diagFDR: generic PSM diagnostics from mzIdentML (.mzid)

This vignette demonstrates how to compute diagFDR diagnostics from a generic mzIdentML file (.mzid).

Many search engines can export identifications in mzIdentML format. When explicit q-values and/or PEPs are not present in the mzIdentML, diagFDR can:

Runnable toy example (no external files required)

This section provides a small toy dataset that mimics the PSM-level competed universe returned by read_dfdr_mzid() after choosing a score and reconstructing TDC q-values.

library(diagFDR)

set.seed(3)

n <- 7000
pi_decoy <- 0.03

# Decoy indicator
is_decoy <- sample(c(FALSE, TRUE), n, replace = TRUE, prob = c(1 - pi_decoy, pi_decoy))

# Targets are a mixture: some null-like (close to decoys), some true (higher score)
# This yields realistic separation and non-trivial discoveries at 1% FDR.
is_true_target <- (!is_decoy) & (runif(n) < 0.30)  # 30% of targets are "true"
is_null_target <- (!is_decoy) & (!is_true_target)

score <- numeric(n)
score[is_decoy]       <- rnorm(sum(is_decoy), mean = 0.0, sd = 1.0)
score[is_null_target] <- rnorm(sum(is_null_target), mean = 0.2, sd = 1.0)
score[is_true_target] <- rnorm(sum(is_true_target), mean = 3.0, sd = 1.0)

toy <- data.frame(
  id = paste0("psm", seq_len(n)),
  is_decoy = is_decoy,
  run = sample(paste0("run", 1:4), n, replace = TRUE),
  score = score,
  pep = NA_real_
)

# Score-based TDC q-values (higher score is better)
toy <- toy[order(toy$score, decreasing = TRUE), ]
toy$D_cum <- cumsum(toy$is_decoy)
toy$T_cum <- cumsum(!toy$is_decoy)
toy$FDR_hat <- (toy$D_cum + 1) / pmax(toy$T_cum, 1)
toy$q <- rev(cummin(rev(toy$FDR_hat)))
toy <- toy[, c("id","is_decoy","q","pep","run","score")]

x_toy <- as_dfdr_tbl(
  toy,
  unit = "psm",
  scope = "global",
  q_source = "toy TDC from score",
  q_max_export = 1,
  provenance = list(tool = "toy")
)

diag <- dfdr_run_all(
  xs = list(mzid_PSM = x_toy),
  alpha_main = 0.01,
  alphas = c(1e-4, 5e-4, 1e-3, 2e-3, 5e-3, 1e-2, 2e-2, 5e-2, 1e-1, 2e-1),
  eps = 0.2,
  win_rel = 0.2,
  truncation = "warn_drop",
  low_conf = c(0.2, 0.5)
)

Headline stability at 1%

diag$tables$headline
#> # A tibble: 1 × 24
#>   alpha T_alpha D_alpha FDR_hat CV_hat FDR_minus1 FDR_plus1 FDR_minusK FDR_plusK
#>   <dbl>   <int>   <int>   <dbl>  <dbl>      <dbl>     <dbl>      <dbl>     <dbl>
#> 1  0.01    2819      27 0.00958  0.192    0.00922   0.00993    0.00603    0.0131
#> # ℹ 15 more variables: k2sqrtD <int>, FDR_minus2sqrtD <dbl>,
#> #   FDR_plus2sqrtD <dbl>, list <chr>, D_alpha_win <int>, effect_abs <dbl>,
#> #   IPE <dbl>, flag_Dalpha <chr>, flag_CV <chr>, flag_Dwin <chr>,
#> #   flag_IPE <chr>, flag_FDR <chr>, flag_equalchance <chr>, status <chr>,
#> #   interpretation <chr>

if (nrow(diag$tables$headline) > 0 && diag$tables$headline$T_alpha[1] == 0) {
  cat("\nNote: No discoveries at alpha_main for this toy run. ",
      "For demonstration, consider using alpha_main = 0.02.\n", sep = "")
}

Tail support and stability versus threshold

diag$plots$dalpha

diag$plots$cv

Local boundary support and elasticity

diag$plots$dwin

diag$plots$elasticity
#> Warning: Removed 3 rows containing missing values or values outside the scale range
#> (`geom_line()`).
#> Warning: Removed 3 rows containing missing values or values outside the scale range
#> (`geom_point()`).

Equal-chance plausibility by q-band

diag$tables$equal_chance_pooled
#> # A tibble: 1 × 12
#>   qmax_export low_lo low_hi N_test N_D_test pi_D_hat effect_abs ci95_lo ci95_hi
#>         <dbl>  <dbl>  <dbl>  <int>    <int>    <dbl>      <dbl>   <dbl>   <dbl>
#> 1           1    0.2    0.5      0        0       NA         NA      NA      NA
#> # ℹ 3 more variables: p_value_binom <dbl>, pass_minN <lgl>, list <chr>
diag$plots$equal_chance__mzid_PSM

Real mzIdentML workflow (.mzid)

The code below shows how to run diagFDR directly from a .mzid file.

library(diagFDR)

mzid_path <- "path/to/search_results.mzid"

x_mzid <- read_dfdr_mzid(
  mzid_path = mzid_path,
  rank = 1L,  # competed universe: take rank-1 SpectrumIdentificationItem

  # Choose a score CV term (priority list) and interpret its direction
  score_accession_preference = c(
    "MS:1002257", # MS-GF:RawScore (higher is better)
    "MS:1001330", # Mascot:score (higher is better)
    "MS:1001328", # SEQUEST:xcorr (higher is better)
    "MS:1002052",
    "MS:1002049",
    "MS:1001331",
    "MS:1001171",
    "MS:1001950",
    "MS:1002466"
  ),
  score_direction = "auto",  # or "higher_better"/"lower_better" if auto fails

  # TDC correction: FDR_hat = (D + add_decoy)/T
  add_decoy = 1L,

  # Strict by default: require score for all PSMs (set <1 to allow missing)
  min_score_coverage = 1.0,

  # Fallback decoy inference if PeptideEvidence@isDecoy is not informative
  decoy_regex = "(^##|_REVERSED$|^REV_|^DECOY_)",

  unit = "psm",
  scope = "global",
  provenance = list(file = basename(mzid_path))
)

diag <- dfdr_run_all(
  xs = list(mzid_PSM = x_mzid),
  alpha_main = 0.01,
  alphas = c(1e-4, 5e-4, 1e-3, 2e-3, 5e-3, 1e-2, 2e-2, 5e-2, 1e-1, 2e-1),
  eps = 0.2,
  win_rel = 0.2,
  truncation = "warn_drop",
  low_conf = c(0.2, 0.5)
)

# Export outputs
dfdr_write_report(
  diag,
  out_dir = "diagFDR_mzid_out",
  formats = c("csv", "png", "manifest", "readme", "summary")
)

# Render a single HTML report (requires rmarkdown in Suggests)
dfdr_render_report(diag, out_dir = "diagFDR_mzid_out")

Optional: pseudo-p-values from the decoy score tail

If you want to run the p-value diagnostics (calibration plots, Storey pi0, BH stability), you can map scores to pseudo-p-values using the empirical decoy null tail.

x_mzid$p <- score_to_pvalue(
  score = x_mzid$score,
  method = "decoy_ecdf",
  is_decoy = x_mzid$is_decoy
)
attr(x_mzid, "meta")$p_source <- "score_to_pvalue(method='decoy_ecdf' on mzid score)"

diag_p <- dfdr_run_all(
  xs = list(mzid_PSM = x_mzid),
  alpha_main = 0.01
)

dfdr_write_report(
  diag_p,
  out_dir = "diagFDR_mzid_out_with_p",
  formats = c("csv","png","manifest","readme","summary")
)
dfdr_render_report(diag_p, out_dir = "diagFDR_mzid_out_with_p")

Interpretation notes / common pitfalls

Score selection matters: mzIdentML files can contain multiple score CV terms. If read_dfdr_mzid() picks an unexpected score, adjust score_accession_preference and/or set score_direction explicitly.
Decoy labeling matters: ideally, mzIdentML includes PeptideEvidence@isDecoy. If not, read_dfdr_mzid() falls back to decoy_regex applied to protein accessions. If your pipeline uses a different decoy naming convention, update decoy_regex.
Missing scores: if some PSMs are missing the chosen score CV term, you can relax min_score_coverage (e.g., 0.95) but then interpret results carefully, as the hypothesis universe changes.
Scope: this vignette constructs a PSM-level universe. Peptide- or protein-level claims require additional aggregation/inference.