Struct LfqConfig 

pub struct LfqConfig {

    pub mz_ppm: f64,
    pub rt_window_pct: f64,
    pub im_tolerance: f64,
    pub n_isotopes: usize,
    pub grid_cols: usize,
    pub grid_scans_per_column: f64,
    pub peak_max_halfwidth_frac: f64,
    pub min_spectral_bhattacharyya: f64,
    pub score_mode: ScoreMode,
    pub run_tdc: bool,
    pub decoy_mz_shift_da: f64,
    pub decoy_rt_shift_pct: f64,
    pub consensus: ConsensusConfig,
    pub normalize: String,
    pub quant_estimator: String,
    pub detected_use_grid: bool,
    pub rt_spread_scoring: bool,
    pub search_scoring: SearchScoring,
    pub search_rt_rescue: bool,
    pub search_rt_rescue_per_run_transfers: bool,
    pub search_expand_charges: bool,
    pub search_gate_transfers_only: bool,
    pub averagine_projection: bool,
    pub decoy_own_template: bool,
    pub lone_coelution: f64,
    pub isotope_model: IsotopeModelSpec,
}

Configuration for LFQ grid extraction and peak integration.

Fields

mz_ppm: f64

m/z tolerance for hill lookup (ppm)

rt_window_pct: f64

Half-window size as a fraction of the run’s total RT span. 0.005 = ±0.5% of the observed RT span → 1% total window, centred on the alignment-predicted native RT. Default 0.005.

im_tolerance: f64

Ion mobility half-window for hill lookup (absolute 1/K0 units). Default 0.015; inert when the input has no ion mobility.

n_isotopes: usize

Number of isotopologue rows: 1 = M only, 2 = M+M1, 3 = M+M1+M2

grid_cols: usize

Retention-time bins across the extraction window. This should track the run’s MS1 scan density: on a two-hour Orbitrap gradient the default window holds roughly fifteen real scans, so a finer grid spreads them over mostly empty bins and the per-column statistics that drive scoring and peak bounds become noise. Measured on IonStar, dropping from 100 to 25 cuts the paired CV difference against FlashLFQ from 2.98 to 0.76 points and improves identification-free median CV from 16.58% to 14.00%. A faster-scanning instrument supports more bins; deriving this from the observed scan rate instead of a fixed count is the proper fix.

grid_scans_per_column: f64

Target MS1 scans averaged into each retention-time column. When above zero the column count is derived per run from the measured scan spacing as window / (target * spacing), so the amount of smoothing per column is the same on a fast instrument as on a slow one, and grid_cols becomes an upper bound. A column fed by a single scan carries that scan’s noise straight into the isotope and hybrid statistics that choose the peak bounds, which is what a fixed column count produced here: on these gradients 100 columns worked out to 1.01 scans each.

peak_max_halfwidth_frac: f64

Furthest a peak may expand from its apex, as a fraction of the grid’s half-width. This used to be a fixed twenty bins, which silently coupled the integration width to the grid resolution: the same setting meant 0.28 minutes at 100 columns and 0.07 at 400, so raising the resolution truncated peaks. Expressed against the window it means the same span whatever the column count. One allows expansion across the half-window, leaving the isotope and hybrid gates as the operative control.

min_spectral_bhattacharyya: f64

Minimum spectral Bhattacharyya score for a grid column to keep extending the integration peak (peak-expansion gate). Named for the metric it actually uses — it is NOT a cosine threshold. Optional (defaults to 0.1 when omitted); the config was previously named spectral_cosine_min, which now errors (see deny_unknown_fields).

score_mode: ScoreMode

Scoring mode used to find the best integration window

run_tdc: bool

Whether to run target-decoy competition and compute q-values

decoy_mz_shift_da: f64

Decoy m/z shift in Da, added to the target m/z and divided by charge (so the decoy m/z = target_mz + decoy_mz_shift_da / charge). Must be large enough to fall outside any plausible isotopologue or adduct envelope of the target.

decoy_rt_shift_pct: f64

Decoy RT shift as a fraction of the run’s RT span, subtracted from the target RT. The default 0.01 is twice the default extraction half-window, so the target and decoy RT windows meet at one boundary. A shift larger than twice rt_window_pct fully separates their RT intervals.

consensus: ConsensusConfig

Tolerances for building the multi-run consensus feature list.

normalize: String

Cross-run intensity normalization applied to the assembled matrix. “none” = raw intensities (original; downstream must normalize). “median_ratios” = DESeq/edgeR-style size factors: each run is scaled by the median, over features detected in ALL runs, of its log2 deviation from the per-feature mean. Robust to a fraction of genuinely-changing features, unlike a plain column-median (which is biased when a sizeable subset of peptides changes systematically). Default “none”.

quant_estimator: String

Per-cell intensity estimator for the consensus matrix. “apex” = apex intensity (detected: feature intensityApex; MBR: total isotopologue intensity at the apex grid column). Default since 0.10.0: lower fold-change error than “sum” on IonStar and on a held-out cohort (see docs/CONFIGURATION.md §4.2). “sum” = integrated peak area (detected: feature intensitySum; MBR: summed XIC grid over start..=end). The default before 0.10.0.

detected_use_grid: bool

Quantify EVERY consensus cell (detected and MBR) by the same grid re-integration, instead of using the per-run feature’s own intensity for detected cells. Keeps detected and MBR cells on one commensurate scale — essential on timsTOF, where the feature integrates the ion-mobility dimension but the 2-D XIC grid does not, so mixing the two inflates CV, and validated as a win on Orbitrap too (PXD003881: gated matrix CV 14.27 → 12.31 %, ECOLI bias −0.067 → −0.020, HUMAN IQR 0.219 → 0.194). Default TRUE since 2026-08-27: one estimator, one scale, everywhere. Legacy false values now warn and use grid intensities: exclusive native signal ownership cannot validate a feature-intensity override.

rt_spread_scoring: bool

Replace the raw RT closeness term in the hybrid score with a σ-normalised Gaussian RT likelihood exp(−½ z²), where z = (column RT − predicted RT) / σ_local(RT) and σ_local is the post-warp RT-residual spread estimated per run from the RANSAC inlier anchors (see alignment::RtSigmaModel). Makes the RT term of the target-decoy discriminant region-aware: strict where alignment is confident, tolerant where it is not. Applied identically to target and decoy cells so the TDC null stays calibrated. Non-reference runs only; falls back to the raw term when no σ model is available. Default false.

search_scoring: SearchScoring

Experimental search-guided scorer; legacy preserves the release path.

search_rt_rescue: bool

Rescue unambiguous same-run IDs when global RT observations conflict. Ambiguous targets remain ineligible for cross-run transfer.

search_rt_rescue_per_run_transfers: bool

Decide transfer eligibility per recipient run rather than per target. Without this, pruning one run’s ambiguous IDs withdraws the peptide from transfer in every run, including runs whose chromatography was clean. Requires search_rt_rescue; runs whose own IDs were withheld as unsupported stay ineligible, and a target whose retained IDs still disagree across runs remains globally ineligible because its consensus coordinate is what a transfer would be centred on.

search_expand_charges: bool

Experimental inferred 2+/3+/4+ targets, with explicit charge provenance.

search_gate_transfers_only: bool

Apply the extraction q-value gate only to transferred cells. A cell with an accepted same-run MS2 identification has already passed peptide-level FDR; gating it again on an exploratory extraction score re-litigates an identification that was accepted, and discards a quantity for a peptide known to be present in that run. Transfers, which have no such evidence, stay gated. FlashLFQ applies FDR to its match-between-runs peaks only. Identification-free mode is unaffected: every cell there is inferred, so every cell stays gated.

averagine_projection: bool

Report the averagine-projected intensity per cell instead of the raw box-sum. For each grid column the observed isotopologue vector is passed through a matched filter for the theoretical averagine pattern (the same mass-derived fingerprint the Bhattacharyya score already uses): proj = <observed, pattern_hat>, where pattern_hat is the L2-normalised pattern. As a matched filter this maximises SNR for on-pattern signal (recovering faint peptides the raw sum buries in noise) and rejects the component of each column orthogonal to the fingerprint (intensity distributed unlike averagine — interference, chemical noise). It is NOT a pattern-match gate: gross mismatches (lone monoisotopes, wrong charge) are still rejected by the Bhattacharyya window gate and the TDC q-value, unchanged. The per-peptide constant ‖pattern‖ factor cancels in cross-run ratios and CV, so LFQ ratios/CV are unaffected in scale. Apex selection and window expansion stay on the raw signal / hybrid score; only the reported value changes. Default false (byte-identical raw box-sum). Mass-based, ID-free. Effect on the benchmark is unmeasured — this is an opt-in knob to be validated, not a new default.

decoy_own_template: bool

(Experiment B / audit A2) Score each decoy cell against an averagine isotope template and theoretical pattern computed from the DECOY’s own shifted mass (neutral_mass + decoy_mz_shift_da) rather than the target’s mass. When false the decoy reuses the target’s template. When true (the default) the +11 Da decoy is judged against the isotope envelope it actually sits on, removing the target-template freebie a mis-massed decoy otherwise inherits on the Bhattacharyya (QDA feature 2) and the co-elution theory weighting (feature 3). Applies only to the TDC decoy grid; target scoring and all reported intensities are unchanged.

lone_coelution: f64

(Experiment B / audit A4) Co-elution value assigned to a grid cell with fewer than two isotope rows carrying signal (a lone monoisotope — nothing to co-elute). Applied identically to target and decoy cells. Default 0.5 (neutral). 1.0 gives an un-judgeable cell the maximal, target-like score, which hands noise-grabbing lone-hill decoys a free target-like coordinate on QDA feature index 3.

isotope_model: IsotopeModelSpec

Which analyte class’s average composition the LFQ isotope templates are built from. Mirrors [features] isotope_model and must match the model the per-run features were detected with, or the consensus cells are scored against a pattern the detector never used. Default "peptide".

Trait Implementations

impl Clone for LfqConfig

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for LfqConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for LfqConfig

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for LfqConfig

where LfqConfig: Default,

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for LfqConfig

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.