mrgsolve
forest_plot
Here we illustrate the varying one covariate at a time approach, keeping all the others at the reference value or category. A two-compartment pharmacokinetic (PK) model defined with ordinary differential equations (ODEs) is used. Covariates Weight, Albumin and Sex had effects on the Clearance (CL) model parameter while Weight and Sex had effects on the Volume of distribution (V) model parameter. For simplicity, there were no included covariates effects on other PK parameters such as peripheral clearance or volume. The approach is general and can be easily extended to any other ODEs model with multiple covariate effects on multiple model parameters.
mrgsolve
codepkmodelcov <- '
$PARAM @annotated
KA : 0.5 : Absorption rate constant Ka (1/h)
CL : 4 : Clearance CL (L/h)
V : 10 : Central volume Vc (L)
Vp : 50 : Peripheral volume Vp (L)
Qp : 10 : Intercompartmental clearance Q (L/h)
CLALB : -0.8 : Ablumin on CL (ref. 45 g/L)
CLSEX : 0.2 : Sex on CL (ref. Female)
CLWT : 1 : Weight on CL (ref. 85 kg)
VSEX : 0.07 : Sex on Vc (ref. Female)
VWT : 1 : Weight on Vc (ref. 85 kg)
$PARAM @annotated // reference values for covariate
WT : 85 : Weight (kg)
SEX : 0 : Sex (0=Female, 1=Male)
ALB : 45 : Albumin (g/L)
$PKMODEL cmt="GUT CENT PER", depot=TRUE, trans=11
$MAIN
double CLi = CL *
pow((ALB/45.0), CLALB)*
(SEX == 1.0 ? (1.0+CLSEX) : 1.0)*
pow((WT/85.0), CLWT)*exp(nCL);
double V2i = V *
(SEX == 1.0 ? (1.0+VSEX) : 1.0)*
pow((WT/85.0), VWT)*exp(nVC);
double KAi = KA;
double V3i = Vp *pow((WT/85.0), 1);
double Qi = Qp *pow((WT/85.0), 0.75);
$OMEGA @annotated @block
nCL : 0.09 : ETA on CL
nVC : 0.01 0.09 : ETA on Vc
$TABLE
double CP = CENT/V2i;
$CAPTURE CP KAi CLi V2i V3i Qi WT SEX ALB
'
modcovsim <- mcode("codepkmodelcov", codepkmodelcov)
partab <- setDT(modcovsim@annot$data)[block=="PARAM", .(name, descr, unit)]
partab <- merge(partab, melt(setDT(modcovsim@param@data), meas=patterns("*"), var="name"))
knitr::kable(partab)
name | descr | unit | value |
---|---|---|---|
ALB | Albumin | g/L | 45.00 |
CL | Clearance CL | L/h | 4.00 |
CLALB | Ablumin on CL | ref. 45 g/L | -0.80 |
CLSEX | Sex on CL | ref. Female | 0.20 |
CLWT | Weight on CL | ref. 85 kg | 1.00 |
KA | Absorption rate constant Ka | 1/h | 0.50 |
Qp | Intercompartmental clearance Q | L/h | 10.00 |
SEX | Sex | 0=Female, 1=Male | 0.00 |
V | Central volume Vc | L | 10.00 |
VSEX | Sex on Vc | ref. Female | 0.07 |
VWT | Weight on Vc | ref. 85 kg | 1.00 |
Vp | Peripheral volume Vp | L | 50.00 |
WT | Weight | kg | 85.00 |
We simulate the reference subject having the reference covariate
values defined in the model which are:
Weight = 85 kg, Sex = Female and Albumin = 45 g/L. We also keep the
between subject variability (BSV) to illustrate its effects on the
concentration-time profiles on linear and log linear scales.
idata <- data.table(ID=1:nbsvsubjects, WT=85, SEX=0, ALB=45)
ev1 <- ev(time = 0, amt = 100, cmt = 1)
data.dose <- ev(ev1)
data.dose <- setDT(as.data.frame(data.dose))
data.all <- data.table(idata, data.dose)
outputsim <- modcovsim %>%
data_set(data.all) %>%
mrgsim(end = 24, delta = 0.25) %>%
as.data.frame %>%
as.data.table
outputsim$SEX <- factor(outputsim$SEX, labels="Female")
# Only plot a random sample of N=500
set.seed(678549)
plotdata <- outputsim[ID %in% sample(unique(ID), 500)]
# New facet label names for dose variable
albumin.labs <- c("albumin: 45 ng/mL")
names(albumin.labs) <- c("45")
wt.labs <- c("weight: 85 kg")
names(wt.labs) <- c("85")
p1 <- ggplot(plotdata, aes(time, CP, group = ID)) +
geom_line(alpha = 0.2, size = 0.1) +
facet_grid(~ WT + ALB + SEX,
labeller = labeller(ALB = albumin.labs,
WT = wt.labs)) +
labs(y = "Plasma Concentrations", x = "Time (h)")
p2 <- ggplot(plotdata, aes(time, CP, group = ID)) +
geom_line(alpha = 0.2, size = 0.1) +
facet_grid(~ WT + ALB + SEX,
labeller = labeller(ALB = albumin.labs,
WT = wt.labs)) +
scale_y_log10() +
labs(y = "Plasma~Concentrations\n(logarithmic scale)", x = "Time (h)")
#labs(y = expression(Log[10]~Plasma~Concentrations), x = "Time (h)")+
egg::ggarrange(p1, p2, ncol = 2)
In this section we compute the PK parameters of interest, provide a
plot of the parameters as well as of the standardized ones. We also
summarize and report the BSV as ranges of 50 and 90% of patients for
each PK parameter. Later on we might choose to include these ranges in
the coveffectsplot
or not.
derive.exposure <- function(time, CP) {
n <- length(time)
x <- c(
Cmax = max(CP),
Clast = CP[n],
AUC = sum(diff(time) * (CP[-1] + CP[-n])) / 2
)
data.table(paramname=names(x), paramvalue=x)
}
refbsv <- outputsim[, derive.exposure(time, CP), by=.(ID, WT, SEX, ALB)]
p3 <- ggplot(refbsv, aes(
x = paramvalue,
y = paramname,
fill = factor(..quantile..),
height = ..ndensity..)) +
facet_wrap(~ paramname, scales="free", ncol=1) +
stat_density_ridges(
geom="density_ridges_gradient", calc_ecdf=TRUE,
quantile_lines=TRUE, rel_min_height=0.001, scale=0.9,
quantiles=c(0.05, 0.25, 0.5, 0.75, 0.95)) +
scale_fill_manual(
name = "Probability",
values = c("white", "#FF000050", "#FF0000A0", "#FF0000A0", "#FF000050", "white"),
labels = c("(0, 0.05]", "(0.05, 0.25]",
"(0.25, 0.5]", "(0.5, 0.75]",
"(0.75, 0.95]", "(0.95, 1]")) +
theme_bw() +
theme(
legend.position = "none",
axis.text.y = element_blank(),
axis.ticks.y = element_blank(),
axis.title.y = element_blank()) +
labs(x="PK Parameters", y="") +
scale_x_log10() +
coord_cartesian(expand=FALSE)
# Obtain the standardized parameter value by dividing by the median.
refbsv[, stdparamvalue := paramvalue/median(paramvalue), by=paramname]
p4 <- ggplot(refbsv, aes(
x = stdparamvalue,
y = paramname,
fill = factor(..quantile..),
height = ..ndensity..)) +
facet_wrap(~ paramname, scales="free", ncol=1) +
stat_density_ridges(
geom="density_ridges_gradient", calc_ecdf=TRUE,
quantile_lines=TRUE, rel_min_height=0.001, scale=0.9,
quantiles=c(0.05, 0.25, 0.5, 0.75, 0.95)) +
scale_fill_manual(
name="Probability",
values=c("white", "#FF000050", "#FF0000A0", "#FF0000A0", "#FF000050", "white"),
labels = c("(0, 0.05]", "(0.05, 0.25]",
"(0.25, 0.5]", "(0.5, 0.75]",
"(0.75, 0.95]", "(0.95, 1]")) +
theme_bw() +
theme(
legend.position = "none",
axis.text.y = element_blank(),
axis.ticks.y = element_blank(),
axis.title.y = element_blank()) +
labs(x="Standardized PK Parameters", y="") +
scale_x_log10() +
coord_cartesian(expand=FALSE, xlim = c(0.3,3))
p3+p4
Ranges of BSV for each PK Parameter:
bsvranges <- refbsv[,list(
P05 = quantile(stdparamvalue, 0.05),
P25 = quantile(stdparamvalue, 0.25),
P50 = quantile(stdparamvalue, 0.5),
P75 = quantile(stdparamvalue, 0.75),
P95 = quantile(stdparamvalue, 0.95)), by = paramname]
bsvranges
paramname | P05 | P25 | P50 | P75 | P95 |
---|---|---|---|---|---|
Cmax | 0.7918743 | 0.9184431 | 1 | 1.077698 | 1.185392 |
Clast | 0.4752010 | 0.7899361 | 1 | 1.286067 | 1.735681 |
AUC | 0.7063883 | 0.8917846 | 1 | 1.123854 | 1.313933 |
Based on our observed covariate data, we compute percentiles of interest that we will use to simulate data at. Common practice is to compute the 5,25,75,95 percentiles (the median being the reference). In some cases, we might want to explore the min, max or other extreme case scenarios. Care should be taken as this approach might generate unrealistic combination of covariates that can never appear in a real patient. The utility function expand.modelframe (written by Benjamin Rich) is defined in the setup section of the vignette and can be found in the source code. It facilitates the creation of a set of covariate values varying one at a time.
reference.values <- data.frame(WT = 85, ALB = 45, SEX = 0)
covcomb <- expand.modelframe( #defined at the top and now expand_modelframe was added
WT = c(56, 72, 98, 128), # P05, P25, P50, P75, P95
ALB = c(40, 50), # P05, P50, P95
SEX = c(1), # Reference is for SEX=0 (female)
rv = reference.values)
# Add the reference
covcomb <- rbind(covcomb, data.table(reference.values, covname="REF"))
covcomb$ID <- 1:nrow(covcomb)
covcomb
WT | ALB | SEX | covname | ID |
---|---|---|---|---|
56 | 45 | 0 | WT | 1 |
72 | 45 | 0 | WT | 2 |
98 | 45 | 0 | WT | 3 |
128 | 45 | 0 | WT | 4 |
85 | 40 | 0 | ALB | 5 |
85 | 50 | 0 | ALB | 6 |
85 | 45 | 1 | SEX | 7 |
85 | 45 | 0 | REF | 8 |
As a first step, we simulate without uncertainty and without BSV
using zero_re()
at unique combination of covariates and
provide a plot to visualize the effects.
idata <- data.table::copy(covcomb)
idata$covname <- NULL
ev1 <- ev(time=0, amt=100, cmt=1)
data.dose <- as.data.frame(ev1)
data.all <- data.table(idata, data.dose)
outcovcomb<- modcovsim %>%
data_set(data.all) %>%
zero_re() %>%
mrgsim(end=24, delta=0.25) %>%
as.data.frame %>%
as.data.table
outcovcomb$SEX <- factor(outcovcomb$SEX, labels=c("Female", "Male"))
albumin.labs <- c("albumin: 40 ng/mL","albumin: 45 ng/mL","albumin: 50 ng/mL")
names(albumin.labs) <- c("40","45","50")
wt.labs <- c("weight: 56 kg","weight: 72 kg","weight: 85 kg","weight: 98 kg","weight: 128 kg")
names(wt.labs) <- c("56","72","85","98","128")
pkprofiletypical <- ggplot(outcovcomb, aes(x=time, y=CP, col=factor(WT), linetype=SEX)) +
geom_line(aes(group=ID), alpha = 1, linewidth = 1) +
facet_nested(ALB + SEX ~ WT,
labeller= labeller(ALB = albumin.labs, WT = wt.labs, SEX = label_value),
switch = "y") +
labs(
x = "Time (h)",
y = "Plasma Concentrations",
linetype = "Sex",
colour = "Weight",
caption = "Simulation without Uncertainty and without BSV") +
coord_cartesian(ylim=c(0,4))
pkprofiletypical <- pkprofiletypical +theme_bw(base_size = 13)+
theme(axis.title.y = element_text(size=15))+
guides(colour = guide_legend(override.aes = list(alpha = 1, linewidth = 1)),
linetype = guide_legend(override.aes = list(linewidth = 1), order = 1))+
coord_cartesian(ylim=c(0,4))
pkprofiletypical
First, we will invent a varcov matrix by assuming 25% relative standard errors and correlations of 0.2 across the board. We then simulate a 100 set of parameters using a multivariate normal (kept at 100 for the vignette, use more replicates for a real project). Also, unless the model was written in a way to allow unconstrained parameter values, care should be taken to make sure the simulated parameters are valid and make sense. In a real modeling applications, the software used to fit the model will provide a variance covariance matrix, if possible. Otherwise, the user can rely on set of parameters generated from bootstrap replicates or SIR run.
Variance Covariance Matrix of fixed effects:
theta <- unclass(as.list(param(modcovsim)))
theta[c("WT", "SEX", "ALB")] <- NULL
theta <- unlist(theta)
as.data.frame(t(theta))
KA | CL | V | Vp | Qp | CLALB | CLSEX | CLWT | VSEX | VWT |
---|---|---|---|---|---|---|---|---|---|
0.5 | 4 | 10 | 50 | 10 | -0.8 | 0.2 | 1 | 0.07 | 1 |
#note that from RsNLME or NONMEM or Monolix the software will give you the varcov just read it it in.
# example NONMEM read in your run .cov
# varcov <- read.csv(paste("runxyz",".cov",sep=""), header=TRUE, skip=1, sep="")
varcov <- cor2cov(
matrix(0.2, nrow=length(theta), ncol=length(theta)),
sd=theta*0.25)
rownames(varcov) <- colnames(varcov) <- names(theta)
as.data.frame(varcov)
KA | CL | V | Vp | Qp | CLALB | CLSEX | CLWT | VSEX | VWT | |
---|---|---|---|---|---|---|---|---|---|---|
KA | 0.0156250 | 0.0250 | 0.06250 | 0.31250 | 0.06250 | -5e-03 | 0.001250 | 0.006250 | 0.0004375 | 0.006250 |
CL | 0.0250000 | 1.0000 | 0.50000 | 2.50000 | 0.50000 | -4e-02 | 0.010000 | 0.050000 | 0.0035000 | 0.050000 |
V | 0.0625000 | 0.5000 | 6.25000 | 6.25000 | 1.25000 | -1e-01 | 0.025000 | 0.125000 | 0.0087500 | 0.125000 |
Vp | 0.3125000 | 2.5000 | 6.25000 | 156.25000 | 6.25000 | -5e-01 | 0.125000 | 0.625000 | 0.0437500 | 0.625000 |
Qp | 0.0625000 | 0.5000 | 1.25000 | 6.25000 | 6.25000 | -1e-01 | 0.025000 | 0.125000 | 0.0087500 | 0.125000 |
CLALB | -0.0050000 | -0.0400 | -0.10000 | -0.50000 | -0.10000 | 4e-02 | -0.002000 | -0.010000 | -0.0007000 | -0.010000 |
CLSEX | 0.0012500 | 0.0100 | 0.02500 | 0.12500 | 0.02500 | -2e-03 | 0.002500 | 0.002500 | 0.0001750 | 0.002500 |
CLWT | 0.0062500 | 0.0500 | 0.12500 | 0.62500 | 0.12500 | -1e-02 | 0.002500 | 0.062500 | 0.0008750 | 0.012500 |
VSEX | 0.0004375 | 0.0035 | 0.00875 | 0.04375 | 0.00875 | -7e-04 | 0.000175 | 0.000875 | 0.0003063 | 0.000875 |
VWT | 0.0062500 | 0.0500 | 0.12500 | 0.62500 | 0.12500 | -1e-02 | 0.002500 | 0.012500 | 0.0008750 | 0.062500 |
Second, we generate the sim_parameters dataset using
mvrnorm
and then incorporate the uncertainty by simulating
using a different set of parameters (row) for each replicate.
First Few Rows of a Dataset Containing Simulated Fixed Effects with Uncertainty:
set.seed(678549)
# mvtnorm::rmvnorm is another option that can be explored
# if you have run a bootstrap just use the parameters data generated using all replicates
sim_parameters <- MASS::mvrnorm(nsim, theta, varcov, empirical=T) %>% as.data.table
head(sim_parameters)
KA | CL | V | Vp | Qp | CLALB | CLSEX | CLWT | VSEX | VWT |
---|---|---|---|---|---|---|---|---|---|
0.3525490 | 3.200401 | 7.920479 | 35.80060 | 9.154611 | -0.7968278 | 0.1469476 | 0.4792149 | 0.0656564 | 1.0745008 |
0.5249089 | 3.677540 | 6.911500 | 45.37793 | 7.117828 | -0.3630370 | 0.0905770 | 0.8847586 | 0.0607989 | 0.3047879 |
0.4832810 | 4.066721 | 9.170236 | 51.89540 | 8.947875 | -1.0678602 | 0.1873986 | 1.1481026 | 0.0657778 | 1.1339100 |
0.5278150 | 2.346888 | 6.033332 | 44.26334 | 7.974338 | -0.6485870 | 0.1496622 | 0.6216810 | 0.0488802 | 0.7770958 |
0.4710727 | 4.065682 | 10.607099 | 38.29416 | 9.425221 | -0.4660593 | 0.3043478 | 0.8386740 | 0.0629995 | 0.8784929 |
0.7963261 | 5.041540 | 13.192930 | 43.99648 | 10.936419 | -0.6903318 | 0.1440387 | 1.0514400 | 0.0560966 | 1.1710921 |
Third, we illustrate how you can iterate over a set of parameters
value using a for
loop. We then overlay the previous
simulation results without uncertainty on the one with uncertainty to
visualize the effect of adding it.
idata <- data.table::copy(covcomb)
idata$covname <- NULL
ev1 <- ev(time=0, amt=100, cmt=1)
data.dose <- as.data.frame(ev1)
iter_sims <- NULL
for(i in 1:nsim) {
data.all <- data.table(idata, data.dose, sim_parameters[i])
out <- modcovsim %>%
data_set(data.all) %>%
zero_re() %>%
mrgsim(start=0, end=24, delta=0.25) %>%
as.data.frame %>%
as.data.table
out[, rep := i]
iter_sims <- rbind(iter_sims, out)
}
iter_sims$SEX <- factor(iter_sims$SEX, labels = c("Female", "Male"))
summary_conc <- function(CP) {
x <- c(
Cmed = median(CP),
Clow = quantile(CP, probs = 0.05),
Cup = quantile(CP, probs = 0.95)
)
data.table(paramname=names(x), paramvalue=x)
}
iter_sims_sum <- iter_sims[, summary_conc(CP), by=.( time, WT, SEX, ALB)]
iter_sims_sum <- spread(iter_sims_sum,paramname,paramvalue)
iter_sims_sum <- as.data.frame(iter_sims_sum)
albumin.labs <- c("albumin: 40 ng/mL","albumin: 45 ng/mL","albumin: 50 ng/mL")
names(albumin.labs) <- c("40","45","50")
wt.labs <- c("weight: 56 kg","weight: 72 kg","weight: 85 kg","weight: 98 kg","weight: 128 kg")
names(wt.labs) <- c("56","72","85","98","128")
pkprofileuncertainty_sum <- ggplot(iter_sims_sum, aes(x=time, col=factor(WT),
fill=factor(WT),
linetype=SEX)) +
geom_ribbon((aes(ymin= `Clow.5%` ,ymax=`Cup.95%`)), alpha = 0.4, linetype = 0)+
geom_line(aes(y=Cmed),alpha=1 , color ="black" , linewidth = 1)+
facet_nested(ALB +SEX ~ WT,
labeller= labeller(ALB = albumin.labs,
WT = wt.labs,
SEX = label_value),switch = "y") +
labs(
x = "Time (h)",
y = "Plasma Concentrations",
linetype = "Sex",
colour = "Uncertainty\n5-95%\nWeight",
fill = "Uncertainty\n5-95%\nWeight",
caption = "Simulation with Uncertainty, without BSV") +
theme_bw(base_size = 13)+
theme(axis.title.y = element_text(size=15))+
guides(fill = guide_legend(override.aes = list(alpha = 0.4, size = 0.4)),
linetype = guide_legend(override.aes = list(linewidth = 1), order = 1))+
coord_cartesian(ylim=c(0,4))
pkprofileuncertainty_sum
Similar to an earlier section, we compute the PK parameters by patient and by replicate standardize by the computed median for reference subject and provide a plot. We add some data manipulation to construct more informative labels that will help in the plotting.
out.df.univariatecov.nca <- iter_sims[, derive.exposure(time, CP), by=.(rep, ID, WT, SEX, ALB)]
refvalues <- out.df.univariatecov.nca[
ALB==45 & WT==85 & SEX=="Female",
.(medparam = median(paramvalue)), by= .(rep,paramname) ]
head(data.frame(refvalues))
rep | paramname | medparam |
---|---|---|
1 | Cmax | 2.0237077 |
1 | Clast | 0.4285660 |
1 | AUC | 23.9392632 |
2 | Cmax | 2.8913245 |
2 | Clast | 0.3178367 |
2 | AUC | 20.7979014 |
Median Parameter Values for the Reference:
covcomb$covvalue[covcomb$covname=="WT"] <- paste(covcomb$WT[covcomb$covname=="WT"],"kg")
covcomb$covvalue[covcomb$covname=="ALB"] <- paste(covcomb$ALB[covcomb$covname=="ALB"],"g/L")
covcomb$covvalue[covcomb$covname=="SEX"] <- "Male"
covcomb$covvalue[covcomb$covname=="REF"] <- "85 kg\nFemale\n45 g/L"
#covcomb[covname=="REF", covvalue := "85 kg Female 45 g/L"]
covcomb <- as.data.table(covcomb)
out.df.univariatecov.nca <- merge(
out.df.univariatecov.nca,
covcomb[, list(ID, covname, covvalue)]
)
setkey(out.df.univariatecov.nca, paramname,rep)
out.df.univariatecov.nca <- merge(
out.df.univariatecov.nca,
refvalues)
out.df.univariatecov.nca[, paramvaluestd := paramvalue/medparam]
boxplotdat <- out.df.univariatecov.nca[covname!="REF"]
boxplotdat[covname=="WT", covname2 := "Weight"]
boxplotdat[covname=="ALB", covname2 := "Albumin"]
boxplotdat[covname=="SEX", covname2 := "Sex"]
boxplotdatREFWT <- out.df.univariatecov.nca[covname=="REF"]
boxplotdatREFWT[, covname2 := "Weight"]
boxplotdatREFWT[, covvalue := covcomb[covname=="REF", covvalue]]
boxplotdatREFSEX <- out.df.univariatecov.nca[covname=="REF"]
boxplotdatREFSEX[, covname2 := "Sex"]
boxplotdatREFSEX[, covvalue := covcomb[covname=="REF", covvalue]]
boxplotdatREFALB <- out.df.univariatecov.nca[covname=="REF"]
boxplotdatREFALB[, covname2 := "Albumin"]
boxplotdatREFALB[, covvalue := covcomb[covname=="REF", covvalue]]
boxplotdat <- rbind(
boxplotdat,
boxplotdatREFWT,
boxplotdatREFSEX,
boxplotdatREFALB)
boxplotdat[paramname=="AUC", paramname2 := "AUC"]
boxplotdat[paramname=="Clast", paramname2 := "C[last]"]
boxplotdat[paramname=="Cmax", paramname2 := "C[max]"]
boxplotdat[, covname2 := factor(covname2, levels=unique(covname2))]
#boxplotdat[, covvalue := factor(covvalue, levels=unique(covvalue))]
boxplotdat[, covvalue := factor(covvalue,
levels=c("56 kg", "72 kg", "40 g/L", "Male", "85 kg\nFemale\n45 g/L", "98 kg", "128 kg", "50 g/L"))]
pkparametersboxplot<- ggplot(boxplotdat, aes(x=covvalue, y=paramvalue))+
facet_grid(paramname2 ~ covname2, scales="free", labeller=label_parsed,
switch="both") +
geom_boxplot() +
labs(y="Parameter Values") +
theme(axis.title=element_blank(),
strip.placement = "outside")
pkparametersboxplot
Here we provide an alternative visual summary of the standardized PK parameters. It shows the distribution, quantiles of interest. It isolates each covariate effects in one panel keeping the reference on its own. It is exactly the same data as the boxplots. Which visual presentation do you prefer? Which one enables you to clearly see and compare the covariate effects?
out.df.univariatecov.nca[covname=="WT", covname2 := "Weight"]
out.df.univariatecov.nca[covname=="ALB", covname2 := "Albumin"]
out.df.univariatecov.nca[covname=="SEX", covname2 := "Sex"]
out.df.univariatecov.nca[covname=="REF", covname2 := "Reference"]
out.df.univariatecov.nca[paramname=="AUC", paramname2 := "AUC"]
out.df.univariatecov.nca[paramname=="Clast", paramname2 := "C[last]"]
out.df.univariatecov.nca[paramname=="Cmax", paramname2 := "C[max]"]
out.df.univariatecov.nca[, covvalue := factor(covvalue, levels=unique(covvalue))]
out.df.univariatecov.nca[, covname2 := factor(covname2, levels=unique(covname2))]
out.df.univariatecov.nca[, paramname2 := factor(paramname2, levels=unique(paramname2))]
ggplot(out.df.univariatecov.nca[out.df.univariatecov.nca$covname!="REF",], aes(
x = paramvaluestd,
y = covvalue,
fill = factor(..quantile..),
height = ..ndensity..)) +
facet_grid(covname2 ~ paramname2,
scales = "free_y",
space = "free",
labeller = label_parsed)+
annotate("rect",
xmin = 0.8,
xmax = 1.25,
ymin = -Inf,
ymax = Inf,
fill = "gray",
alpha = 0.4) +
stat_density_ridges(
geom = "density_ridges_gradient",
calc_ecdf = TRUE,
quantile_lines = TRUE,
rel_min_height = 0.001,
scale = 0.9,
quantiles = c(0.05,0.5, 0.95)) +
scale_x_continuous(
breaks = c(0.25, 0.5, 0.8, 1/0.8, 1/0.5, 1/0.25),
trans = "log") +
scale_fill_manual(
name = "Probability",
values = c("white", "#0000FFA0", "#0000FFA0", "white"),
labels = c("(0, 0.05]", "(0.05, 0.5]","(0.5, 0.95]", "(0.95, 1]")) +
geom_vline(aes(xintercept=1), size=1) +
theme_bw() +
labs(x="Effects Relative to Parameter Reference Value", y="")+
scale_x_continuous(breaks=c(0.25,0.5,0.8,1,1/0.8,1/0.5,1/0.25),
labels=c("1/4","1/2","0.8","1","1.25","2","4"),
trans ="log" )
forest_plot
To contrast the covariate effects with random unexplained variability
we add to the data the BSV intervals computed in an earlier section. We
then do some data manipulation and formatting to produce a plot from the
package function forest_plot
. To simplify we will only keep
AUC before revisiting more than one parameter plots at the end. The user
can also compute an show the BSV for all presented effects not just for
the reference. Refer to the approach that incorporate BSV and full
covariate distributions presented in another vignette.
fpdata <- out.df.univariatecov.nca[,
setNames(as.list(quantile(paramvaluestd, probs=c(0.5, 0.05, 0.95))), c("mid", "lower", "upper")),
by=.(paramname2, covname2, covvalue)]
bsvranges[paramname=="AUC", paramname2 := "AUC"]
bsvranges[paramname=="Clast", paramname2 := "C[last]"]
bsvranges[paramname=="Cmax", paramname2 := "C[max]"]
setkey(bsvranges, paramname2)
fpdataBSV50 <- fpdata[covname2 == "Reference"]
fpdataBSV50$covname2 <- "BSV"
fpdataBSV50$covvalue <- "50% of patients"
setkey(fpdataBSV50, paramname2)
fpdataBSV50$lower <- bsvranges[,"P25"]
fpdataBSV50$upper <- bsvranges[,"P75"]
fpdataBSV90 <- fpdata[covname2 == "Reference"]
fpdataBSV90$covname2 <- "BSV"
fpdataBSV90$covvalue <- "90% of patients"
setkey(fpdataBSV90, paramname2)
fpdataBSV90$lower <- bsvranges[,"P05"]
fpdataBSV90$upper <- bsvranges[,"P95"]
fpdata <- rbind(fpdata, fpdataBSV90, fpdataBSV50)
fpdata[, LABEL := sprintf("%s [%s, %s]",
round_pad(mid, 2),
round_pad(lower, 2),
round_pad(upper, 2)) ]
setnames(fpdata, "paramname2", "paramname")
setnames(fpdata, "covname2", "covname")
setnames(fpdata, "covvalue", "label")
fpdata[, label := factor(label, levels=unique(label))]
interval_legend_text <- "Median (points)\n90% CI (horizontal lines)"
interval_bsv_text <- "BSV (points)\nPrediction Intervals (horizontal lines)"
ref_legend_text <- "Reference (vertical line)\nClinically relevant limits\n(gray area)"
area_legend_text <- "Reference (vertical line)\nClinically relevant limits\n(gray area)"
png("./Figure4_6.png",width =9 ,height = 6,units = "in",res=72)
coveffectsplot::forest_plot(fpdata[paramname=="AUC" &
covname!="Reference" &
covname!="BSV",],
ref_area = c(0.8, 1/0.8),
x_range = c(0.5, 2),
strip_placement = "inside",
base_size = 18,
y_label_text_size = 12,
y_facet_text_angle = 0,
xlabel = "Fold Change Relative to Reference",
ref_legend_text = ref_legend_text,
area_legend_text = area_legend_text,
interval_legend_text = interval_legend_text,
interval_bsv_text = interval_bsv_text,
plot_title = "",
facet_formula = "covname ~ paramname",
facet_switch = "y",
facet_scales = "free_y",
facet_space = "free",
paramname_shape = FALSE,
table_position = "right",
table_text_size = 4,
plot_table_ratio = 3,
show_table_facet_strip = "none",
logxscale = TRUE,
major_x_ticks = c(0.5, 0.8,1, 1/0.8, 1/0.5),
major_x_labels = c("1/2", "0.8","1", "1.25", "2"),
return_list = FALSE)
dev.off()
#> png
#> 2
In this section, we first show a forest_plot
built-in
theme, then how you return the ggplot objects as a list for further
editing using regular ‘ggplot2’ code.
theme_benrich
along Additional OptionsThis is achieved by setting theme_benrich = TRUE
,
specifying that you want no legend
legend_position = "none"
.
With this theme active you can also control the table_title
text and table_title_size
arguments.
png("./coveffectsplot4.png",width =9 ,height = 6,units = "in",res=72)
coveffectsplot::forest_plot(fpdata[paramname=="AUC"],
ref_area = c(0.8, 1/0.8),
x_range = c(0.5,2),
xlabel = "Fold Change Relative to Reference",
x_label_text_size= 10,
y_facet_text_angle = 0,
facet_formula = "covname~paramname",
theme_benrich = TRUE,
table_title_size = 15,
table_title = "Median [90% CI]",
interval_legend_text = interval_legend_text,
interval_bsv_text = interval_bsv_text,
legend_position = "none",
strip_placement = "outside",
base_size = 12,
facet_switch = "y",
facet_scales = "free_y",
facet_space = "free",
paramname_shape = FALSE,
table_position = "right",
table_text_size=4,
plot_table_ratio = 3,
show_table_facet_strip = "none",
logxscale = TRUE,
major_x_ticks = c(0.25,0.5,0.8,1,1/0.8,1/0.5,1/0.25),
major_x_labels = c("1/4","1/2","0.8","1","1.25","2","4"),
return_list = FALSE)
dev.off()
#> png
#> 2
You can get the underlying ggplots as a list for further editing by
setting return_list = TRUE
and saving it into an object.
The list will contain two objects the first being the main plot and the
second the table. We illustrate how you can modify the look of the plots
using regular ggplot code that modify the facet text color to
gray
and italic. Finally we recombine the plots using
egg::ggarrange
.
png("./coveffectsplot0.png",width = 9 ,height = 6,units = "in",res=72)
plotlists <- coveffectsplot::forest_plot(fpdata[paramname=="AUC"],
ref_area = c(0.8, 1/0.8),
xlabel = "Fold Change Relative to Reference",
ref_legend_text = "Reference (vertical line)\nClinically
relevant limits\n(gray area)",
area_legend_text = "Reference (vertical line)\nClinically
relevant limits\n(gray area)",
interval_legend_text = interval_legend_text,
plot_title = "",
interval_bsv_text = interval_bsv_text,
facet_formula = "covname~paramname",
facet_switch = "y",
facet_scales = "free_y",
facet_space = "free",
paramname_shape = FALSE,
table_position = "right",
table_text_size=4,
plot_table_ratio = 4,
show_table_facet_strip = "none",
logxscale = TRUE,
major_x_ticks = c(0.25,0.5,0.8,1,1/0.8,1/0.5,1/0.25),
major_x_labels = c("1/4","1/2","0.8","1","1.25","2","4"),
return_list = TRUE)
plotlists
#> [[1]]
#>
#> [[2]]
dev.off()
#> png
#> 2
main_plot <- plotlists[[1]] + theme(
panel.spacing=unit(10, "pt"),
panel.grid=element_blank(),
panel.grid.minor=element_blank(),
legend.position="bottom",
strip.placement.y="outside",
strip.background.y=element_blank(),
strip.text.y=element_text(
hjust=1,
vjust=1,
face="italic",color="gray",
size=rel(1)),
legend.text = element_text(size=rel(0.5)),
plot.margin = margin(t=0,r=0,b=0,l=5,unit="pt")) +
scale_y_discrete(
breaks=c("90% of patients",
"50% of patients",
"85 kg\nFemale\n45 g/L",
"40 g/L","50 g/L","Male",
"56 kg","72 kg","98 kg","128 kg"
),
labels=c("90% of patients",
"50% of patients",
"85 kg-Female-45 g/L",
"40 g/L","50 g/L","Male",
"56 kg","72 kg","98 kg","128 kg"
)
)
table_plot <- plotlists[[2]] + theme(
panel.border=element_blank(),
panel.spacing=unit(10, "pt"),
strip.background.y=element_blank(),
legend.text = element_text(size=rel(0.5)),
plot.margin = margin(t=0,r=5,b=0,l=0,unit="pt"))
png("./coveffectsplot5.png",width =8.5 ,height = 6,units = "in",res=72)
egg::ggarrange(
main_plot,
table_plot,
nrow = 1,
widths = c(3, 1)
)
dev.off()
#> png
#> 2
You can also have plots with more than one PK parameter. You may want to facet by parameter, or to use different shape by parameter.
This is achieved by setting paramname_shape = FALSE
and
facet_formula = "covname~paramname"
. We also suppress the
table by using table_position = "none"
and reduce the plot
text sizes using base_size = 11
.
png("./coveffectsplot6.png",width =9.5 ,height = 6,units = "in",res=72)
forest_plot(fpdata,
ref_area = c(0.8, 1/0.8),
x_range = c(0.5,2),
xlabel = "Fold Change Relative to Reference",
facet_formula = "covname~paramname",
interval_legend_text = interval_legend_text,
interval_bsv_text = interval_bsv_text,
facet_switch = "y",
facet_scales = "free_y",
facet_space = "free",
facet_labeller = "label_parsed",
paramname_shape = FALSE,
table_position = "none",
table_text_size=4,
base_size = 11,
plot_table_ratio = 4,
show_table_facet_strip = "none",
logxscale = TRUE,
major_x_ticks = c(0.5,0.8,1,1/0.8,1/0.5),
major_x_labels = c("1/2","0.8","1","1.25","2"),
x_label_text_size = 10,
return_list = FALSE)
dev.off()
#> png
#> 2
This is achieved by setting paramname_shape = TRUE
we
also illustrate how you can use legend_order
to choose the
legend ordering and few other options.
png("./coveffectsplot7.png",width =9.5 ,height = 6,units = "in",res=72)
forest_plot(fpdata[paramname!="AUC"],
ref_area = c(0.8, 1/0.8),
x_range = c(0.35,1/0.35),
xlabel = "Fold Change Relative to Reference",
ref_legend_text = "Reference\nClinically relevant limits\n(0.8-1.25)",
area_legend_text = "Reference\nClinically relevant limits\n(0.8-1.25)",
interval_legend_text = "Median\n90% CI",
interval_bsv_text = "BSV\nPrediction Intervals",
facet_formula = "covname~.",
paramname_shape = TRUE,
legend_order =c("shape","pointinterval","ref", "area"),
legend_shape_reverse = TRUE,
bsv_col = scales::muted("red"),
interval_col = scales::muted("blue"),
facet_switch = "y",
facet_scales = "free_y",
facet_space = "free",
table_position = "none",
table_text_size=4,
base_size = 9,
plot_table_ratio = 4,
show_table_facet_strip = "none",
logxscale = TRUE,
major_x_ticks = c(0.5,0.8,1,1/0.8,1/0.5),
major_x_labels = c("1/2","0.8","1","1.25","2"),
legend_space_x_mult = 0.01,
legend_position = "right",
return_list = FALSE)
dev.off()
#> png
#> 2
This is achieved by setting paramname_color = TRUE
(on
top of paramname_shape = TRUE
or not) and the the user can
then pass a vector of colors in the interval_col
argument
png("./coveffectsplot_color.png",width =9.5 ,height = 6,units = "in",res=72)
forest_plot(fpdata[paramname!="AUC" &
covname!="BSV"&
covname!="Reference",],
ref_area = c(0.8, 1/0.8),
x_range = c(0.35,1/0.35),
xlabel = "Fold Change Relative to Reference",
ref_legend_text = "Reference\nClinically relevant limits\n(0.8-1.25)",
area_legend_text = "Reference\nClinically relevant limits\n(0.8-1.25)",
facet_formula = "covname~.",
paramname_shape = TRUE,
paramname_color = TRUE,
combine_interval_shape_legend = TRUE,
legend_order =c("shape","pointinterval","ref", "area"),
legend_shape_reverse = TRUE,
bsv_col = scales::muted("red"),
facet_switch = "y",
facet_scales = "free_y",
facet_space = "free",
table_position = "none",
base_size = 12,
logxscale = TRUE,
major_x_ticks = c(0.5,0.8,1,1/0.8,1/0.5),
major_x_labels = c("1/2","0.8","1","1.25","2"),
legend_space_x_mult = 0.01,
legend_position = "right",
return_list = TRUE)[[1]]+
scale_color_manual(labels = c(expression(C[last]),expression(C[max])),
values = c(scales::muted("blue"),scales::muted("red")))+
scale_shape_discrete(labels = c(expression(C[last]),expression(C[max])))
dev.off()
#> png
#> 2
While we covered varying one at a time covariate value (marginal effects), we can use observed or simulated distribution of correlated covariates and simulate joint covariate effects as illustrated in the Pediatric Application vignette.It can be misleading to show the BSV around the reference only as it operates at all levels of covariate as such we recommend to fully represent the intervals with BSV and Uncertainty at all levels / covariate splits (full Distribution PK example vignette.)