Aster: Adaptive Statistical framework for Therapeutic Efficacy and Recrudescence

Package asterTES performs recurrence classification and failure rate estimation for therapeutic efficacy studies (TES). To install it from GitHub:

# install.packages("remotes")
remotes::install_github("EPPIcenter/asterTES")

Data

As an example, we use a simulated study that could be conducted in a high transmission setting. Consider TES data that are stored in three data frames: dfppl (where rows represent individuals), dfsmp (rows represent genotyped samples), and dfgen (genotyping data, rows represent alleles).

library(asterTES)
head(dfppl, 3)
#>   id lastday outcome
#> 1 A1      NA       0
#> 2 A2      NA       0
#> 3 A3      28       1
head(dfsmp, 4)
#>   id smp_id day parasite_dens
#> 1 A1  A1-D0   0  5.190163e+04
#> 2 A2  A2-D0   0  2.024178e+04
#> 3 A3  A3-D0   0  1.501553e+05
#> 4 A3 A3-D28  28  4.521802e+00
head(dfgen, 3)
#>   smp_id locus allele
#> 1  A1-D0  loc1     t3
#> 2  A1-D0  loc1     t5
#> 3  A1-D0  loc1     t7

The outcome in dfppl is coded as 0 for successful treatment, 1 for recurrence, and 2 for loss to follow-up (censoring).

Prep

To analyze these data, we first need to estimate COI, population allele frequencies, and a level of relatedness between separate infections in the population (background relatedness). Various methods can be used; to illustrate, we use a package dcifer:

# install.packages("dcifer")
# Stay tuned for upcoming dcifer updates! Changes may affect the code below
library(dcifer)

For estimation of allele frequencies and background relatedness, we use Day 0 samples only:

dsmp <- formatDat(dfgen, svar = "smp_id", lvar = "locus", avar = "allele")
dsmp <- dsmp[dfsmp$smp_id]                     # same order as dfsmp
coi  <- getCOI(dsmp, lrank = 2)

# D0 only to calculate afreq and rbg
ismp0 <- dfsmp$day == 0
afreq <- calcAfreq(dsmp[ismp0], coi[ismp0], tol = 1e-5)
minfr <- 1/sum(coi)
afreq <- lapply(afreq, function(af, minfr) {af[af == 0] <- minfr; af/sum(af)},
                minfr = minfr)                 # fill in 0-frequencies
rhat <- ibdDat(dsmp[ismp0], coi[ismp0], afreq, pval = FALSE)
rbg  <- mean(rhat, na.rm = TRUE)

Next, we estimate detection probabilities. Here, we use a simple example where the probability of detecting a strain at a locus depends on discretized parasite density:

pdbreaks <- c(0, 10, 1000, 10000, Inf)
pmiss    <- c(0.5, 0.2, 0.1, 0.05)
pd2pdet <- function(x, breaks, labs) {
  y <- cut(x, breaks, labels = labs, include.lowest = FALSE)
  return(1 - as.numeric(as.character(y)))
}

dfsmp$pdet <- pd2pdet(dfsmp$parasite_dens, pdbreaks, pmiss)

Process

To obtain individual recurrence classification and conditional probabilities of genotyping data given recrudescence, a pair of samples for the same individual is needed - the Day 0 sample and the recurrent sample. For example, to classify a recurrent infection from an individual with ID "A3":

iid <- which(dfsmp$id == "A3")
i0 <- iid[1]; ir <- iid[2]
(res1 <- A0A1(dsmp[[i0]], dsmp[[ir]], coi[i0], coi[ir], afreq, dfsmp$pdet[i0],
              dfsmp$pdet[ir], rbg))
#> $recrud
#> [1] TRUE
#> 
#> $cond_prob
#> [1] -132.366 -127.106

We can also process the whole dataset and estimate the therapeutic failure rate along with classification for each pair and corresponding posterior probability of recrudescence. Two such probabilities are provided for each recurrence: one with a user-specified prior probability and another using an empirical Bayes approach. Classification can be updated based on the latter (using a 0.5 cut-off). The failure rate is calculated using the Kaplan-Meier survival estimator:

res <- asterKM(dsmp, dfppl, dfsmp, coi, afreq, rbg, pprior = 0.5, 
               update_est = TRUE)
res$failure_rate
#> [1] 0.1447888
head(res$classification)
#>   id lastday outcome recrud     logA0     logA1 ppost_user ppost_empB
#> 1 A1      NA       0     NA        NA        NA         NA         NA
#> 2 A2      NA       0     NA        NA        NA         NA         NA
#> 3 A3      28       1   TRUE -132.3660 -127.1060  0.9948313  0.9768664
#> 4 A4      NA       0     NA        NA        NA         NA         NA
#> 5 A5      28       1   TRUE -132.1396 -127.6724  0.9886500  0.9502748
#> 6 A6      NA       0     NA        NA        NA         NA         NA

The package also provides functions to calculate various probabilities and functions used in the likelihood (for more details, refer to the documentation).