A genomics biobank holds hundreds of thousands of variables and hundreds of phenotypes. Running the full T-Rex calibration on every phenotype is costly, and for most phenotypes the answer is that nothing is selected. T-Rex+Screening [1] is a single-pass version of the random experiments without the L-loop and the T-loop: \(K\) experiments with \(p\) dummies each, stopped at the first dummy, and an estimate of the false discovery rate (FDR) read off the vote. The estimate is one over the number of selected variables, and it is 1 when nothing is selected. The screening reports this estimated FDR instead of controlling a target, so it is a fast first pass. The biobank screening wraps it in a decision rule per phenotype: when the estimated FDR of the screening falls into an acceptance window, the screening result is kept, otherwise the full T-Rex selector runs as a fallback.
The screening selector takes no target FDR.
set.seed(2024)
n <- 100
p <- 300
true_support <- c(3, 50, 120, 200, 280)
beta <- numeric(p)
beta[true_support] <- 2
X <- matrix(rnorm(n * p), n, p)
y <- as.numeric(X %*% beta + rnorm(n))scr <- TRexScreeningSelector$new(X, y, seed = 3, verbose = FALSE,
control = trex_control(K = 20))
scr$select()
scr$selected_indices
#> [1] 3 25 50 120 200 231 280
scr$estimated_FDR
#> [1] 0.1428571beta_mat holds the coefficients of the original
variables at the stop of each experiment (one column per experiment) and
dummy_betas those of the dummies, the material the estimate
is read off. The dummies are fresh per experiment under the default
L-loop strategy "STANDARD".
use_bootstrap_CI = TRUE selects by the coefficients
instead of the vote. The non-zero dummy coefficients of the experiments
are resampled R_boot times, at least twice, for a
confidence interval of their mean, and a variable is selected when its
averaged coefficient lies outside that interval.
confidence_level is the smallest level at which the
selection is no larger than the ordinary one, searched on a grid of step
ci_grid_step in \((0,
1]\). The estimated FDR is again one over the number of selected
variables. The confidence-based screening is meant for the plain
screening. With a dependency-aware variant it mostly selects nothing,
because the filter lowers the ordinary count that the confidence level
has to reach.
boot <- TRexScreeningSelector$new(X, y, seed = 3, verbose = FALSE,
screen_control = trex_screen_control(use_bootstrap_CI = TRUE,
R_boot = 200),
control = trex_control(K = 20))
boot$select()
boot$used_bootstrap
#> [1] TRUE
boot$confidence_level
#> [1] 0.389
boot$estimated_FDR
#> [1] 0.1428571
boot$selected_indices
#> [1] 3 25 50 120 200 231 280trex_method makes the screening dependency-aware. The
dummies stay the same independent dummies, and the variants lower the
relative occurrence of a variable that resembles its correlated
neighbours. "TREX_DA_AR1" compares a variable with the
neighbours inside a window, which holds the lags whose correlation
abs(cor_coef)^lag reaches rho_thr_DA.
"TREX_DA_NN" keeps the neighbours of that window whose
sample correlation reaches rho_thr_DA, and a variable
without such a neighbour keeps its relative occurrence. With a threshold
at or above the correlation of the data no neighbour is left, and
"TREX_DA_NN" selects as "TREX", so the
threshold stays below the correlation that the filter is meant to guard
against. At the default the variant selected as
"TREX_DA_AR1" in simulations. "TREX_DA_EQUI"
compares a variable with all other variables and
"TREX_DA_BLOCK_EQUI" with the variables of its block out of
n_blocks, and both are skipped when
abs(cor_coef) does not exceed rho_thr_DA.
Setting cor_coef to -2, the default, estimates
the coefficient from the data, and estimated_correlation
reports the coefficient that was used.
The screening stops at the first dummy, so about one null variable is expected to enter before it. On a response of pure noise it selected at least one variable in a quarter to a half of the runs in simulations, with fewer than one selected variable on average. A small selection is therefore no evidence of a signal. The estimated FDR rests on that expected number of false discoveries and is no bound on the FDP of a run. With two or more variables selected on noise it is 0.5 or less, while every one of them is false. The biobank screening below accepts a screening result only when the estimate lies in its window and runs the T-Rex selector otherwise, which selected nothing on noise in about 99 % of the runs.
TRexBiobankScreeningSelector screens one or many
phenotypes against one genotype matrix. Y is a vector for
one phenotype or a matrix with one column per phenotype. The control is
trex_biobank_control(): the acceptance window of the
estimated FDR is lower_bound_FDR to
upper_bound_FDR, and target_FDR_trex is the
target of the fallback selector. The phenotypes draw independent
dummies, because the phenotype in column k runs with
seed + k - 1, and a single-phenotype run with that seed
reproduces it.
Y <- cbind(y, as.numeric(X %*% (0.5 * beta) + rnorm(n)), rnorm(n))
colnames(Y) <- c("strong", "weak", "null")
bio <- TRexBiobankScreeningSelector$new(X, Y, seed = 3,
biobank_control = trex_biobank_control(0.05, 0.15,
target_FDR_trex = 0.1),
control = trex_control(K = 20))
res <- bio$select()
length(res)
#> [1] 3Each record names the phenotype, the method that produced the result
and whether the fallback ran. For a screening result
estimated_FDR is the estimate of the accepted variant. For
a fallback result it is the target level target_FDR_trex,
at which the T-Rex selector controls the FDR. An estimate can equal the
target, so used_fallback_trex tells the two cases
apart.
summary_tab <- do.call(rbind, lapply(res, function(r)
data.frame(phenotype = r$phenotype_index, selected = length(r$selected_indices),
estimated_FDR = round(r$estimated_FDR, 3), method = r$method_used,
fallback = r$used_fallback_trex)))
summary_tab
#> phenotype selected estimated_FDR method fallback
#> 1 1 7 0.143 Screen-TRex (bootstrap-CI) FALSE
#> 2 2 5 0.100 T-Rex (fallback) TRUE
#> 3 3 0 0.100 T-Rex (fallback) TRUE
res[[1]]$selected_indices
#> [1] 3 25 50 120 200 231 280The record also keeps the ordinary and the bootstrap screening
results (selected_indices_screen_ordinary,
estimated_FDR_screen_bootstrap, …), so the decision can be
audited. For a single phenotype select() returns one record
instead of a list.
Both screening selectors accept a memory-mapped design
(mem_matrix()), which the memory-mapping page describes,
and the biobank screening a memory-mapped response matrix
Y, which it only reads. The combination of the screening
and a memory-mapped design is what solved sparse regression problems
with five million variables on a laptop [2].
The TRexSelector_Examples repository holds runnable demos and Monte Carlo studies of T-Rex+Screening and the biobank screening, in memory and memory-mapped, with their results and a README per demo, in R, Python and C++.