. That is, we only allow γ with for some. Thus, instead of being exponential in p, the model space Γ has polynomial size of order, thereby allowing explicit calculation of posterior inclusion probabilities via Equation 3. We take d max = 4, giving for the p = 52 predictors in the cancer drug response application below; the original size of Γ was of order 1015.
Biologically informative model priors
We now turn our attention to the model prior P(γ). In many molecular biology settings, there is much valuable information available which may be used to construct biologically informative model priors. This could be network and pathway structures, providing information on relationships between predictors, or information from publicly available datasets. However, it may not be obvious precisely how such information should be used and it is usually possible to encode several different, apparently plausible priors. We are therefore interested in investigating the question of how to choose between such priors.
where m is a hyperparameter (the ‘source parameter’) that selects amongst priors and λ is a hyperparameter controlling the overall strength of the prior.
Number of pathways (f1)
This definition prevents the empty model being a priori most probable and avoids double counting (proteins that are members of multiple pathways are considered to be a member of only one pathway for the purpose of calculating f1(γ), and this single pathway is selected to minimise the pathway count; see Figure1). If the strength parameter λ is negative, the prior increasingly penalises models as number of pathways increases, whereas a positive value results in a prior that prefers models representing many pathways.
Intra-pathway distance (f2)
The second feature we consider is that variables which jointly influence the response may either be close to each other in a network sense, or may in fact be far apart in the network. This is done by a function f2(γ) which gives the average distance between pairs of proteins that are both in γ and in the same pathway. Specifically, the distance between two proteins j1 and j2, denoted d(j1,j2), is the number of edges in the shortest (undirected) path between them. Then, we define where D γ is the average of all d(j1,j2) with for some k. In order for the distance to be defined for any two proteins in a pathway, we assume that the network topology for a pathway consists of a single connected component (in the undirected sense). We term a protein included in γ as a singleton if there are no other included proteins in the same pathway (i.e. protein j is a singleton if for some k). For models that only contain singletons or the empty model we set D γ = 0. The function f2(γ) defined in this way satisfies a number of natural desiderata. It is agnostic to |γ| and to the pathway count K γ (see Figure1). Also, it avoids double counting (the distance between each protein pair contributes to f2(γ) once only) and is indifferent between models including only singletons and models with the smallest possible average distance of D γ = 1. A negative strength parameter λ results in a prior that penalises larger intra-pathway distances, whilst a positive value encourages larger distances.
For a given choice of hyperparameters, the marginal likelihood can be calculated exactly by exploiting the model space restriction described above. The score is calculated for varying hyperparameters and those resulting in the largest score are used for variable selection.
We first show an application of our proposed approach to synthetic response data generated from a published study of cell signalling, and then further illustrate the approach with an analysis of proteomic data and drug response from breast cancers.
Synthetic response data
In ongoing studies, such as that presented below, truly objective performance comparisons may be challenging, since we usually do not know which molecules are truly influential in driving biological response. At the same time, in fully synthetic data it can be difficult to mimic realistic correlations between variables within a pathway or across a network. For this reason, we empirically assessed the methods proposed using published single-cell, phospho-proteomic data with responses generated from that data. This preserved pathway-related correlation structure between predictors but permitted objective assessment. The dataset consists of 11 proteins and n tot = 853 samples. (The complete dataset from contains data obtained under nine different conditions, corresponding to different interventions. Here, we use the baseline dataset which contains 853 samples).
We first considered two simulation models, and, each of which is a predictor subset consisting of three proteins; PIP3, ERK1/2, p38 for Simulation 1, and RAF, MEK1/2, PKA for Simulation 2. In each case, the three proteins were chosen to be favoured by a particular prior. is favoured by the intra-pathway distance prior (f2) with positive λ; the proteins included in had a large average intra-pathway distance and incorporated a medium number of pathways. is favoured by either the number of pathways prior (f1) with negative λ or the intra-pathway distance prior (f2) with negative λ; the proteins included in had both a small intra-pathway distance and incorporated a small number of pathways. Since, by construction, each model is favoured by a particular prior, we can test the ability of the empirical Bayes approach to select appropriate hyperparameter values. Response data Y were generated using a linear model with interaction terms (4); Y=A + BC + ε, where A,B,C are the three influential variables.
We are especially interested in the small-sample regime that is often of interest in molecular studies. We therefore subsampled (without replacement) n=35 training data from the complete dataset (this matched the sample size of the drug response study reported below), and assessed predictive ability on the remaining, held-out data ().
Subsampling was repeated to give 5,000 training/test pairs, over which results are reported below. At each iteration, only small-sample training data was used for inference. The empirical Bayes method was employed to set prior source and strength parameters (using training data only), with (this specification permits a flat prior if empirical Bayes analysis supports neither prior). Posterior inclusion probabilities were then calculated as described above.
BVS with flat prior and linear model with interaction terms (‘BVS: flat +int’);
BVS with a prior that is incorrect with respect to the true, underlying model: intra-pathway distance prior (f 2) favouring small distances (λ = − 5) for Simulation 1 and large distances (λ = 5) for Simulation 2 (‘BVS: incorrect +int’);
BVS with flat prior and linear model with no interaction terms (‘BVS: flat -int’);
BVS with a Markov random field prior  and linear model with interaction terms (see below for further details; ‘BVS: MRF prior +int’);
penalised-likelihood Lasso regression  using a linear model with no interaction terms (see below for further details; ‘Lasso -int’);
penalised-likelihood Lasso regression  using a linear model with pair-wise interaction terms (‘Lasso +int’);
a penalised-likelihood approach, proposed by Li and Li , based on the Lasso and also incorporates network structure information (‘Li&Li’); and
absolute correlation coefficients between each predictor and response (‘corr’).
The strength parameter λ is usually constrained to be non-negative, resulting in a prior that encourages selection of predictors whose neighbours in G are also included in the model. Here, we do not enforce this constraint and also allow negative values for λ. Negative values result in a prior that penalises models containing predictors that are neighbours in G. As with the proposed Bayesian variable selection method, we use a linear model with interaction terms and set λ with empirical Bayes. The graph structure G is obtained from the structure shown in Figure2 by converting all directed edges to undirected edges.
Lasso regression performs variable selection by placing an ℓ1 penalty on the regression coefficients. This has the effect of shrinking a subset of regression coefficients to exactly zero; the predictors with non-zero coefficients are taken as the inferred model. Sparsity of the inferred model is controlled by a tuning parameter, which we set by 5-fold cross-validation. This method results in a single inferred model (i.e. point estimate). However, a full ROC curve can still be obtained by thresholding absolute regression coefficients.
The penalised-likelihood method proposed by Li and Li combines a Lasso penalty with an additional penalty term that incorporates predictor network structure. Together, these penalties encourage sparse estimates for regression coefficients that are also ‘smooth’ over the network structure; that is, coefficients for neighbours in the network are encouraged to be similar. In terms of variable selection, the approach promotes models containing predictors that are neighbours in the graph. The two penalty terms each have their own tuning parameter, controlling sparsity and smoothness over the network respectively; we set these parameters by 5-fold cross-validation. (This approach, and the standard Lasso regression approach, were implemented using Matlab package glmnet.)
We observe that, in both simulations, the automated empirical Bayes analysis, with pathway-based priors, improves performance over the flat prior and provides substantial gains over an incorrect prior. The empirical Bayes approach selected the correct prior in 85% of iterations for Simulation 1 and 96% of iterations for Simulation 2 (for Simulation 1 correct prior parameters were m = 2 with λ > 0, median value of λ selected was λ = 3.5; for Simulation 2 correct prior parameters were m = 1 or m = 2 with λ < 0, median value of λ selected was λ = −5 for both m = 1 and m = 2). Since the Lasso regression method (with interaction terms) does not incorporate prior information, it is unsurprising that it is also outperformed by the empirical Bayes approach. Hence, it is fairer to compare it to Bayesian variable selection with a flat prior (and interaction terms). In Simulation 1 these regimes both show a similar performance with the Lasso approach displaying some gains at small numbers of false positives. However, in Simulation 2 the Bayesian approach offers a clear improvement in performance over Lasso regression (AUC scores of 0.91 and 0.83 respectively). Due to its inability to model combinatorial interplay, the linear model without interaction terms is outperformed by the linear model with interaction terms for both Bayesian and Lasso approaches.
In Simulation 1, the strength parameter for the Markov random field prior was set to λ = −5 by empirical Bayes in 86% of iterations, thereby correctly promoting models that do not contain predictors that are neighbours in the network. However, although the Markov random field prior offers improvements over a flat prior, it is outperformed by the proposed pathway-based priors at small numbers of false negatives. This is due to the fact that our intra-pathway distance prior is able to promote models with large distances between predictors, whilst the Markov random field prior can only penalise models that contain neighbours. The Markov random field prior is in general less flexible because it considers neighbours rather than distances. Intriguingly, in Simulation 2, λ > 0 (which correctly promotes models containing neighbours in the network) was only selected by empirical Bayes in 41% of iterations. As a result, performance of the Markov random field prior is inferior to a flat prior. We discuss this further in Discussion below.
The penalised-likelihood approach proposed in, incorporating network information, performs poorly in both simulations, with similar performance compared to simply looking at correlations between predictors. Whilst in Simulation 2, a clear improvement is observed over standard Lasso regression (without interaction terms), this is not the case in Simulation 1. This is because the approach promotes models containing predictors that are neighbours in the network, which reflects the true underlying model for Simulation 2 only. The general poor performance of this approach is likely due to its inability to capture combinatorial interplay since it necessarily employs a linear model without interaction terms.
Synthetic response data, predictive errors from held-out test data
BVS: EB prior†
BVS: flat prior†
BVS: ‘incorrect’ prior†
BVS: MRF prior†
The failure of the incorrect prior illustrates the importance of prior elicitation. Moreover, our results demonstrate that the proposed empirical Bayes approach can select a suitable prior automatically, even under very small sample conditions (here n = 35). If the data is not in agreement with a proposed prior, then it is desirable that λ = 0 is selected by empirical Bayes, resulting in a flat prior. To test this, we used the model in Simulation 2 with a prior that favoured models with predictors from many pathways (i.e. number of pathways prior with λ restricted to be non-negative). This prior does not reflect the true, underlying model, which contains a small number of pathways. Empirical Bayes analysis successfully selected λ = 0 in 95% of iterations.
The only user-set parameters in the proposed method are d max (the maximum number of predictors allowed in a model), and the range of values for the prior strength parameter λ to optimise over in empirical Bayes. We sought to check the sensitivity of our results to these parameters. As described in ‘Methods’ above, we set d max = 4 and considered. We compared the posterior inclusion probabilities inferred from 50 iterations of Simulation 1 to those obtained using (i) an increased maximum number of included predictors of d max = 5; (ii) Markov chain Monte Carlo-based (MCMC) inference with no restriction on number of included predictors, and (iii) an increased range for the prior strength (see Figure4). We found very close agreement in all cases, indicating that results reported do not depend on the sparsity restriction or the chosen range for λ.
In Simulation 2 and Simulation 3, the smallest value of λ = −5 was selected by empirical Bayes in a majority of iterations. The true, underlying model has the minimum possible number of pathways and intra-pathway distance. Hence, the strong (negative) prior strength is appropriate because it causes the prior to heavily penalise any model not satisfying these minima. Under the increased range for λ, the smallest value (λ = −10) was still selected in these iterations, but results were almost identical (as seen for Simulation 1 in Figure4). This indicates that the prior was already having close to maximal influence at the lower value of λ = −5.
Cancer drug response data
Aberrant signalling is heavily implicated in almost every aspect of cancer biology[24, 25] and, as a result, signalling proteins are targets for many emerging cancer therapies. Here, we apply the methods proposed to probing phospho-proteomic influences on response to an anti-cancer agent Triciribine.
Phospho-protein abundance was assayed in a high-throughput manner using the KinetWorksTM system (Kinexus Inc, Vancouver, Canada), for p = 52 proteins related to epidermal growth factor receptor (EGFR) signalling, in each of n = 35 breast cancer cell lines (see Additional File1:Sections 1.1-1.2 for details). The EGFR signalling network plays a central role in breast cancer biology and the cell lines used have previously been shown to retain much of the biological heterogeneity of primary tumours. GI50 (log transformed) was used to quantify response to Triciribine for each of the 35 cell lines. GI50 is the concentration that causes 50% growth inhibition compared to a baseline. A network (with a total of five pathways) was constructed usinghttp://cellsignal.com (see Figure5).
Drug response data, predictive errors from cross-validation
BVS: EB prior +int
BVS: flat prior +int
BVS: ‘incorrect’ prior +int
BVS: MRF prior +int
Figure7 shows posterior inclusion probabilities obtained under three prior regimes: empirical Bayes (intra-pathway distance prior with λ = 5), flat prior and an "incorrect" prior that is not optimal according to the empirical Bayes analysis (number of pathways prior with λ = −5). Phospho-IR and phospho-RB(S259) stand out in the empirical Bayes analysis. Triciribine targets AKT, which inhibits apoptotic processes and is heavily implicated in cancer signalling. IR (insulin receptor) is a tyrosine kinase receptor, known to stimulate the AKT pathway, and it has been suggested that the RB/E2F pathway, which is also known to play a role in cancer, has an effect on AKT activity via transcriptional regulation. Hence, the salience of IR and RB accords with known biology and drug mechanism. The MAP model for each prior regime is highlighted in red in Figure7. We note that these models do not always contain the proteins with highest inclusion probabilities.
We performed Leave-One-Out-Cross-Validation (LOOCV), making predictions for the held-out test sample using both posterior model averaging (Equation 10) and the MAP model (Equation 11). The full variable selection approach, including selection of hyperparameters with empirical Bayes, was carried out at each cross-validation iteration. Table2 shows mean absolute predictive errors, with comparisons made as in the synthetic response data study above. For the ‘incorrect’ prior, the prior source parameter not selected by empirical Bayes was used, along with the optimal strength parameter λ for that prior. Mirroring the synthetic data results (Simulations 1 and 2), we observe that prior elicitation with empirical Bayes provides an increase in mean predictive accuracy over a flat prior, an ‘incorrect’ prior, the Markov random field prior and the approach proposed by Li and Li, whilst Lasso regression has lowest mean predictive error. We note, however, that due to the very small sample size, differences in mean predictive error between these regimes are not conclusive. Yet, they all show an improvement over the baseline linear approach, and model averaging results in an average 26% decrease in predictive error over using MAP models. The prior strength parameter for the Markov random field prior was set to λ = −5 by empirical Bayes in every cross-validation iteration (this agrees with the selection of a pathway-based prior promoting large distances).
Illustrative computation times
Linear model without interaction terms
Linear model with interaction terms
d max = 2
d max = 3
d max = 4
d max = 5
d max = 2
d max = 3
d max = 4
d max = 5
Model priors incorporating biological information can play an important role in variable selection, especially at the small sample sizes characteristic of molecular studies. In applications where there are multiple sources of prior information, or multiple possible prior specifications, the empirical Bayes approach we put forward permits objective selection and weighting. This aids prior elicitation and guards against the use of mis-specified priors. We demonstrated that a biologically informative prior, with hyperparameters set by empirical Bayes, can have benefits over both a flat prior and a subjectively formed prior which is incorrect with respect to the underlying system. We also observed that, whilst Lasso regression can offer some improvement in predictive performance over the Bayesian approaches, its accuracy in selecting the correct underlying model (i.e. variable selection) can be inferior to the proposed empirical Bayes approach, thereby affecting interpretability of results. Empirical Bayes approaches have previously been used in variable selection, but with standard Bernoulli-distributed priors[33, 34].
We developed informative priors in the context of protein signalling based on two high-level features derived from network information: the number of pathways a subset of predictors incorporates and the intra-pathway distance between proteins in a proposed model. This formulation used the entire network structure in an intuitive way, removing the the need to specify individual prior probabilities for each variable and avoiding assumptions of prior independence between variables.
Our pathway-based priors form part of a growing literature on exploiting existing domain knowledge to aid inference, especially in the small sample setting. For example, recent variable selection studies also make use of graph structure within a Bayesian Markov random field prior[19, 35, 36] and within a non-Bayesian framework[21, 37, 38], essentially preferring models containing predictors that are neighbours in the graph. This is similar in spirit to the special case of our prior where the network consists of a single pathway and short intra-pathway distances are strongly preferred.
We compared our pathway-based priors to the Markov random field prior, but found in Simulation 2 that empirical Bayes frequently set the prior strength parameter to an incorrect value, resulting in a prior that penalises models containing predictors that are neighbours in the network, instead of promoting them. This is likely due to the parameterisation of the Markov random field prior, which is not agnostic to the number of included predictors in the model |γ|; addition of a predictor to a model could lead to a substantial increase in the prior score. Indeed, it has previously been noted that Markov random field priors can be unstable with the occurance of phase transitions in |γ|. Hence, the prior prefers less sparse models, but these models do not agree well with the data, as more complex models are penalised by the marginal likelihood. In contrast, our distance prior is based on an average distance measure and so is somewhat indifferent to |γ|. In Simulation 3, we do not observe this behaviour of the Markov random field prior; since the linear model does not include interaction terms, the model complexity does not increase as sharply with |γ| and so there is less disagreement between the prior and the marginal likelihood. We note that biologically informative priors have also been used for classification[22, 39, 40] and network inference[41–43].
We also compared our approach to the network-based penalised-likelihood method proposed by Li and Li. It performed poorly in Simulations 1 and 2, primarily due to its inability to capture nonlinear interplay. However, even in Simulation 3, with no interaction terms in the data-generating model, it failed to match the performance of our proposed empirical Bayes approach with pathway-based priors. This could be due to lack of similarity between the regression coefficients in the data-generating model, which goes against an assumption of the penalised-likelihood approach; that coefficients are similar for predictors that are neighbours in the network. This could also explain its poor predictive performance.
We used a continuous regression framework with interaction terms. Whilst discrete models are naturally capable of capturing non-linear interplay between components, the discretisation process results in a loss of information. Continuous models avoid this loss, but the response is usually assumed to depend linearly on predictors. The product terms in our model provide the possibility of capturing influences on the response of interest by interplay between predictors, including higher-order interactions. Chipman and Jensen et al. have employed a related approach allowing pairwise interactions only. We note that, under our formulation, model complexity grows rapidly with number of included predictors. However, complex models are naturally penalised by the marginal likelihood formulation giving overall sparse, parsimonious models, yet allowing for complex interplay via product terms.
We carried out variable selection using exact model averaging. This was made possible by means of a sparsity restriction. Sparsity constraints have been employed in previous work in Bayesian variable selection[4, 45] and also in the related setting of inference of gene regulatory networks[42, 46]. The sparsity-constrained approach proposed is attractive as it yields exact posterior probabilities and facilitates exact empirical Bayes analysis. Sparsity is a reasonable assumption in settings where it is likely that only a few predictors play a key role in influencing a response. In such settings, and where data is of small-to-moderate dimensionality, our exact approach is fast and deterministic with no requirement of MCMC convergence diagnostics. This, together with empirical Bayes and the choice of parameter priors, results in the overall approach having very few user-set parameters.
In applications of higher dimensionality, where the exact calculation is no longer feasible, empirical Bayes can still be performed using an approximate conditional marginal ‘likelihood’ approach as seen in George and Foster and Yuan and Lin. This involves optimisation over the model space instead of averaging. MCMC, with the selected hyperparameter values, can then be used to estimate inclusion probabilities. Alternatively, a fully Bayes MCMC approach could be taken, which places a prior on the hyperparameters and integrates them out (see e.g.).
Illustrative computational times for our approach are shown in Table3, for four values of p (number of predictors) and four values of d max (maximum number of predictors allowed in a model). We also considered linear models with and without interaction terms. Empirical Bayes was used to select between two priors (M = 2) and to set the prior strength parameter (optimisation performed over ten values of λ). The computation time scales as for the model without interaction terms and for the model with interaction terms. We see that the approach is fast on datasets of moderate dimensionality (∼100 variables) with d max = 3. We note that shortage of memory was the limiting factor on our machine. Computational time could also be improved by using multiple cores to calculate empirical Bayes marginal likelihood scores for multiple values of λ simultaneously.
We showed examples of automated selection between multiple sources of ancillary information, but, rather than selecting a single source, the methods proposed could be generalised to allow combinations of complementary information sources as seen in Jensen et al.. Whilst our priors were based on pathway and network structure, the methods can also permit integration and weighting of publicly available data, which while plentiful, can be of uncertain relevance to a given study.
In this paper we have proposed an empirical Bayes method for objective selection and weighting of biologically informative prior information for integration within Bayesian variable selection. The method is computationally efficient, exact and has very few user-set parameters. We developed informative pathway-based priors in the context of protein signalling and illustrated our method on synthetic repsonse data. We demonstrated that in situations where there are several plausible formulations for the prior, it is capable of selecting the most appropriate. In particular, the approach has potential to significantly improve results by guarding against mis-specification of priors. Comparisons were made to alternative methods, demonstrating that the proposed approach offers a competitive variable selection performance. We have also shown an application on cancer drug response data and obtained biologically plausible results. Our method is general and can be applied in any setting with multiple sources of prior knowledge.
SMH designed the priors, carried out all computational analyses and wrote the paper. SM conceived the study, provided feedback on all aspects and revised the manuscript. RMN and NB carried out the cell line assays. WLK and SZ carried out the drug response assays. PTS and JWG led the biological aspects of the work and provided feedback. All authors read and approved the final manuscript.
This work was supported by the Director, Office of Science, Office of Basic Energy Sciences, of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231, by the National Institutes of Health, National Cancer Institute grants U54 CA 112970 and P50 CA 58207 to JWG, and the Cancer Systems Biology Center grant from the Netherlands Organisation for Scientific Research. SMH and SM were supported under EPSRC EP/E501311/1.
- Lee KE, Sha N, Dougherty ER, Vannucci M, Mallick BK: Gene selection: a Bayesian variable selection approach. Bioinformatics 2003, 19: 90–97. 10.1093/bioinformatics/19.1.90View ArticlePubMed
- Rogers S, Girolami M: A Bayesian regression approach to the inference of regulatory networks from gene expression data. Bioinformatics 2005, 21(14):3131–3137. 10.1093/bioinformatics/bti487View ArticlePubMed
- Jensen ST, Chen G, Stoeckert CJ: Bayesian variable selection and data integration for biological regulatory networks. Ann Appl Stat 2007, 1: 612–633. 10.1214/07-AOAS130View Article
- Mukherjee S, et al.: Sparse combinatorial inference with an application in cancer biology. Bioinformatics 2009, 25: 265–271. 10.1093/bioinformatics/btn611PubMed CentralView ArticlePubMed
- Wu TT, Chen YF, Hastie T, Sobel E, Lange K: Genome-wide association analysis by lasso penalized logistic regression. Bioinformatics 2009, 25: 714–721. 10.1093/bioinformatics/btp041PubMed CentralView ArticlePubMed
- Ai-Jun Y, Xin-Yuan S: Bayesian variable selection for disease classification using gene expression data. Bioinformatics 2010, 26: 215–222. 10.1093/bioinformatics/btp638View Article
- Brown PJ, Vannucci M, Fearn T: Bayes model averaging with selection of regressors. J R Stat Soc B 2002, 64: 519–536. 10.1111/1467-9868.00348View Article
- George EI, McCulloch RE: Approaches for Bayesian variable selection. Stat Sin 1997, 7: 339–373.
- Chipman H, et al.: The practical implementation of Bayesian model selection. IMS Lecture Notes-Monograph Ser 2001, 38: 65–134.View Article
- Sachs K, Perez O, Pe’er D, Lauffenburger DA, Nolan GP: Causal protein-signaling networks derived from multiparameter single-cell data. Science 2005, 308: 523–529. 10.1126/science.1105809View ArticlePubMed
- Madigan D, York J, Allard D: Bayesian graphical models for discrete data. Int Stat Rev 1995, 63: 215–232. 10.2307/1403615View Article
- Raftery AE, Madigan D, Hoeting JA: Bayesian model averaging for linear regression models. J Am Stat Assoc 1997, 92: 179–191. 10.1080/01621459.1997.10473615View Article
- Robert CP, Casella G: Monte Carlo Statistical Methods. New York: Springer; 2004.View Article
- Nott DJ, Green PJ: Bayesian variable selection and the Swendsen-Wang algorithm. J Comput Graph Stat 2004, 13: 141–157. 10.1198/1061860042958View Article
- Smith M, Kohn R: Nonparametric regression using Bayesian variable selection. J Econometrics 1996, 75: 317–343. 10.1016/0304-4076(95)01763-1View Article
- Denison DGT, Holmes CC, Mallick BK, Smith AFM: Bayesian Methods for Nonlinear Classification and Regression. London: Wiley; 2002.
- Kohn R, Smith M, Chan D: Nonparametric regression using linear combinations of basis functions. Stat Comput 2001, 11: 313–322. 10.1023/A:1011916902934View Article
- Carlin BP, Louis TA: Bayesian Methods for Data Analysis. Chapman & Hall; 2008.
- Li F, Zhang NR: Bayesian variable selection in structured high-dimensional covariate spaces with applications in genomics. J Am Stat Assoc 2010, 105: 1202–1214. 10.1198/jasa.2010.tm08177View Article
- Tibshirani R: Regression shrinkage and selection via the lasso. J R Stat Soc B 1996, 58: 267–288.
- Li C, Li H: Network-constrained regularization and variable selection for analysis of genomic data. Bioinformatics 2008, 24: 1175–1182. 10.1093/bioinformatics/btn081View ArticlePubMed
- Stingo FC, Vannucci M: Variable selection for discriminant analysis with Markov random field priors for the analysis of microarray data. Bioinformatics 2011, 27: 495–501. 10.1093/bioinformatics/btq690PubMed CentralView ArticlePubMed
- Friedman J, Hastie T, Tibshirani R: Regularization paths for generalized linear models via coordinate descent. J Stat Soft 2010, 33: 1–22.View Article
- Weinberg RA: The Biology of Cancer. New York: Garland Science; 2006.
- Hanahan D, Weinberg RA: The hallmarks of cancer. Cell 2000, 100: 57–70. 10.1016/S0092-8674(00)81683-9View ArticlePubMed
- Yarden Y, Sliwkowski MX: Untangling the ErbB signalling network. Nat Rev Mol Cell Biol 2001, 2: 127–137. 10.1038/35052073View ArticlePubMed
- Neve RM, et al.: A collection of breast cancer cell lines for the study of functionally distinct cancer subtypes. Cancer Cell 2006, 10: 515–527. 10.1016/j.ccr.2006.10.008PubMed CentralView ArticlePubMed
- Heiser LM, et al.: Subtype and pathway specific responses to anticancer compounds in breast cancer. Proceedings of the National Academy of Sciences 2011,. in press in press
- Yang L, et al.: Akt/protein kinase B signaling inhibitor-2, a selective small molecule inhibitor of Akt signaling with antitumor activity in cancer cells overexpressing Akt. Cancer Res 2004, 64: 4394–4399. 10.1158/0008-5472.CAN-04-0343View ArticlePubMed
- Burgering BMT, Coffer PJ: Protein kinase B (c-Akt) in phosphatidylinositol-3-OH kinase signal transduction. Nature 2002, 376: 599–602.View Article
- Nevins JR: The Rb/E2F pathway and cancer. Hum Mol Genet 2001, 10: 699–703. 10.1093/hmg/10.7.699View ArticlePubMed
- Chaussepied M, Ginsberg D: Transcriptional Regulation of AKT Activation by E2F. Mol Cell 2004, 16: 831–837. 10.1016/j.molcel.2004.11.003View ArticlePubMed
- George EI, Foster DP: Calibration and empirical Bayes variable selection. Biometrika 2000, 87: 731–747. 10.1093/biomet/87.4.731View Article
- Yuan M, Lin Y: Efficient empirical Bayes variable selection and estimation in linear models. J Am Stat Assoc 2005, 100: 1215–1225. 10.1198/016214505000000367View Article
- Wei Z, Li H: A hidden spatial-temporal Markov random field model for network-based analysis of time course gene expression data. Ann Appl Stat 2008, 2: 408–429.View Article
- Li H, Monni S: Bayesian methods for network-structured genomics data. In Frontiers of Statistical Decision Making and Bayesian Analysis. Edited by: Chen MH, Müller P, Sun D, Ye K, Dey DK. New, York: Springer; 2010:303–315.
- Binder H, Schumacher M: Incorporating pathway information into boosting estimation of high-dimensional risk prediction models. BMC Bioinf 2009, 10: 18. 10.1186/1471-2105-10-18View Article
- Slawski M, zu Castell W, Tutz G: Feature selection guided by structural information. Ann Appl Stat 2010, 4: 1056–1080. 10.1214/09-AOAS302View Article
- Zhu Y, Shen X, Pan W: Network-based support vector machine for classification of microarray samples. BMC Bioinf 2009, 10: S21.View Article
- Guillemot V, Tenenhaus A, Le Brusquet L, Frouin V: Graph constrained discriminant analysis: A new method for the integration of a graph into a classification process. PLoS ONE 2011, 6: e26146. 10.1371/journal.pone.0026146PubMed CentralView ArticlePubMed
- Hartemink AJ, Bernard A: Informative structure priors: Joint learning of dynamic regulatory networks from multiple types of data. In Pac Symp Biocomput 2005. Singapore: World Scientific; 2005:459–470.
- Werhli AV, Husmeier D: Reconstructing gene regulatory networks with Bayesian networks by combining expression data with multiple sources of prior knowledge. Stat Appl Genet Mol Biol 2007, 6: 15.
- Mukherjee S, Speed TP: Network inference using informative priors. Proc Natl Acad Sci USA 2008, 105: 14313–14318. 10.1073/pnas.0802272105PubMed CentralView ArticlePubMed
- Chipman H: Bayesian variable selection with related predictors. Can J Stat 1996, 24: 17–36. 10.2307/3315687View Article
- Jiang W: Bayesian variable selection for high dimensional generalized linear models: convergence rates of the fitted densities. Ann Stat 2007, 35: 1487–1511. 10.1214/009053607000000019View Article
- Husmeier D: Sensitivity and specificity of inferring genetic regulatory interactions from microarray experiments with dynamic Bayesian networks. Bioinformatics 2003, 19: 2271–2282. 10.1093/bioinformatics/btg313View ArticlePubMed
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.