1 Introduction

Bayesian ANalysis of Differential Localisation Experiments (BANDLE) is an integrative semi-supervised functional mixture model, developed by Crook et al (Crook et al. 2022), to obtain the probability of a protein being differentially localised between two conditions.

In this vignette we walk users through how to install and use the R (R Development Core Team 2011) Bioconductor (Gentleman et al. 2004) bandlepackage by simulating a well-defined differential localisation experiment from spatial proteomics data from the pRolocdata package (Gatto et al. 2014).

The BANDLE method uses posterior Bayesian computations performed using Markov-chain Monte-Carlo (MCMC) and thus uncertainty estimates are available (Gilks, Richardson, and Spiegelhalter 1995). It is inspired by the T-augmented Gaussuan mixture model (TAGM) by Crook et al. 2018 (Crook et al. 2018) which was developed to allow interrogation of multiply localised proteins through uncertainty estimation. Throughout this vignette we use the term differentially localised to pertain to proteins which are assigned to different sub-cellular localisations between two conditions.

The output of BANDLE provides users with:

  1. Protein subcellular localisation predictions. For each protein the full probability distribution over all possible organelles/complexes is computed. Through uncertainty quantification users can gain insight into proteins that may localise to more than one location.

  2. Differential localisation predictions. The full probability of a protein being differentially localised between two conditions is computed. Uncertainty estimates are available for the differential localisation probability to aid filtering of candidate movers for validation.

2 Installation

The package can be installed with the BiocManager package:

if (!requireNamespace("BiocManager", quietly=TRUE))
    install.packages("BiocManager")
BiocManager::install("bandle")

and then loaded,

library("bandle")

For visualisation we also load the packages,

library("pheatmap")
library("viridis")
library("dplyr")
library("ggplot2")
library("gridExtra")

3 The data

In this vignette and Crook et al. (2022), the main data source that we use to study differential protein sub-cellular localisation are data from high-throughput mass spectrometry-based experiments. The data from these types of experiments traditionally yield a matrix of measurements wherein we have, for example, PSMs, peptides or proteins along the rows, and samples/channels/fractions along the columns. The bandle package uses the MSnSet class as implemented in the Bioconductor MSnbase package and thus requires users to import and store their data as a MSnSet instance. For more details on how to create a MSnSet see the relevant vignettes in pRoloc. The pRolocdata experiment data package is a good starting place to look for test data. This data package contains tens of quantitative proteomics experiments, stored as MSnSets.

3.1 A well-defined theoretical example

To get started with the basics of using bandle we begin by generating a simple example dataset which simulates a differential localisation experiment (please see the second vignette in this package for a full real-life biological use case). In this example data, the key elements are replicates, and a perturbation of interest. There is code within the bandle package to simulate an example experiment.

In the code chunk below we begin by loading the pRolocdata package to obtain a spatial proteomics dataset. This will be the basis of our simulation which will use boostrapping to generate new datasets. The dataset we have chosen to load is a dataset from 2009 (tan2009r1). This is data from a early LOPIT experiment performed on Drosophila embryos by Tan et al. (2009). The aim of this experiment was to apply LOPIT to an organism with heterogeneous cell types. This experiment used four isotopes across four distinct fractions and thus yielded four measurements (features) per protein profile. We visualise the data by using principal components analysis.

library("pRolocdata")
data("tan2009r1")

## Let's set the stock colours of the classes to plot to be transparent
setStockcol(NULL)
setStockcol(paste0(getStockcol(), "90")) 

## Plot the data using plot2D from pRoloc
plot2D(tan2009r1,
       main = "An example spatial proteomics datasets", 
       grid = FALSE)
addLegend(tan2009r1, where = "topleft", cex = 0.7, ncol = 2)

The following code chuck simulates a differential localisation experiment. It will generate numRep/2 of each a control and treatment condition. We will also simulate relocalisations for numDyn proteins.

set.seed(1)
tansim <- sim_dynamic(object = tan2009r1,
                      numRep = 6L,
                      numDyn = 100L)
## [1] "markers"

The list of the 6 simulated experiments are found in tansim$lopitrep. Each one is an MSnSet instance (the standard data container for proteomics experimental data). The first 3 are the simulated control experiments (see tansim$lopitrep[1:3]), and the following 3 in the list are the treatment condition simulated experiments (see tansim$lopitrep[4:6]).

# To access the first replicate
tansim$lopitrep[[1]]
## MSnSet (storageMode: lockedEnvironment)
## assayData: 888 features, 4 samples 
##   element names: exprs 
## protocolData: none
## phenoData
##   sampleNames: X114 X115 X116 X117
##   varLabels: Fractions
##   varMetadata: labelDescription
## featureData
##   featureNames: P20353 P53501 ... P07909 (888 total)
##   fvarLabels: FBgn Protein.ID ... knn.scores (18 total)
##   fvarMetadata: labelDescription
## experimentData: use 'experimentData(object)'
##   pubMedIds: 19317464 
## Annotation:  
## - - - Processing information - - -
## Added markers from  'mrk' marker vector. Thu Jul 16 22:53:44 2015 
## Performed knn prediction (k=10) Tue Nov  5 16:11:29 2024 
##  MSnbase version: 1.17.12

We can plot them using the plot2D function from pRoloc.

plot_title <- c(paste0("Replicate ", seq(3), " condition", " A"), 
               paste0("Replicate ", seq(3), " condition", " B"))

par(mfrow = c(2, 3))
out <- lapply(seq(tansim$lopitrep), function(z) 
    plot2D(tansim$lopitrep[[z]], grid = FALSE, main = plot_title[z]))

For understanding, exploring and visualizing individual spatial proteomics experiments, see the vignettes in pRoloc and MSnbase packages.

4 Preparing for bandle analysis

The main function of the package is bandle, this uses a complex model to analyse the data. Markov-Chain Monte-Carlo (MCMC) is used to sample the posterior distribution of parameters and latent variables. From which statistics of interest can be computed. Here we only run a few iterations for brevity but typically one needs to run thousands of iterations to ensure convergence, as well as multiple parallel chains.

4.1 Fitting Gaussian processes

First, we need to fit non-parametric regression functions to the markers profiles, upon which we place our analysis. This uses Gaussian processes. The fitGPmaternPC function can be used and fits some default penalised complexity priors (see ?fitGP), which works well. However, these can be altered, which is demonstrated in the next code chunk and also further in Vignette 2 of this package.

par(mfrow = c(4, 3))
gpParams <- lapply(tansim$lopitrep, function(x) 
  fitGPmaternPC(x, hyppar = matrix(c(10, 60, 250), nrow = 1)))

We apply the fitGPmaternPC function to each datasets by calling lapply over the tansim$lopitrep list of datasets. The output of fitGPmaternPC returns a list of posterior predictive means and standard deviations. As well as MAP hyperparamters for the GP.

Note here we the use the default parameters for the fitGPmaternPC function are hyppar = matrix(c(10, 60, 250), nrow = 1). This is provided as a starting point for fitting the GPs to the marker profile distributions. In the Crook et al. (2022) we found that these values worked well for smaller spatial proteomics datasets and those with fewer channels/fractions. This was visually assessed by passing these values and visualising the GP fit using the plotGPmatern function.

The plotGPmatern function can be used to plot the profiles for each class in each replicate condition with the posterior predictive distributions overlayed with the markers protein profiles.

For example, to plot the predictive distributions of the first dataset,

par(mfrow = c(4, 3))
plotGPmatern(tansim$lopitrep[[1]], params = gpParams[[1]])

In the next code chunk we set some priors on the data. The prior needs to form a K*3 matrix. K corresponds to the number of subcellular classes in the data, and 3 columns for the prior; (1) length-scale, (2) amplitude and (3) variance (see hyppar in ?fitGPmaternPC). Increasing these values, increases the shrinkage. For more details see the manuscript by Crook et al. (2022). We strongly recommend users start with the recommended hyppar parameters and change and assess them as necessary for their dataset by visually evaluating the fit of the GPs using the plotGPmatern function.

K <- length(getMarkerClasses(tansim$lopitrep[[1]], fcol = "markers"))
pc_prior <- matrix(NA, ncol = 3, K)
pc_prior[seq.int(1:K), ] <- matrix(rep(c(10, 60, 250),
                                       each = K), ncol = 3)

Now we have generated these complexity priors we can pass them as an argument to the fitGPmaternPC function. For example,

gpParams <- lapply(tansim$lopitrep,
                   function(x) fitGPmaternPC(x, hyppar = pc_prior))

By looking at the plot of posterior predictives using the gpParams we can see the GP fit looks sensible.

4.2 Setting the prior on the weights

The next step is to set up the matrix Dirichlet prior on the mixing weights. These weights are defined across datasets so these are slightly different to mixture weights in usual mixture models. The \((i,j)^{th}\) entry is the prior probability that a protein localises to organelle \(i\) in the control and \(j\) in the treatment. This mean that off-diagonal terms have a different interpretation to diagonal terms. Since we expect re-localisation to be rare, off-diagonal terms should be small. The following functions help set up the priors and how to interpret them. The parameter q allow us to check the prior probability that more than q differential localisations are expected.

set.seed(1)
dirPrior = diag(rep(1, K)) + matrix(0.001, nrow = K, ncol = K)
predDirPrior <- prior_pred_dir(object = tansim$lopitrep[[1]],
                               dirPrior = dirPrior,
                               q = 15)

The mean number of re-localisations is small:

predDirPrior$meannotAlloc
## [1] 0.3457542

The prior probability that more than q differential localisations are expected is small

predDirPrior$tailnotAlloc
## [1] 4e-04

The full prior predictive can be visualised as histogram. The prior probability that proteins are allocated to different components between datasets concentrates around 0.

hist(predDirPrior$priornotAlloc, col = getStockcol()[1])

For most use-cases we indeed expect the number of differential localisations to be small. However, there may be specific cases where one may expect the either a smaller or larger number of differential localisations. Users could try testing different values for the dirPrior for example, replacing 0.001 with 0.0005 or smaller, for larger datasets to bring the number of expected re-localisations inline with the biological expectation, and visa-versa when we expect the number of proteins to have changed to be higher.

5 Running the bandle function

We are now ready to run the main bandle function. Remember to carefully select the datasets and replicates that define the control and treatment. As a reminder, in this introductory vignette we have used a small dataset and generated theoretical triplicates of each theoretical condition. Please see the second vignette in this package for a more detailed workflow and real biological use-case. In the below code chunk we run bandle for only 100 iterations for the convenience of building the vignette, but typically we’d recommend you run the number of iterations (numIter) in the 1000s.

Remember: the first 3 datasets are the first 3 elements of tansim and the final 3 elements are the “treatment” triplicate datasets.

## Split the datasets into two separate lists, one for control and one for treatment
control <- tansim$lopitrep[1:3] 
treatment <- tansim$lopitrep[4:6]

## Run bandle
bandleres <- bandle(objectCond1 = control,
                    objectCond2 = treatment,
                    numIter = 100,   # usually 10,000
                    burnin = 5L,    # usually 5,000
                    thin = 1L,      # usually 20
                    gpParams = gpParams,
                    pcPrior = pc_prior,
                    numChains = 3,  # usually >=4
                    dirPrior = dirPrior, 
                    seed = 1)       # set a random seed for reproducibility)

The bandle function generates an object of class bandleParams. The show method indicates the number of parallel chains that were run, this should typically be greater than 4 (here we use 3 just as a demo).

bandleres
## Object of class "bandleParams"
## Method: bandle 
## Number of chains: 3

6 Analysing bandle output

6.1 Assesing the model for convergence

The bandle method uses of Markov Chain Monte Carlo (MCMC) and therefore before we can extract our classification and differential localisation results we first need to check the algorithm for convergence of the MCMC chains. This is standard practice in Bayesian inference and there are many ways to assess for convergence as discussed in Crook et al. (2019).

The two main functions we can use to help us assess convergence are:

  • calculateGelman which calculates the Gelman diagnostics for all pairwise chain combinations
  • plotOutliers which generates trace and density plots for all chains

Let’s start with the calculateGelman which uses code the coda R package. It allows us to compare the inter and intra chain variances. If the chains have converged the ratio of these quantities should be < 1.2, ideally, close to 1, as discuss in Crook et al. (2019) and general Bayesian literature.

calculateGelman(bandleres)
## $Condition1
##            comb_12  comb_13  comb_23
## Point_Est 1.197407 1.135958 1.092632
## Upper_CI  1.616689 1.494546 1.092794
## 
## $Condition2
##            comb_12  comb_13   comb_23
## Point_Est 1.064983 1.052406 0.9960637
## Upper_CI  1.064988 1.053878 0.9967413

In this example, to demonstrate how to use bandle we have only run 100 MCMC iterations for each chain. As already mentioned, in practice we suggest running a minimum of 1000 iterations and a minimum of 4 chains.

In this toy example we see that the point estimate Gelman diagnostics are <1.2 which suggests convergence. The upper confidence intervals however are higher than 1.2 for some pairs of chains. We may wish to consider removing some chains which when paired with another chain exhibit values > 1. First let’s generate trace and density plots for all chains using the plotOutliers function.

plotOutliers(bandleres)

  • Trace plots are subjective but can used to help visually assess the sample path of the chains. Textbooks on Bayesian inference often tell us that a good trace plots should look like a “hairy” or “fuzzy caterpillar”.
  • The density plots generated from the plotOutliers function show the number of outliers for each iteration expressed as a probability density. For convergence we expect to see a normally distributed plot centered around roughly the same number of outliers in each chain.

As expected, as this is a toy theoretical example, only run for a few chains and iterations, the trace and density plots do not show convergence i.e. over the 100 iterations they do not reach a stationary distribution.

In practice, we would do is remove bad chains i.e. those that have not converged.

6.2 Removing unconverged chains

Chains which have not converged can be removed by standard subsetting. As an example, let’s remove chain 2. We generate a new object called bandlres_opt,

bandleres_opt <- bandleres[-2]

We see the object bandlres_opt has now only 2 chains,

bandleres_opt
## Object of class "bandleParams"
## Method: bandle 
## Number of chains: 2

In practice if we remove chains we would re-compute the Gelman and plot the outliers as we did above to check again for convergence.

6.3 Populating a bandleres object

Currently, the summary slots of the bandleres object are empty. The summaries function accesses them.

summaries(bandleres_opt)
## [[1]]
## An object of class "bandleSummary"
## Slot "posteriorEstimates":
## <S4 Type Object>
## attr(,"elementType")
## [1] "ANY"
## attr(,"elementMetadata")
## `\001NULL\001`
## attr(,"metadata")
## list()
## 
## Slot "diagnostics":
## <0 x 0 matrix>
## 
## Slot "bandle.joint":
## <0 x 0 matrix>
## 
## 
## [[2]]
## An object of class "bandleSummary"
## Slot "posteriorEstimates":
## <S4 Type Object>
## attr(,"elementType")
## [1] "ANY"
## attr(,"elementMetadata")
## `\001NULL\001`
## attr(,"metadata")
## list()
## 
## Slot "diagnostics":
## <0 x 0 matrix>
## 
## Slot "bandle.joint":
## <0 x 0 matrix>

These can be populated as follows,

bandleres_opt <- bandleProcess(bandleres_opt)

These slots have now been populated

summaries(bandleres_opt)

7 Predicting the subcellular location

We can append the results to our original MSnSet datasets using the bandlePredict function.

xx <- bandlePredict(control, 
                    treatment, 
                    params = bandleres_opt, 
                    fcol = "markers")
res_control <- xx[[1]]
res_treatment <- xx[[2]]

The output is a list of MSnSets. In this example, we have 3 for the control and 3 for the treatment.

length(res_control)
## [1] 3
length(res_treatment)
## [1] 3

The results are appended to the first MSnSet feature data slot for each condition.

fvarLabels(res_control[[1]])
##  [1] "FBgn"                             "Protein.ID"                      
##  [3] "Flybase.Symbol"                   "AccessionNo"                     
##  [5] "EntryName"                        "AccessionNoAll"                  
##  [7] "EntryNameAll"                     "No.peptide.IDs"                  
##  [9] "Mascot.score"                     "No.peptide.quantified"           
## [11] "PLSDA"                            "pd.2013"                         
## [13] "pd.markers"                       "markers.orig"                    
## [15] "markers"                          "markers.tl"                      
## [17] "knn"                              "knn.scores"                      
## [19] "bandle.allocation"                "bandle.probability"              
## [21] "bandle.probability.lowerquantile" "bandle.probability.upperquantile"
## [23] "bandle.mean.shannon"              "bandle.differential.localisation"
## [25] "bandle.outlier"                   "bandle.joint"

To access them use the fData function

fData(res_control[[1]])$bandle.probability
fData(res_control[[1]])$bandle.allocation

7.1 Thresholding on protein allocations

It is common practice in supervised machine learning to set a specific threshold on which to define new assignments/allocations, below which classifications are left unassigned/unknown. Indeed, we do not expect the whole subcellular diversity to be represented by the 11 niches defined here, we expect there to be many more, many of which will be multiply localised within the cell. It is important to allow for the possibility of proteins to reside in multiple locations, this information is available in the bandle.joint slot and also can be extracted from the bandleParams object.

As we are using a Bayesian model the outputs of the classifier are probabilities. This not only allows us to look at the distribution of probabilities over all subcellular classes but also allows us to extract a probability threshold on which we can define new assignments.

The subcellular allocations are located in the bandle.allocation column of the fData slot and the posteriors are located in the bandle.probability slot. We can use the getPredictions function from the pRoloc package to return a set of predicted localisations according to if they meet a probability threshold.

For example, in the below code chunk we set a 1% FDR (t = .99) for assigning proteins a subcellular niche, below which we leave them unlabelled termed “unknown” in the pRoloc literature.

res_control[[1]] <- getPredictions(res_control[[1]], 
                                   fcol = "bandle.allocation",                   
                                   scol = "bandle.probability",                   
                                   mcol = "markers",                   
                                   t = .99)
## ans
##  Cytoskeleton            ER         Golgi      Lysosome       Nucleus 
##            12           216           134             8            39 
##            PM    Peroxisome    Proteasome  Ribosome 40S  Ribosome 60S 
##           134             8            41            72            49 
## mitochondrion       unknown 
##            82            93
res_treatment[[1]] <- getPredictions(res_treatment[[1]], 
                                   fcol = "bandle.allocation",                   
                                   scol = "bandle.probability",                   
                                   mcol = "markers",                   
                                   t = .99)
## ans
##  Cytoskeleton            ER         Golgi      Lysosome       Nucleus 
##            17           190           123             9            41 
##            PM    Peroxisome    Proteasome  Ribosome 40S  Ribosome 60S 
##           132            14            44            78            47 
## mitochondrion       unknown 
##            85           108

We see a new column has been appended to the fData slot, called, bandle.allocation.pred which contains the subcellular predictions after thresholding.

We can use the function unknownMSnSet to subset the data so that we only examine proteins which were not included as markers in the analysis.

## Remove the markers from the MSnSet
res_control_unknowns <- unknownMSnSet(res_control[[1]], fcol = "markers")
res_treated_unknowns <- unknownMSnSet(res_treatment[[1]], fcol = "markers")

Now we can use the getMarkers function to summarise the new allocations by class.

getMarkers(res_control_unknowns, "bandle.allocation.pred")
## organelleMarkers
##  Cytoskeleton            ER         Golgi       Nucleus            PM 
##             5           188           121            18           100 
##    Peroxisome    Proteasome  Ribosome 40S  Ribosome 60S mitochondrion 
##             4            26            52            17            53 
##       unknown 
##            93
getMarkers(res_treated_unknowns, "bandle.allocation.pred")
## organelleMarkers
##  Cytoskeleton            ER         Golgi      Lysosome       Nucleus 
##            10           162           110             1            20 
##            PM    Peroxisome    Proteasome  Ribosome 40S  Ribosome 60S 
##            98            10            29            58            15 
## mitochondrion       unknown 
##            56           108

Note: We may also wish to take into account the probability of the protein being an outlier and thus use the results in the bandle.outlier column of the feature data. We could calculate the product of the posterior and the outlier (as they are both probabilities i.e. bandle.outlier x bandle.outlier) to obtain a localisation score that takes into account the outlier model. More details on this are found in the second vignette of this package.

7.2 Distribution on allocations

We can examine the distribution of allocations that:

  1. have been assigned to a single location with high confidence and,
  2. those which did not meet the threshold and thus have high uncertainty i.e. are labelled as “unknown”.

7.2.1 Proteins assigned to one main location

The column in the fData called bandle.allocation.pred contains the predicted subcellular location for proteins. Let’s plot this,

Extract the predictions,

res1 <- fData(res_control_unknowns)$bandle.allocation.pred
res2 <- fData(res_treated_unknowns)$bandle.allocation.pred

res1_tbl <- table(res1)
res2_tbl <- table(res2)

Construct a quick barplot,

par(mfrow = c(1, 2))
barplot(res1_tbl, las = 2, main = "Predicted location: control",
        ylab = "Number of proteins")
barplot(res2_tbl, las = 2, main = "Predicted location: treatment",
        ylab = "Number of proteins")

The barplot tells us for this example that after thresholding with a 1% FDR on the posterior probability bandle has allocated the majority of unlabelled proteins to the ER, followed by the Golgi. We also see many proteins have been be unassigned and given the class label “unknown”. As previously mentioned the class label “unknown” is a historic term from the pRoloc package to describe proteins that are left unassigned following thresholding and thus proteins which exhibit uncertainty in their allocations. In the section after this one we examine these proteins and what their distributions may mean.

The associated posterior estimates are located in the bandle.probabilitycolumn and we can construct a boxplot to examine these probabilities by class,

pe1 <- fData(res_control_unknowns)$bandle.probability
pe2 <- fData(res_treated_unknowns)$bandle.probability

par(mfrow = c(1, 2))
boxplot(pe1 ~ res1, las = 2, main = "Posterior: control",
        ylab = "Probability")
boxplot(pe2 ~ res2, las = 2, main = "Posterior: treatment",
        ylab = "Probability")

7.2.2 Proteins with uncertainty

We can use the unknownMSnSet function once again to extract proteins which did not get a main location when we performed thresholding i.e. those labelled “unknown”.

res1_unlabelled <- unknownMSnSet(res_control_unknowns, 
                                 fcol = "bandle.allocation.pred")
res2_unlabelled <- unknownMSnSet(res_treated_unknowns, 
                                 fcol = "bandle.allocation.pred")

We see we have 91 and 99 proteins for the control and treatment respectively, which do not get assigned one main location.

nrow(res1_unlabelled)
## [1] 93
nrow(res2_unlabelled)
## [1] 108

Let’s extract the names of these proteins,

fn1 <- featureNames(res1_unlabelled)
fn2 <- featureNames(res2_unlabelled)

Let’s plot the the first 9 proteins that did not meet the thresholding criteria. We can use the mcmc_plot_probs function to generate a violin plot of the localisation distribution. We need to recall the main bandleParams object we created bandleres_opt to extract the full distribution and visualise the uncertainty.

Let’s first look at these proteins in the control condition,

g <- vector("list", 9)
for (i in 1:9) g[[i]] <- mcmc_plot_probs(bandleres_opt, fn1[i], cond = 1)
do.call(grid.arrange, g)

Now the treated,

g <- vector("list", 9)
for (i in 1:9) g[[i]] <- mcmc_plot_probs(bandleres_opt, fn1[i], cond = 2)
do.call(grid.arrange, g)

If we wanted to visualise one specific protein of interest e.g. "Q24253"

grid.arrange(
  mcmc_plot_probs(bandleres_opt, fname = "Q24253", cond = 1) +
  ggtitle("Distribution of Q24253 in control"),
  mcmc_plot_probs(bandleres_opt, fname = "Q24253", cond = 2) +
    ggtitle("Distribution of Q24253 in treated")
)

We can also get a summary of the full probability distribution by looking at the joint estimates stored in the bandle.joint slot of the MSnSet.

For example, again we can extract the joint posteriors for a specific protein,

fData(res_control_unknowns)$bandle.joint["Q24253", ]
##  Cytoskeleton            ER         Golgi      Lysosome       Nucleus 
## 3.678575e-107  1.046798e-27  6.149034e-03  4.221267e-32  3.089950e-68 
##            PM    Peroxisome    Proteasome  Ribosome 40S  Ribosome 60S 
##  9.938510e-01  5.865184e-33  2.405721e-37  6.856680e-47  6.332074e-23 
## mitochondrion 
##  4.014260e-75

Or full a set of proteins we could visualise the joint posteriors on a heatmap.

bjoint_control <- fData(res_control_unknowns)$bandle.joint
pheatmap(bjoint_control, cluster_cols = FALSE, color = viridis(n = 25), 
         show_rownames = FALSE, main = "Joint posteriors in the control")

bjoint_treated <- fData(res_treated_unknowns)$bandle.joint
pheatmap(bjoint_treated, cluster_cols = FALSE, color = viridis(n = 25),
         show_rownames = FALSE, main = "Joint posteriors in the treated")

8 Differential localisation probability

As previously mentioned the term “differentially localised” is used to pertain to proteins which are assigned to different sub-cellular localisations between two conditions. For the majority of users this is the main output they are keen to extract using the BANDLE method. The differential localisation probability can be found in either (1) the bandle.differential.localisation column of the MSnSet that we generated following prediction, or (2) obtained directly from the bandleParams object using the diffLocalisationProb function. The latter is useful for users who are only interested in running bandle for obtaining differential localisation information and not in using bandle as a method for protein localisation prediction.

To obtain the differential localisation probability from a bandleParams object,

dl <- diffLocalisationProb(bandleres_opt)
head(dl)
##     P20353     P53501     Q7KU78     P04412     Q7KJ73     Q9VM65 
## 0.00000000 0.03157895 0.00000000 0.06315789 0.03157895 0.00000000

Or from the MSnSet,

dl <- fData(res_control_unknowns)$bandle.differential.localisation
names(dl) <- featureNames(res_control_unknowns)
head(dl)
##     P20353     P53501     Q7KU78     P04412     Q7KJ73     Q9VM65 
## 0.00000000 0.03157895 0.00000000 0.06315789 0.03157895 0.00000000

The differential localisation probability tells us which proteins are most likely to differentially localise. This can also be seen on a rank plot,

plot(dl[order(dl, decreasing = TRUE)],
     col = getStockcol()[3], pch = 19, ylab = "Probability",
     xlab = "Rank", main = "Differential localisation rank plot")

In-line with our expectations, the rank plot indicates that most proteins are not differentially localised.

We can for example, examine how many proteins get a differential probability greater than 0.99 to look for the most confident differentially localised candidates.

## Subset MSnSets for DL proteins > 0.99
ind <- which(dl > 0.99)
res_control_dl0.99 <- res_control_unknowns[ind, ]
res_treated_dl0.99 <- res_treated_unknowns[ind, ]

## Get DL results
dl0.99 <- fData(res_control_dl0.99)$bandle.differential.localisation
(names(dl0.99) <- featureNames(res_control_dl0.99))
##  [1] "P26308"   "Q9VP77"   "Q9VLJ6"   "Q9VU35"   "Q9VC06"   "Q9NJH0"  
##  [7] "Q8INP8"   "A8JNJ6"   "Q8SZ38"   "Q24276"   "Q9VKD3"   "Q9VI10"  
## [13] "P42207"   "Q9VE79"   "Q7KMM4"   "Q9VN21"   "P11584"   "Q9VUR0"  
## [19] "Q9VJ46"   "Q9VZS3"   "O46111"   "P29310"   "Q05783"   "Q8MLV1"  
## [25] "M9PBB3"   "Q9VXE5"   "Q9VYT4"   "Q86PC7"   "Q9V4E0"   "Q9V4T5"  
## [31] "P08879"   "O97066"   "Q9VN86"   "P18431"   "P08111"   "Q8SYR7"  
## [37] "Q95TQ6"   "B7Z0X1"   "Q9VFP1"   "NO_ID_10" "E1JHT6"   "M9NF21"  
## [43] "Q9VXI6"   "P22700"   "Q8IPU3"   "D5AEK7"   "Q9VRL2"   "M9PIC3"  
## [49] "Q9VF87"   "Q7K3E2"   "Q8SWX8"   "Q9W1H1"   "Q9W1K0"   "P20240"  
## [55] "Q9VPH7"   "O18335"   "Q24007"   "Q9V3R8"   "E1JHY0"   "Q9VVI2"  
## [61] "Q9VJC7"   "M9PFY2"   "M9MSL3"   "Q7KVX1"   "B7Z0E0"   "Q9VVL7"  
## [67] "Q9VVJ7"   "Q95SY0"   "B7Z0D3"   "P82295"   "Q9VLQ1"   "P32234"  
## [73] "Q9VS57"   "Q94901"   "Q9VFV9"   "Q9U9Q4"   "O96051"   "Q9VFQ9"  
## [79] "Q9VMD5"

We find there are 79 proteins above this threshold.

8.1 Visualising differential localisation

We can visualise the changes in localisation between conditions on an alluvial plot using the plotTranslocations function

## Create an list of the two MSnSets
dl_msnsets <- list(res_control_dl0.99, res_treated_dl0.99)

## Set colours for organelles and unknown
mrkCl <- getMarkerClasses(res_control[[1]], fcol = "markers")
dl_cols <- c(getStockcol()[seq(mrkCl)], "grey")
names(dl_cols) <- c(mrkCl, "unknown")

## Now plot
plotTranslocations(dl_msnsets, 
                   fcol = "bandle.allocation.pred", 
                   col = dl_cols)

Or alternatively, on a chord (circos) diagram

plotTranslocations(dl_msnsets, 
                   fcol = "bandle.allocation.pred", 
                   col = dl_cols,
                   type = "chord")

A table summarising the differential localisations can be computed with the plotTable function

plotTable(dl_msnsets, fcol = "bandle.allocation.pred")
## 79 features in common
## ------------------------------------------------
## If length(fcol) == 1 it is assumed that the
## same fcol is to be used for both datasets
## setting fcol = c(bandle.allocation.pred, bandle.allocation.pred)
## ----------------------------------------------
##        Condition1    Condition2 value
## 1              ER         Golgi     1
## 3              ER            PM     3
## 4              ER    Peroxisome     7
## 5              ER    Proteasome     4
## 6              ER  Ribosome 40S     4
## 7              ER  Ribosome 60S     1
## 9              ER       unknown     6
## 10             ER  Cytoskeleton     2
## 11             ER      Lysosome     1
## 12          Golgi            ER     1
## 13          Golgi       Nucleus     3
## 19          Golgi mitochondrion     1
## 20          Golgi       unknown     2
## 25        Nucleus            PM     1
## 26        Nucleus    Peroxisome     1
## 38             PM    Proteasome     2
## 39             PM  Ribosome 40S     4
## 42             PM       unknown     2
## 46     Peroxisome         Golgi     1
## 59     Proteasome            PM     2
## 61     Proteasome  Ribosome 40S     2
## 63     Proteasome mitochondrion     2
## 64     Proteasome       unknown     1
## 68   Ribosome 40S         Golgi     1
## 74   Ribosome 40S mitochondrion     1
## 75   Ribosome 40S       unknown     4
## 83   Ribosome 60S    Proteasome     1
## 84   Ribosome 60S  Ribosome 40S     1
## 85   Ribosome 60S mitochondrion     1
## 97  mitochondrion       unknown     2
## 98  mitochondrion  Cytoskeleton     1
## 100       unknown            ER     2
## 102       unknown       Nucleus     2
## 103       unknown            PM     1
## 105       unknown    Proteasome     3
## 106       unknown  Ribosome 40S     1
## 107       unknown  Ribosome 60S     1
## 108       unknown mitochondrion     1
## 109       unknown  Cytoskeleton     1

9 Additional analysis

One advantage of using Bayesian methods over classic machine learning is the ability to quantify the uncertainty in our estimates. This can be useful to help pare down and select the proteins that are predicted to differentially localise.

9.1 Estimating uncertainty in differential localisation

There are several ways we can go about performing uncertainty quantification on the differential localisation probability. Several functions are available in the bandle package, namely, the binomDiffLoc function which allows users to sample credible intervals from a binomial distribution, or the bootstrapdiffLocprob which uses a non-parametric bootstrap on the Monte-Carlo samples.

9.1.1 The bootstrapdiffLocprob function

We can examine the top n proteins (here we use an example of top = 100) and produce bootstrap estimates of the uncertainty (note here the uncertainty is likely to be underestimated as we did not produce many MCMC samples). These can be visualised as ranked boxplots.

set.seed(1)
boot_t <- bootstrapdiffLocprob(params = bandleres, top = 100,
                               Bootsample = 5000, decreasing = TRUE)

boxplot(t(boot_t), col = getStockcol()[5],
        las = 2, ylab = "Probability", ylim = c(0, 1),
        main = "Differential localisation \nprobability plot (top 100 proteins)")

9.1.2 The binomDiffLoc function

Instead of applying the bootstrapdiffLocprob we could use the binomDiffLoc function to obtain credible intervals from the binomial distribution.

bin_t <- binomialDiffLocProb(params = bandleres, top = 100,
                             nsample = 5000, decreasing = TRUE)

boxplot(t(bin_t), col = getStockcol()[5],
        las = 2, ylab = "Probability", ylim = c(0, 1),
        main = "Differential localisation \nprobability plot (top 100 proteins)")

9.1.3 Obtaining probability estimates

There are many ways we could obtain probability estimates from either of the above methods. We could, for example, take the mean of each protein estimate, or compute the cumulative error (there is not really a false discovery rate in Bayesian statistics) or we could threshold on the interval to reduce the number of differential localisations if you feel the model has been overconfident.

# more robust estimate of probabilities
dprobs <- rowMeans(bin_t)

# compute cumulative error, there is not really a false discovery rate in
# Bayesian statistics but you can look at the cumulative error rate
ce <- cumsum(1  - dprobs)

# you could threshold on the interval and this will reduce the number of
# differential localisations
qt <- apply(bin_t, 1, function(x) quantile(x, .025))

9.2 The expected false discovery rate

Instead of estimating the false discovery rate we can estimate the expected false discovery rate from the posterior probabilities at a particular threshold. This mean that for fixed threshold, we compute the expected proportion of false discoveries. Here is an example below. We can see that setting a probability threshold of 0.95 leads to an expected false discovery rate of less than \(0.5\%\)

EFDR(dl, threshold = 0.95)
## [1] 0.001871345

(We remind users that this data is a simulated and uses very few iterations and chains).

10 Description of bandle parameters

The bandle function has a significant number of parameters to allow flexible and bespoke analysis. Here, we describe these parameters in more detail to allow user to make decisions on the level of flexibility they wish to exploit.

  1. objectCond1. This is a list of MSnSets containing the first condition.

  2. objectCond2. This is a list of MSnSets containing the second condition.

    1. These object should have the same observations and features. These will be checked during bandle analysis.
  3. fcol indicates the feature column in the MSnSets that indicated the markers. Proteins that are not markers should be labels unknown. The default is markers.

  4. hyperLearn is the algorithm used to learn the hyperparameters of the Gaussian processes. For speed the default is an optimization algorithm called “LBFGS”, however is users want to perform uncertainty quantification on these parameters we can use Markov-chain Monte Carlo (MCMC) methods. This is implemented using the Metropolis-Hastings algorithm. Though this latter methodology provides more information, it is much more costly. The analysis is expected to take several days rather than hours.

  5. numIter is the number of MCMC iterations for the algorithm. We typically suggest around 10,000 iterations is plenty for convergence. Though some cases may take longer. If resources are constrained, we suggest 4,000 iterations as acceptable. A minimum number of iterations is around 1,000 though at this level we expect the posterior estimates to suffer considerably. If possible more parallel chains should be run in this case by changing numChains to, say, 9. The more chains and iterations the more computationally expensive the algorithm. The time taken for the algorithm scales roughly linearly in the number of iterations

  6. burnin is the number of samples that should be discarded from the beginning of the chain due to the bias induced by the starting point of the algorithm. We suggest sensible burnin values to be roughly \(10-50\%\) of the number of iterations

  7. thin reduces auto-correlation in the MCMC samples. The default is \(5\), which means every 5th sample is taken. If memory requirements are an issue, we suggest to increase the thinning amount. Though above \(20\), you will see a decrease in performance.

  8. u and v represent the prior hyperparameters of the proportion of outliers. This is modelled using a Beta(u,v) with u = 2 and v = 10 a default. This suggest that roughly \(\frac{u}{u = V} = 16%\) of proteins are believed to be outliers and that it is quite unlikely that more than \(50%\) of proteins are outliers. Users can examine the quantiles of the Beta(u,v) distribution if they wish to place a more bespoke prior. For example, increasing u will increase the number of a prior believed outliers.

  9. lambda is a ridge parameter used for numerical stability and is set to \(0.01\). If you experience the algorithm fails due to numerical issue then you can set this value larger. If you require values above \(1\) it is likely that there are other issues with the analysis. We suggest checking the method is appropriate for your problem and opening issue detailing the problems.

  10. gpParams results from fitting Gaussian proccess (Gaussian random fields). We refer the users to those functions. The default is NULL which will fit GPs internally but we recommend setting these outside the bandle function because it leads to more stable results.

  11. hyperIter if the hyperparameters of the GP are learnt using MH algorithm then this is the frequency at which these are updated relative to the bandle algorithm. By default this is unused, but if hyperLearn is set to MH then this proceed at every 20 iterations.

  12. hyperMean is the mean of the log normal prior used on the hyperparameters. Though by default this is not used unless PC is set to false

  13. hyperSd is the standard deviation of the log normal prior used on the hyperparameters. The default is c(1,1,1) for the 3 hyperparameters, increasing these values increases the uncertainty in the prior values of the hyperparameters.

  14. seed is the random-number seed.

  15. pg indicates whether or not to use the Polya-Gamma (PG) prior. The default is false and a Dirichlet prior is used instead. If set to true the pg is used. In which case a default PG prior is used. This prior attempts to match the default Dirichlet prior that is used when PG prior is set to false. The PG prior is more computationally expensive but can provide prior information on correlations

  16. pgPrior is by default NULL. We suggest using the pg_prior function to help set this parameter and the documentation therein. This function uses an empirical approach to compute a sensible default.

  17. tau is a parameter used by the Polya-Gamma prior and we refer to BANDLE manuscript for details. By default it is only used if pg prior is true, when the default becomes 0.2. At this value the pg prior is similar to the Dirichlet prior but with information on correlations.

  18. dirPrior is the Dirichlet matrix prior on the correlations. This should be provided as a K by K matrix, where K is the number of subcellular niches. The diagonal component should represent the prior belief that organelles do not re-localise (same compartment), where as the off-diagonal terms represent the prior terms of re-localisation. The prior_pred_dir can be used to provide a prior predictive check based on the provided prior. It is recommended that the off-diagonal terms are at least two orders of magnitude smaller than the diagonal terms. An example is given in the vignette.

  19. maternCov is this true the covariance function is the matern covariance, otherwise a Gaussian covariance is used.

  20. PC indicates whether a penalised complexity (PC) is used. The default is true and otherwise log normal priors are used.

  21. pcPrior is a numeric of length 3 indicating the parameters of the PC prior. The prior is placed on the parameters of length-scale, amplitude, and variance in that order. The default values are \(0.5,3,100\), and increasing the value increases the shrinkage towards straight-lines with zero variance.

  22. nu which defaults to 2 is the smoothness of the matern covariance. By increasing nu you encourage smoother solutions. nu should be an integer, though for values of nu above 3, we have observed numerical instability.

  23. propSd is the standard deviation of the random-walk update used in the MH algorithm. We do not recommend changing this unless you are familiar with Bayesian analysis. The default is c(0.3,0.1,0.05) for the 3 hyperparameters. Changing these will alter the efficiency of the underlying samplers.

  24. numChains is the number of parrallel chains and defaults to 4. We recommend using as much processing resources as you have and frequently have used 9 in practise.

  25. BPPARAM is the BiocParallel back-end which defaults to BiocParallel::bpparam(). We refer you to the BiocParallel package for details on setting this dependent on your computing system.

11 Session information

All software and respective versions used to produce this document are listed below.

sessionInfo()
## R Under development (unstable) (2024-10-21 r87258)
## Platform: x86_64-pc-linux-gnu
## Running under: Ubuntu 24.04.1 LTS
## 
## Matrix products: default
## BLAS:   /home/biocbuild/bbs-3.21-bioc/R/lib/libRblas.so 
## LAPACK: /usr/lib/x86_64-linux-gnu/lapack/liblapack.so.3.12.0
## 
## locale:
##  [1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C              
##  [3] LC_TIME=en_GB              LC_COLLATE=C              
##  [5] LC_MONETARY=en_US.UTF-8    LC_MESSAGES=en_US.UTF-8   
##  [7] LC_PAPER=en_US.UTF-8       LC_NAME=C                 
##  [9] LC_ADDRESS=C               LC_TELEPHONE=C            
## [11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C       
## 
## time zone: America/New_York
## tzcode source: system (glibc)
## 
## attached base packages:
## [1] stats4    stats     graphics  grDevices utils     datasets  methods  
## [8] base     
## 
## other attached packages:
##  [1] pRolocdata_1.45.0    gridExtra_2.3        ggplot2_3.5.1       
##  [4] dplyr_1.1.4          viridis_0.6.5        viridisLite_0.4.2   
##  [7] pheatmap_1.0.12      bandle_1.11.0        pRoloc_1.47.0       
## [10] BiocParallel_1.41.0  MLInterfaces_1.87.0  cluster_2.1.6       
## [13] annotate_1.85.0      XML_3.99-0.17        AnnotationDbi_1.69.0
## [16] IRanges_2.41.0       MSnbase_2.33.0       ProtGenerics_1.39.0 
## [19] mzR_2.41.0           Rcpp_1.0.13-1        Biobase_2.67.0      
## [22] S4Vectors_0.45.0     BiocGenerics_0.53.1  generics_0.1.3      
## [25] BiocStyle_2.35.0    
## 
## loaded via a namespace (and not attached):
##   [1] splines_4.5.0               filelock_1.0.3             
##   [3] tibble_3.2.1                hardhat_1.4.0              
##   [5] preprocessCore_1.69.0       pROC_1.18.5                
##   [7] rpart_4.1.23                lifecycle_1.0.4            
##   [9] httr2_1.0.6                 doParallel_1.0.17          
##  [11] globals_0.16.3              lattice_0.22-6             
##  [13] MASS_7.3-61                 MultiAssayExperiment_1.33.0
##  [15] dendextend_1.18.1           magrittr_2.0.3             
##  [17] limma_3.63.1                plotly_4.10.4              
##  [19] sass_0.4.9                  rmarkdown_2.29             
##  [21] jquerylib_0.1.4             yaml_2.3.10                
##  [23] MsCoreUtils_1.19.0          DBI_1.2.3                  
##  [25] RColorBrewer_1.1-3          lubridate_1.9.3            
##  [27] abind_1.4-8                 zlibbioc_1.53.0            
##  [29] GenomicRanges_1.59.0        purrr_1.0.2                
##  [31] mixtools_2.0.0              AnnotationFilter_1.31.0    
##  [33] nnet_7.3-19                 rappdirs_0.3.3             
##  [35] ipred_0.9-15                circlize_0.4.16            
##  [37] lava_1.8.0                  GenomeInfoDbData_1.2.13    
##  [39] ggrepel_0.9.6               listenv_0.9.1              
##  [41] gdata_3.0.1                 parallelly_1.38.0          
##  [43] ncdf4_1.23                  codetools_0.2-20           
##  [45] DelayedArray_0.33.1         xml2_1.3.6                 
##  [47] shape_1.4.6.1               tidyselect_1.2.1           
##  [49] farver_2.1.2                UCSC.utils_1.3.0           
##  [51] matrixStats_1.4.1           BiocFileCache_2.15.0       
##  [53] jsonlite_1.8.9              caret_6.0-94               
##  [55] e1071_1.7-16                ggalluvial_0.12.5          
##  [57] survival_3.7-0              iterators_1.0.14           
##  [59] foreach_1.5.2               segmented_2.1-3            
##  [61] tools_4.5.0                 progress_1.2.3             
##  [63] lbfgs_1.2.1.2               glue_1.8.0                 
##  [65] prodlim_2024.06.25          SparseArray_1.7.1          
##  [67] xfun_0.49                   MatrixGenerics_1.19.0      
##  [69] GenomeInfoDb_1.43.0         withr_3.0.2                
##  [71] BiocManager_1.30.25         fastmap_1.2.0              
##  [73] fansi_1.0.6                 digest_0.6.37              
##  [75] timechange_0.3.0            R6_2.5.1                   
##  [77] colorspace_2.1-1            gtools_3.9.5               
##  [79] lpSolve_5.6.21              biomaRt_2.63.0             
##  [81] RSQLite_2.3.7               utf8_1.2.4                 
##  [83] tidyr_1.3.1                 hexbin_1.28.4              
##  [85] data.table_1.16.2           recipes_1.1.0              
##  [87] FNN_1.1.4.1                 class_7.3-22               
##  [89] prettyunits_1.2.0           PSMatch_1.11.0             
##  [91] httr_1.4.7                  htmlwidgets_1.6.4          
##  [93] S4Arrays_1.7.1              ModelMetrics_1.2.2.2       
##  [95] pkgconfig_2.0.3             gtable_0.3.6               
##  [97] timeDate_4041.110           blob_1.2.4                 
##  [99] impute_1.81.0               XVector_0.47.0             
## [101] htmltools_0.5.8.1           bookdown_0.41              
## [103] MALDIquant_1.22.3           clue_0.3-65                
## [105] scales_1.3.0                png_0.1-8                  
## [107] gower_1.0.1                 knitr_1.48                 
## [109] reshape2_1.4.4              coda_0.19-4.1              
## [111] nlme_3.1-166                curl_6.0.0                 
## [113] GlobalOptions_0.1.2         proxy_0.4-27               
## [115] cachem_1.1.0                stringr_1.5.1              
## [117] parallel_4.5.0              mzID_1.45.0                
## [119] vsn_3.75.0                  pillar_1.9.0               
## [121] grid_4.5.0                  vctrs_0.6.5                
## [123] pcaMethods_1.99.0           randomForest_4.7-1.2       
## [125] dbplyr_2.5.0                xtable_1.8-4               
## [127] evaluate_1.0.1              magick_2.8.5               
## [129] tinytex_0.54                mvtnorm_1.3-2              
## [131] cli_3.6.3                   compiler_4.5.0             
## [133] rlang_1.1.4                 crayon_1.5.3               
## [135] future.apply_1.11.3         labeling_0.4.3             
## [137] LaplacesDemon_16.1.6        mclust_6.1.1               
## [139] QFeatures_1.17.0            affy_1.85.0                
## [141] plyr_1.8.9                  stringi_1.8.4              
## [143] munsell_0.5.1               Biostrings_2.75.0          
## [145] lazyeval_0.2.2              Matrix_1.7-1               
## [147] hms_1.1.3                   bit64_4.5.2                
## [149] future_1.34.0               KEGGREST_1.47.0            
## [151] statmod_1.5.0               highr_0.11                 
## [153] SummarizedExperiment_1.37.0 kernlab_0.9-33             
## [155] igraph_2.1.1                memoise_2.0.1              
## [157] affyio_1.77.0               bslib_0.8.0                
## [159] sampling_2.10               bit_4.5.0

References

Crook, Oliver M., Lisa M. Breckels, Kathryn S. Lilley, Paul D. W. Kirk, and Laurent Gatto. 2019. “A Bioconductor Workflow for the Bayesian Analysis of Spatial Proteomics.” F1000Research 8 (April): 446. https://doi.org/10.12688/f1000research.18636.1.
Crook, Oliver M., Colin T. R. Davies, Lisa M. Breckels, Josie A. Christopher, Laurent Gatto, Paul D. W. Kirk, and Kathryn S. Lilley. 2022. “Inferring Differential Subcellular Localisation in Comparative Spatial Proteomics Using BANDLE.” Nature Communications 13 (1). https://doi.org/10.1038/s41467-022-33570-9.
Crook, Oliver M., Claire M. Mulvey, Paul D. W. Kirk, Kathryn S. Lilley, and Laurent Gatto. 2018. “A Bayesian Mixture Modelling Approach for Spatial Proteomics.” Edited by Christine Vogel. PLOS Computational Biology 14 (11): e1006516. https://doi.org/10.1371/journal.pcbi.1006516.
Gatto, Laurent, Lisa M. Breckels, Samuel Wieczorek, Thomas Burger, and Kathryn S. Lilley. 2014. “Mass-Spectrometry Based Spatial Proteomics Data Analysis Using pRoloc and pRolocdata.” Bioinformatics.
Gentleman, Robert C., Vincent J. Carey, Douglas M. Bates, Ben Bolstad, Marcel Dettling, Sandrine Dudoit, Byron Ellis, et al. 2004. “Bioconductor: Open Software Development for Computational Biology and Bioinformatics.” Genome Biol 5 (10): –80. https://doi.org/10.1186/gb-2004-5-10-r80.
Gilks, Walter R, Sylvia Richardson, and David Spiegelhalter. 1995. Markov Chain Monte Carlo in Practice. CRC press.
R Development Core Team. 2011. R: A Language and Environment for Statistical Computing. Vienna, Austria: R Foundation for Statistical Computing. http://www.R-project.org/.
Tan, Denise JL, Heidi Dvinge, Andrew Christoforou, Paul Bertone, Alfonso Martinez Arias, and Kathryn S Lilley. 2009. “Mapping Organelle Proteins and Protein Complexes in Drosophila Melanogaster.” Journal of Proteome Research 8 (6): 2667–78.