This tutorial is quite fast and on a very simple data set (2
conditions only), for a more complicated tutorial on the setup please
see vignette('baldur_ups_tutorial')
. First we load
baldur
and setup the model dependent variables we need,
then normalize the data and add the mean-variance trends.
library(baldur)
# Setup design matrix
yeast_design <- model.matrix(~0+factor(rep(1:2, each = 3)))
colnames(yeast_design) <- paste0('ng', c(50, 100))
# Compare the first and second column of the design matrix
# with the following contrast matrix
yeast_contrast <- matrix(c(-1, 1), nrow = 2)
# Set id column
id_col <- colnames(yeast)[1] # "identifier"
# Define the number of parallel workers to use
workers <- floor(parallel::detectCores()/2)
# Since baldur itself does not deal with missing data we remove the
# rows that have missing values for the purpose of the tutorial.
# Else, one would replace the filtering step with imputation but that is outside
# the scope of baldur
yeast_norm <- yeast %>%
# Remove missing data
tidyr::drop_na() %>%
# Normalize data (this might already have been done if imputation was performed)
psrn(id_col) %>%
# Add mean-variance trends
calculate_mean_sd_trends(yeast_design)
Importantly, note that the column names of the design matrix are unique subsets of the names of the columns within the conditions:
colnames(yeast)
#> [1] "identifier" "ng50_1" "ng50_2" "ng50_3" "ng100_1" "ng100_2" "ng100_3"
colnames(yeast_design)
#> [1] "ng50" "ng100"
This is essential for baldur
to know which columns to
use in calculations and to perform transformations.
Next is to infer the mixture of the data and to estimate the
parameters needed for baldur
. First we will setup the
needed variables for using baldur
without partitioning the
data. Then, partitioning and setting up baldur
after
trend-partitioning
Finally we sample the posterior of each row in the data. First we sample assuming a single trend, then using the partitioning.
# Single trend
gr_results <- gr_model %>%
# Add hyper-priors for sigma
estimate_gamma_hyperparameters(yeast_norm) %>%
infer_data_and_decision_model(
id_col,
yeast_design,
yeast_contrast,
unc_gr,
clusters = workers # I highly recommend using parallel workers/clusters
) # this will greatly reduce the time of running baldur
# The top hits then looks as follows:
gr_results %>%
dplyr::arrange(err)
#> # A tibble: 1,802 × 22
#> identifier comparison err lfc lfc_025 lfc_50 lfc_975 lfc_eff lfc_rhat sigma sigma_025 sigma_50 sigma_975 sigma_eff sigma_rhat lp
#> <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
#> 1 Cre09.g40… ng100 vs … 3.78e-213 6.18 5.79 6.18 6.56 1780. 1.00 0.123 0.0664 0.112 0.244 986. 1.00 14.5
#> 2 Cre12.g55… ng100 vs … 1.05e-178 1.61 1.50 1.61 1.72 3090. 1.00 0.0492 0.0265 0.0446 0.0979 1115. 1.00 28.9
#> 3 sp|P37302… ng100 vs … 4.15e-156 1.51 1.40 1.51 1.63 3107. 1.00 0.0464 0.0249 0.0421 0.0958 1158. 1.00 29.7
#> 4 sp|P38788… ng100 vs … 3.49e-151 1.07 0.994 1.07 1.16 3109. 1.00 0.0359 0.0190 0.0328 0.0715 1345. 1.00 32.4
#> 5 Cre14.g61… ng100 vs … 5.51e-143 -4.54 -4.89 -4.54 -4.17 2439. 1.00 0.142 0.0774 0.130 0.276 1361. 1.00 15.6
#> 6 Cre10.g42… ng100 vs … 9.10e-134 4.16 3.82 4.16 4.51 2959. 1.00 0.147 0.0828 0.135 0.275 1468. 1.00 17.7
#> 7 Cre12.g53… ng100 vs … 1.52e- 90 1.41 1.28 1.41 1.55 2580. 1.00 0.0587 0.0313 0.0533 0.117 1029. 1.00 26.8
#> 8 sp|P07259… ng100 vs … 4.38e- 87 1.14 1.02 1.14 1.26 3264. 1.00 0.0518 0.0281 0.0475 0.102 1299. 1.00 27.7
#> 9 Cre06.g30… ng100 vs … 5.85e- 86 4.20 3.80 4.21 4.62 2471. 1.00 0.150 0.0818 0.136 0.311 964. 1.00 13.9
#> 10 sp|P19882… ng100 vs … 2.18e- 85 0.883 0.794 0.882 0.976 2997. 1.00 0.0412 0.0225 0.0377 0.0794 1530. 1.00 33.1
#> # ℹ 1,792 more rows
#> # ℹ 6 more variables: lp_025 <dbl>, lp_50 <dbl>, lp_975 <dbl>, lp_eff <dbl>, lp_rhat <dbl>, warnings <list>
Here err
is the probability of error, i.e., the two
tail-density supporting the null-hypothesis, lfc
is the
estimated log2-fold change,
sigma
is the common variance, and lp
is the
log-posterior. Columns without suffix shows the mean estimate from the
posterior, while the suffixes _025
, _50
, and
_975
, are the 2.5, 50.0, and 97.5, percentiles,
respectively. The suffixes _eff
and _rhat
are
the diagnostic variables returned by rstan
(please see the
Stan manual for details). In general, a larger _eff
indicates a better sampling efficiency, and _rhat
compares
the mixing within chains against between the chains and should be
smaller than 1.05.
First we fit the LGMR model:
We can print the model with print
and extract parameters
of interest with coef
:
print(yeast_lgmr, pars = c("coef", "aux"))
#>
#> LGMR Model
#> μ = exp(-1.847 - 0.325 f(ȳ)) + κ exp(θ(7.52 - 0.474 f(ȳ)))
#>
#> Coefficients:
#> mean se_mean sd 2.5% 25% 50% 75% 97.5% n_eff Rhat
#> γ_0L 7.520 0.000783 0.0504 7.424 7.486 7.520 7.554 7.621 4141 1
#> γ_ȳ 0.325 0.000361 0.0246 0.277 0.308 0.325 0.342 0.374 4659 1
#> γ_ȳL 0.474 0.000542 0.0453 0.386 0.444 0.474 0.505 0.563 6999 1
#> γ_0 -1.847 0.000731 0.0295 -1.904 -1.867 -1.847 -1.827 -1.788 1629 1
#>
#>
#> Auxiliary:
#> mean se_mean sd 2.5% 25% 50% 75% 97.5% n_eff Rhat
#> α 4.48 0.007693 0.2980 3.918 4.269 4.466 4.673 5.079 1500 1
#> NRMSE 0.52 0.000405 0.0248 0.475 0.503 0.519 0.536 0.572 3761 1
# Extract the regression, alpha, and theta parameters and the NRMSE.
yeast_lgmr_coef <- coef(yeast_lgmr, pars = "all")
Baldur allows for two ways to plot the LGMR model,
plot_lgmr_regression
, and
plot_regression_field
. The first plots lines of three cases
of θ, 0
,
0.5
, and 1
, and colors each peptide according
to their infered θ. They can
be plotted accordingly:
In generall, a good fit spreads out and captures the overall M-V trend. The main M-V density is captured by the common trend while the sparser part is captured by the latent trend.
We can then estimate the uncertainty similar to the GR case:
Then running the data and decision model:
baldur
have two ways of visualizing the results 1)
plotting sigma vs LFC and 2) Volcano plots. To plot sigma against LFC we
use plot_sa
:
gr_results %>%
plot_sa(
alpha = .05, # Level of significance
lfc = 1 # Add LFC lines
)
lgmr_results %>%
plot_sa(
alpha = .05, # Level of significance
lfc = 1 # Add LFC lines
)
While it is hard to see with this few examples, in general a good
decision is indicated by a lack of a trend between σ and LFC. To make a volcano plot
one uses plot_volcano
in a similar fashion to
plot_sa
: