Struct ScoreConfig 

pub struct ScoreConfig {

    pub matrix_joint_agreement: bool,
    pub matrix_joint_cell_quality: bool,
    pub matrix_joint_cell_counts: bool,
    pub matrix_joint_cell_q: bool,
    pub matrix_joint_alternative_groups: bool,
    pub diagnostic_peptides_path: Option<String>,
    pub matrix_recurrence: bool,
    pub top_k: usize,
    pub n_folds: usize,
    pub calibration_fdr: f64,
    pub bootstrap_rounds: usize,
    pub iterations: usize,
    pub max_depth: u32,
    pub shrinkage: f32,
    pub max_train_rows: usize,
    pub seed_min_runs_frac: f64,
    pub seed: u64,
    pub rt_seed: Option<u64>,
    pub cv_seed: Option<u64>,
    pub classifier_seed: Option<u64>,
    pub mass_anchor_ppm: f64,
    pub discriminant: String,
    pub ondata_rt: bool,
    pub ondata_rt_lambda: f64,
    pub ondata_rt_finetune: bool,
    pub mass_anchor: bool,
    pub rt_calibrated: bool,
    pub ml_rescore: bool,
}

Fields

matrix_joint_agreement: bool

Joint-input ablation: use candidate-specific observations but disable cross-run agreement descriptors in the discriminant.

matrix_joint_cell_quality: bool

Experimental row-level MBR/extraction-q descriptors in joint searches.

matrix_joint_cell_counts: bool

Keep fill-count descriptors when cell-quality scoring is enabled.

matrix_joint_cell_q: bool

Keep q-derived descriptors when cell-quality scoring is enabled.

matrix_joint_alternative_groups: bool

Let coherent alternate matrix rows preserve a peptide during competition.

diagnostic_peptides_path: Option<String>

Optional diagnostic export from the selected calibration only.

matrix_recurrence: bool

Ablate only run-recurrence classifier columns; preserve quantities and rows.

top_k: usize

Keep this many best-mass-error targets and decoys per feature. This caps the candidate set the discriminant sees, so setting it too low silently drops real peptides before RT can rescue them. Entrapment (human+yeast on single-run HeLa) showed 25 identified only ~half of what the search honestly supports: 25→105 IDs, 50→208, 100→275, all at ~1% entrapment FDP. 50 is the default (≈2× IDs over 25 at negligible cost); raise toward 100 for maximum depth at a runtime cost.

n_folds: usize

calibration_fdr: f64

FDR used to select the provisional set for RT calibration.

bootstrap_rounds: usize

Number of re-score rounds after the first calibration.

iterations: usize

max_depth: u32

shrinkage: f32

max_train_rows: usize

Cap on training rows per fold-model (0 = no cap). Bounds training time.

seed_min_runs_frac: f64

Seed features must appear in at least this fraction of samples. Lower for many-sample designs (e.g. dose-response) where a peptide need not be in every run; the RANSAC seed tolerates the extra contamination.

seed: u64

Fallback seed for scoring’s random components. 0 = the shipped default streams (see crate::seed_mix). uno --seed sets this value; explicit rt_seed, cv_seed, or classifier_seed values take precedence.

rt_seed: Option<u64>

Override only RANSAC RT-calibration pair sampling. None uses seed.

cv_seed: Option<u64>

Override only cross-validation fold assignment. None uses seed.

classifier_seed: Option<u64>

Override classifier row/column sampling and LightGBM hyperparameter search. Applies to both Sequoia and LightGBM. None uses seed.

mass_anchor_ppm: f64

Mass tolerance (ppm) the mass-anchor bootstrap starts at, and the ceiling it tightens toward as it recalibrates the per-run mass shift + sigma. Default 10 (Orbitrap). For lower-accuracy instruments (e.g. timsTOF) set this wider — e.g. 20 — plus [matching] ppm_tolerance = 20; the anchor then seeds wide and tightens to single-digit ppm once the mass error is recalibrated on-data.

discriminant: String

Discriminant: “deviation” (robust orthogonal RT+IM consistency, ms1searchpy style — default) or “gbdt” (semi-supervised boosted trees). On the MS1-only/saturated-mass regime the gbdt washes out the weak RT+IM signal, so the deviation score is the default.

ondata_rt: bool

Refit the RT predictor on this run’s observed RTs (ridge on composition features) using the confident set, each iteration. The DirectMS1 trick: the frozen DeepLC model only rank-correlates ~0.3 on an unseen gradient, but a model fit to the data makes RT a strong discriminator. Default on.

ondata_rt_lambda: f64

Ridge penalty for the on-data RT fit.

ondata_rt_finetune: bool

Fine-tune mode for the on-data RT fit (requires ondata_rt and an MLP RT model). Instead of fitting a fresh ridge on raw composition features (a weak linear model that discards the pretrained representation), refit the readout over the MLP’s frozen penultimate embedding — keeping its learned nonlinear RT structure while adapting offset/scale/shape to this run.

mass_anchor: bool

Bootstrap the RT calibration from MASS-ONLY protein identifications (ms1searchpy stage-1) instead of the per-peptide RANSAC seed. On dense single-run data the per-peptide seed is ~99% contaminated; confident proteins (mass multiplicity) give far cleaner RT anchors.

rt_calibrated: bool

The RT library is already calibrated to this run’s RT (e.g. DeepLC calibrate_preds). uno then uses rt_diff = obs - pred directly and does NOT apply its own calibration on top (which would double-calibrate and distort the predictions).

ml_rescore: bool

CV-gbdt PFM peptide-quality ranking, fed through a 10-step cumulative qpred-bin ensemble protein score (ms1searchpy -ml 1 + final_iteration). VALIDATED (2026-07-07, PXD015669 human+yeast entrapment, native biosaur2 features with isoerror/isoerror2/rt_width): 223 real/4 yeast (deviation alone) -> 251 real/1 yeast with this on — deeper AND more honest. The per-PFM classifier itself still barely separates decoys from dim-real peptides on uno’s pure-Rust gbdt (dim_mean ~= DECOY_mean); the gain comes from the ENSEMBLE STRUCTURE (10 cumulative thresholds, averaged, like ms1searchpy’s final_iteration), not classifier quality. When a feature file lacks the quality columns it correctly no-ops (guarded by a strict better-than-baseline check).

ON by default (see ScoreConfig::default), and active in every benchmark reported in the manuscript. It costs a re-run of the protein engine per inclusion band, so it trades latency for depth. Set false for the plain deviation path.

Implementations

impl ScoreConfig

pub fn effective_rt_seed(&self) -> u64

pub fn effective_cv_seed(&self) -> u64

pub fn effective_classifier_seed(&self) -> u64

Trait Implementations

impl Clone for ScoreConfig

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 ScoreConfig

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

Formats the value using the given formatter.

impl Default for ScoreConfig

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for ScoreConfig

where ScoreConfig: Default,

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

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.