Introduction to corncob, no phyloseq
Bryan D Martin
2024-01-10
Introduction
Effectively modeling microbial relative abundance poses a number of
statistical challenges, including:
Here, we introduce corncob, an individual taxon
regression model that uses abundance tables and sample data.
corncob is able to model differential abundance and
differential variability, and addresses each of the challenges presented
above.
Install corncob using:
remotes::install_github("statdivlab/corncob")
To begin, we load our example data set we load our example data set
as three different data frames
library(corncob)
library(magrittr)
data(soil_phylo_sample)
data(soil_phylo_otu)
data(soil_phylo_taxa)
We can see that we have 5 sample variables. Let’s look at the first
few observations.
## Plants DayAmdmt Amdmt ID Day
## S009 1 01 1 D 0
## S204 1 21 1 D 2
## S112 0 11 1 B 1
## S247 0 22 2 F 2
## S026 0 00 0 A 0
## S023 1 00 0 C 0
Our covariates are as follows:
Additionally, we have a table that gives us counts of the phyla by
sample. Let’s take a look at the first 5 phyla and the first 5
samples.
## S009 S204 S112 S247 S026
## OTU.43 350 74 300 70 43
## OTU.2 1796 4204 1752 695 945
## OTU.187 280 709 426 100 139
## OTU.150 33 151 18 13 28
## OTU.91 0 0 184 0 0
Finally, we have a taxonomy table with 7 taxonomic ranks.
## Kingdom Phylum Class Order
## OTU.43 "Bacteria" "Nitrospirae" "Nitrospira" "Nitrospirales"
## OTU.2 "Bacteria" "Proteobacteria" "Alphaproteobacteria" "Rhizobiales"
## OTU.187 "Bacteria" "Acidobacteria" "Acidobacteriia" "Acidobacteriales"
## Family Genus Species
## OTU.43 "Nitrospiraceae" "Nitrospira" ""
## OTU.2 "Bradyrhizobiaceae" "Bradyrhizobium" ""
## OTU.187 "Koribacteraceae" "" ""
Fitting a Model
First, let’s subset our samples to only include those with the
DayAmdmt covariate equal to 11 or 21 and then collapse the
samples to the phylum level. We have already done this and saved the
resulting sample data frame and otu data frame.
data(soil_phylum_small_sample)
sample_data <- soil_phylum_small_sample
data(soil_phylum_small_otu)
data <- soil_phylum_small_otu
Note that collapsing the samples is not necessary, and this model can
work at any taxonomic rank. However, we will later be fitting a model to
every taxa. We can see that by agglomerating taxa to the phylum level,
we have gone from from 7770 to 40 taxa. Thus we collapse in order to
increase the speed for the purposes of this tutorial.
Now we fit our model. We will demonstrate with Proteobacteria. We
will need to make a data frame that includes our sample data as well as
the Proteobacteria counts and sequencing depths for each sample.
pro_data <- cbind(sample_data,
W = unlist(data["Proteobacteria", ]),
M = colSums(data))
For now, we will not include any covariates, so we use
~ 1 as our model formula responses.
corncob <- bbdml(formula = cbind(W, M - W) ~ 1,
phi.formula = ~ 1,
data = pro_data)
Interpreting a Model
First, let’s plot the data with our model fit on the relative
abundance scale. To do this, we simply type:
The points represent the relative abundances. The bars represent the
95% prediction intervals for the observed relative abundance by sample.
The parameter B determines the number of bootstrap
simulations used to approximate the prediction intervals. For purposes
of this tutorial, we use a small value B = 50 for
computational purposes, but recommend a higher setting for more accurate
intervals, such as the default B = 1000.
Now let’s look at the same plot, but on the counts scale with 95%
prediction intervals (since counts is not a parameter). To do this, we
add the option total = TRUE to our plotting code.
plot(corncob, total = TRUE, B = 50)
Finally, let’s color the plot by the DayAmdmt covariate.
To do this, we add the option color = "DayAmdmt" to our
plotting code.
plot(corncob, total = TRUE, color = "DayAmdmt", B = 50)
plot(corncob, color = "DayAmdmt", B = 50)
Notice that this plot also reorders our samples so that groups appear
together so that they are easier to compare.
We can observe on this plot that it might be of interest to
distinguish between the two groups with covariates. The average
empirical relative abundance for the samples with
DayAmdmt = 21 tends to be lower and less variable than the
samples with DayAmdmt = 11.
Adding covariates
Let’s try modeling the expected relative abundance and the
variability of the counts with DayAmdmt as a covariate. We
do this by modifying formula and phi.formula
as:
corncob_da <- bbdml(formula = cbind(W, M - W) ~ DayAmdmt,
phi.formula = ~ DayAmdmt,
data = pro_data)
Let’s also plot this data on both the total count and relative
abundance scales.
plot(corncob_da, color = "DayAmdmt", total = TRUE, B = 50)
plot(corncob_da, color = "DayAmdmt", B = 50)
Visually, the model with covariates seems to provide a much better
fit to the data, but how can we compare the two models
statistically?
Model Selection
Let’s use a likelihood ratio test to select our final model for this
taxon. We want to test the null hypothesis that the likelihood of the
model with covariates is equal to the likelihood of the model without
covariates. To do this test, we use:
lrtest(mod_null = corncob, mod = corncob_da)
## [1] 4.550571e-05
We obtain a p-value much smaller than a cut-off of 0.05. Therefore we
conclude that there is a statistically significant difference in the
likelihood of the two models. Thus, we probably want to use the model
with covariates for this taxon.
Parameter Interpretation
Now that we have chosen our model, let’s interpret our model output.
To see a summary of the model, type:
##
## Call:
## bbdml(formula = cbind(W, M - W) ~ DayAmdmt, phi.formula = ~DayAmdmt,
## data = pro_data)
##
##
## Coefficients associated with abundance:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) -0.44595 0.03604 -12.375 7.18e-13 ***
## DayAmdmt21 -0.16791 0.04067 -4.129 0.000297 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
##
## Coefficients associated with dispersion:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) -5.3077 0.3537 -15.008 6.44e-15 ***
## DayAmdmt21 -1.3518 0.5029 -2.688 0.012 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
##
## Log-likelihood: -286.53
This output will look familiar if you have done regression analysis
in R in the past. Covariates associated with the expected relative
abundance are presented separately from covariates associated with the
variance of the counts are preceded by.
From this model summary, we can see that the DayAmdmt21
abundance coefficient is negative and statistically significant. This
suggests that this taxon is differentially-abundant across
DayAmdmt, and that samples with DayAmdmt = 21
are expected to have a lower relative abundance. This matches what we
saw from the observed abundances.
We can also see that the DayAmdmt21 dispersion
coefficient is negative and statistically significant. This suggests
that this taxon is differentially-variable across DayAmdmt,
and that samples with DayAmdmt = 21 are expected to have a
lower variability. This matches what we saw from the observed
abundances.
Analysis for Multiple Taxa
What if we want to test all the taxa in our data to see if they are
differentially-abundant or differentially-variable? We use the
differentialTest function. It will perform the above tests
on all taxa, and it will control the false discovery rate to account for
multiple comparisons.
Next, we use the differentialTest command. We specify
the covariates of our model using formula and
phi.formula as before, except we no longer include the
response term because we are testing multiple taxa. We also specify
which covariates we want to test for by removing them in the
formula_null and phi.formula_null
arguments.
The difference between the formulas and the null version of the
formulas are the variables that we test. We will go through several
examples, starting with a test for differential abundance across the
DayAmdmt coefficient.
We set fdr_cutoff to be our controlled false discovery
rate.
Here, we will pass in our entire count table for all phyla as the
data argument (setting taxa_are_rows to
TRUE), and the sample data frame as the
sample_data argument.
set.seed(1)
da_analysis <- differentialTest(formula = ~ DayAmdmt,
phi.formula = ~ DayAmdmt,
formula_null = ~ 1,
phi.formula_null = ~ DayAmdmt,
test = "Wald", boot = FALSE,
data = data,
sample_data = sample_data,
taxa_are_rows = TRUE,
fdr_cutoff = 0.05)
We can see the output of the function by calling it:
## Object of class differentialTest
##
## $p: p-values
## $p_fdr: FDR-adjusted p-values
## $significant_taxa: taxa names of the statistically significant taxa
## $significant_models: model summaries of the statistically significant taxa
## $all_models: all model summaries
## $restrictions_DA: covariates tested for differential abundance
## $restrictions_DV: covariates tested for differential variability
## $discriminant_taxa_DA: taxa for which at least one covariate associated with the abundance was perfectly discriminant
## $discriminant_taxa_DV: taxa for which at least one covariate associated with the dispersion was perfectly discriminant
##
## plot( ) to see a plot of tested coefficients from significant taxa
We can see a list of differentially-abundant taxa using:
da_analysis$significant_taxa
## [1] "Proteobacteria" "Gemmatimonadetes" "Bacteroidetes" "Cyanobacteria"
## [5] "Firmicutes" "Planctomycetes" "Armatimonadetes" "Spirochaetes"
## [9] "Elusimicrobia" "BRC1" "OP3" "FBP"
## [13] "Chlorobi" "TM6"
In this case, we identified 14 taxa that are differentially-abundant
across DayAmdmt (out of the 39 taxa tested).
We can see a list of differentially-variable taxa using:
set.seed(1)
dv_analysis <- differentialTest(formula = ~ DayAmdmt,
phi.formula = ~ DayAmdmt,
formula_null = ~ DayAmdmt,
phi.formula_null = ~ 1,
test = "LRT", boot = FALSE,
data = data,
sample_data = sample_data,
taxa_are_rows = TRUE,
fdr_cutoff = 0.05)
dv_analysis$significant_taxa
## [1] "Acidobacteria" "Cyanobacteria" "Spirochaetes" "Elusimicrobia"
## [5] "FBP"
In this case, we identified 5 taxa that are differentially-variable
across DayAmdmt (out of the 40 taxa tested).
We can examine a subset of the p-values of our tests using:
## Acidobacteria Proteobacteria Gemmatimonadetes Actinobacteria
## 6.509418e-01 3.642734e-05 3.270448e-13 3.703096e-01
## [Thermi]
## 1.031419e-01
We can examine a subset of the p-values after controlling for the
false discovery rate using:
## Acidobacteria Proteobacteria Gemmatimonadetes Actinobacteria
## 7.811302e-01 1.457094e-04 3.924537e-12 5.172744e-01
## [Thermi]
## 2.062838e-01
where the values are now adjusted to control the false discovery rate
at 0.05.
We can also plot the model coefficients of our results:
Here, we can see that for Bacteria_Armatimonadetes, the
effect of DayAmdmt21 is positive compared to the baseline
(in this case, DayAmdmt11).
Finally, we can see a list of any taxa for which we were not able to
fit a model using:
which(is.na(da_analysis$p)) %>% names
## [1] "GN04" "GN02" "MVP-21"
In this case, we weren’t able to fit GN04 automatically.
It’s worthwhile to investigate the OTU individually if this is the
case.
It may be that the model is overparameterized because there aren’t
enough observations, or it may just be that the initializations were
invalid for that taxa and it needs to be re-evaluated with new
initializations.
Let’s first try examining the data.
## S204 S112 S134 S207 S202 S139 S122 S212 S117 S104 S214 S109 S217 S229 S132
## GN04 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
## S209 S227 S107 S237 S224 S127 S137 S114 S124 S119 S219 S232 S129 S102 S234
## GN04 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0
## S222 S239
## GN04 0 0
We see that the observed counts of OTU is zero in all samples except
for S102, where we observed a single count. We should be
skeptical of any statistical model fit on a single
observed count!
corncob is stable, but if you notice any issues, please
log them on
Github to help us help you!
Examples of Answering Scientific Questions
We will now walk through several scientific questions of interest and
show how they can be answered using hypothesis testing with
corncob. Note that Day and Amdmt
are both factor covariates with levels 0, 1, and 2.
Note that some of these are rather strange tests, and shown for
demonstration of the flexibility of the model only. Normally, when
testing for differential variability across a covariate, we recommend
always controlling for the effect of that covariate on the abundance. We
first demonstrate examples with the small version of the soil dataset
(note, this is a slightly different subset of the data from that used in
corncob-intro.Rmd).
Testing for differential abundance across Day, without
controlling for anything else:
ex1 <- differentialTest(formula = ~ Day,
phi.formula = ~ 1,
formula_null = ~ 1,
phi.formula_null = ~ 1,
data = data,
taxa_are_rows = TRUE,
sample_data = sample_data,
test = "Wald", boot = FALSE,
fdr_cutoff = 0.05)
plot(ex1)
Testing for differential abundance across Day,
controlling for the effect of Day on dispersion:
ex2 <- differentialTest(formula = ~ Day,
phi.formula = ~ Day,
formula_null = ~ 1,
phi.formula_null = ~ Day,
data = data,
taxa_are_rows = TRUE,
sample_data = sample_data,
test = "Wald", boot = FALSE,
fdr_cutoff = 0.05)
plot(ex2)
Jointly testing for differential abundance and differential
variability across Day:
ex3 <- differentialTest(formula = ~ Day,
phi.formula = ~ Day,
formula_null = ~ 1,
phi.formula_null = ~ 1,
data = data,
taxa_are_rows = TRUE,
sample_data = sample_data,
test = "Wald", boot = FALSE,
fdr_cutoff = 0.05)
plot(ex3)
More examples of answering scientific questions with a larger subset
of this soil data set and with an IBD data set can be seen in
corncob-intro.Rmd.