Package 'MIML'

Metastasis-free survival panel (GSE65622)

Description

The same patients, visits and 507 biomarkers as RFS_data, paired instead with metastasis-free survival. The default input to miml_mfs().

Usage

MFS_data

Format

A data frame of 288 rows and 516 columns, identical in structure to RFS_data except that the survival column is MFS and event is the metastasis indicator. Twenty-six of the 80 patients record a metastasis.

Source

Gene Expression Omnibus accession GSE65622, released April 2016.

See Also

RFS_data, miml_mfs()

Examples

data(MFS_data)
dim(MFS_data)
MFS_data[1:5, 1:6]

---

Metastasis-free survival analysis of the GSE65622 panel

Description

Identical to miml_rfs() but run on MFS_data against metastasis-free survival. Twenty-six of the eighty patients record a metastasis, against five recurrences for the other endpoint, so this is the better powered of the two comparisons.

Usage

miml_mfs(
  data = NULL,
  m = 5L,
  maxit = 10L,
  nrounds = 200L,
  n_clusters = 2L,
  n_bcc_markers = 3L,
  predictors = c("manuscript", "all"),
  seed = 123L,
  verbose = TRUE
)

Arguments

data	Panel to analyse. Defaults to MFS_data.
m	Number of completed datasets. The published analysis used 5.
maxit	Refit cycles per completion. The published analysis used 10.
nrounds	Maximum boosting rounds. The published analysis used 200.
n_clusters	Number of consensus clusters. The published analysis used 2.
n_bcc_markers	Biomarkers carried into the clustering model, chosen by descending variance. The published analysis used 3.
predictors	Predictor pool for the imputation models. "manuscript" uses the patient identifier, the survival endpoint, the event indicator and the six phenotype columns, which is what the published analysis did; "all" lets biomarkers predict each other and drops the outcome from the pool. See the leakage note below.
seed	Random seed. The caller's RNG state is restored on exit.
verbose	Print progress.

Value

An object of class miml_analysis, as for miml_rfs().

Predictor pool and outcome leakage

Under predictors = "manuscript" the survival time and the event indicator are inside the predictor pool, and no biomarker predicts another. Any association later estimated between the imputed values and survival is therefore inflated by construction. The default reproduces the published numbers; use predictors = "all" for new work.

Run time

At the defaults the imputation fits 507 biomarkers over 10 cycles for each of 5 completions and takes roughly an hour.

See Also

miml_rfs(), MFS_data

Examples

data(MFS_data)

small <- MFS_data[1:60, c("Patient_ID", "MFS", "event",
                          names(MFS_data)[4:6])]
miml_mfs(small, m = 1, maxit = 1, nrounds = 3,
         n_clusters = 0, verbose = FALSE)


sub <- MFS_data[, c("Patient_ID", "MFS", "event", names(MFS_data)[4:23])]
res <- miml_mfs(sub, m = 1, maxit = 2, nrounds = 20)
res

## the published analysis is miml_mfs() at its defaults.

---

Recurrence-free survival analysis of the GSE65622 panel

Description

Runs the published MIML workflow on RFS_data: LightGBM imputation of every incomplete biomarker, Bayesian consensus clustering of the completed panel, and a Kaplan-Meier, log-rank and Cox comparison of the resulting clusters against recurrence-free survival.

Usage

miml_rfs(
  data = NULL,
  m = 5L,
  maxit = 10L,
  nrounds = 200L,
  n_clusters = 2L,
  n_bcc_markers = 3L,
  predictors = c("manuscript", "all"),
  seed = 123L,
  verbose = TRUE
)

Arguments

data	Panel to analyse. Defaults to RFS_data.
m	Number of completed datasets. The published analysis used 5.
maxit	Refit cycles per completion. The published analysis used 10.
nrounds	Maximum boosting rounds. The published analysis used 200.
n_clusters	Number of consensus clusters. The published analysis used 2.
n_bcc_markers	Biomarkers carried into the clustering model, chosen by descending variance. The published analysis used 3.
predictors	Predictor pool for the imputation models. "manuscript" uses the patient identifier, the survival endpoint, the event indicator and the six phenotype columns, which is what the published analysis did; "all" lets biomarkers predict each other and drops the outcome from the pool. See the leakage note below.
seed	Random seed. The caller's RNG state is restored on exit.
verbose	Print progress.

Details

Imputed values are the conditional-mean predictions of the fitted LightGBM models. No residual is drawn and no donor is sampled, so the completions carry no between-imputation variance and must not be combined by Rubin's rules. This is machine-learning single imputation.

Value

An object of class miml_analysis: a list with endpoint, missingness, completed, clusters, survival and settings.

Predictor pool and outcome leakage

Under predictors = "manuscript" the survival time and the event indicator are inside the predictor pool, and no biomarker predicts another. Any association later estimated between the imputed values and survival is therefore inflated by construction. The default reproduces the published numbers; use predictors = "all" for new work.

Run time

At the defaults the imputation fits 507 biomarkers over 10 cycles for each of 5 completions and takes roughly an hour.

See Also

miml_mfs(), RFS_data

Examples

data(RFS_data)

## small and fast: three biomarkers, one cycle, no clustering
small <- RFS_data[1:60, c("Patient_ID", "RFS", "event",
                          names(RFS_data)[4:6])]
miml_rfs(small, m = 1, maxit = 1, nrounds = 3,
         n_clusters = 0, verbose = FALSE)


## twenty biomarkers, two cycles, clustered and compared by survival
sub <- RFS_data[, c("Patient_ID", "RFS", "event", names(RFS_data)[4:23])]
res <- miml_rfs(sub, m = 1, maxit = 2, nrounds = 20)
res

## the published analysis is miml_rfs() at its defaults: 507 biomarkers,
## 10 cycles, 5 completions, roughly an hour.

---

Recurrence-free survival panel (GSE65622)

Description

Longitudinal serum proteomic measurements from patients with locally advanced rectal cancer undergoing multimodal treatment, paired with recurrence-free survival. The default input to miml_rfs().

Usage

RFS_data

Format

A data frame of 288 rows and 516 columns: 80 patients observed at three or four sampling points each. Patient_ID is an integer identifier repeated across a patient's visits; RFS is recurrence-free survival in days measured from that visit, so it decreases across a patient's rows; event is the recurrence indicator, constant within a patient. Then 507 protein biomarkers named as in the original submission, then six phenotype columns: diarrhea_grade, sampling_point, slide_print_batch, trg_score, ypn_stage, ypt_stage.

Biomarker names are not syntactic R names, so quote them or use [[ ]]. About 10.3 percent of the biomarker measurements are missing. Five of the 80 patients record a recurrence.

Source

Gene Expression Omnibus accession GSE65622, released April 2016.

See Also

MFS_data, miml_rfs()

Examples

data(RFS_data)
dim(RFS_data)
RFS_data[1:5, 1:6]

---