Data Schemas

This page documents the schemas (the data dictionary) for the key AAanalysis DataFrames: the documented, test-guarded contract for the columns each frame carries. It is generated from aaanalysis.utils.DICT_DF_SCHEMAS and kept in sync by a contract test, so the documentation cannot drift from the code.

Every column lists its dtype, whether it is required / nullable / unique, the allowed values or numeric range where applicable, a one-line meaning, and an example value.

The df_feat contract expands the df_feat entry below with the feature-id grammar, the per-residue positions encoding, and the stability policy.

• df_feat: the CPP Output Contract

df_seq

Sequence-level input table; one row per protein/sequence. Accepts four column formats (besides the always-required ‘entry’): position-based [‘sequence’, ‘tmd_start’, ‘tmd_stop’], part-based [‘jmd_n’, ‘tmd’, ‘jmd_c’], sequence-based [‘sequence’], or sequence-TMD-based [‘sequence’, ‘tmd’].

Accepted column formats (besides the always-required columns):

• position_based: sequence, tmd_start, tmd_stop

• part_based: jmd_n, tmd, jmd_c

• sequence_based: sequence

• sequence_tmd_based: sequence, tmd

Column	Type	Required	Nullable	Unique	Description	Allowed / range / example
entry	str	yes	no	yes	Unique protein/sequence identifier.	e.g. P05067
sequence	str	no	no	no	Amino acid sequence (one-letter code).	e.g. MLPGLALLLL…
tmd_start	int	no	no	no	1-based start position of the TMD in ‘sequence’ (position-based format).	range: [1, inf]; e.g. 37
tmd_stop	int	no	no	no	1-based stop position of the TMD in ‘sequence’, inclusive (position-based format).	range: [1, inf]; e.g. 59
jmd_n	str	no	no	no	N-terminal juxtamembrane domain subsequence (part-based format).	e.g. NSPFYYDWHS
tmd	str	no	no	no	Transmembrane domain subsequence (part-/sequence-TMD formats).	e.g. LQVGGLICAGVL
jmd_c	str	no	no	no	C-terminal juxtamembrane domain subsequence (part-based format).	e.g. KCKCKFGQKS
label	int	no	no	no	Class/group label. Binary contrasts by convention (reference=0, test=1); multi-class via the SequenceFeature.get_labels_* helpers.	e.g. 1
name	str	no	yes	no	Human-readable protein name.	e.g. APP

df_parts

Sequence parts table consumed by CPP / SequenceFeature; one row per sequence. Columns are DYNAMIC: one column per selected sequence part, named by the part vocabulary; each value is the part’s amino acid subsequence.

Dynamic columns (dtype: str, nullable: False): Amino acid subsequence of the named sequence part. Column names are drawn from the part vocabulary: tmd, tmd_e, tmd_n, tmd_c, jmd_n, jmd_c, ext_c, ext_n, tmd_jmd, jmd_n_tmd_n, tmd_c_jmd_c, ext_n_tmd_n, tmd_c_ext_c.

df_scales

Amino acid scale matrix (AAontology / AAindex-derived). Loaded via load_scales; consumed by CPP / SequenceFeature.

Matrix / array contract:

• index: amino acid (one-letter; 20 canonical)

• columns: scale id (e.g. ‘KLEP840101’)

• values: min-max normalised scale value in [0, 1] (dtype: float)

df_cat

AAontology scale metadata (one row per scale id). Loaded via load_scales(name=’scales_cat’).

Column	Type	Required	Nullable	Unique	Description	Allowed / range / example
scale_id	str	yes	no	yes	Amino acid scale identifier; the join key to df_scales columns.	e.g. KLEP840101
category	str	yes	no	no	AAontology main category.	allowed: ASA/Volume, Composition, Conformation, Energy, Others, Polarity, …; e.g. Energy
subcategory	str	yes	no	no	AAontology subcategory.	e.g. Charge
scale_name	str	yes	no	no	Human-readable scale name.	e.g. Charge
scale_description	str	yes	no	no	One-sentence scale description.	e.g. Net charge (Klein et al., 1984).

df_subcat

AAontology subcategory overview (one row per subcategory). Loaded via load_scales(name=’subcat’).

Column	Type	Required	Nullable	Unique	Description	Allowed / range / example
category	str	yes	no	no	AAontology main category.	allowed: ASA/Volume, Composition, Conformation, Energy, Others, Polarity, …; e.g. Energy
subcategory	str	yes	no	yes	AAontology subcategory.	e.g. Charge
cluster	str	yes	no	no	Cluster the subcategory belongs to.	e.g. Electrostatics
interpret_grade	int	yes	no	no	Interpretability grade, 1-10 (1 = best).	range: [1, 10]; e.g. 2
top_explain	float	yes	yes	no	Interpretability-tier threshold; NaN for unclassified subcategories.	e.g. 0.5
n_scales	int	yes	no	no	Number of scales in the subcategory.	range: [1, inf]; e.g. 12
n_scales_aaindex	int	yes	no	no	Number of those scales originating from AAindex.	range: [0, inf]; e.g. 10
subcategory_description	str	yes	no	no	Subcategory description.	e.g. Net charge of the residue.
key_references	str	yes	yes	no	Key literature references.	e.g. Klein84

df_feat

CPP feature table (the primary downstream contract; see also the simple DICT_DF_FEAT). One row per feature. Columns 1-13 are the canonical lower bound (LIST_COLS_FEAT); optional/dynamic columns follow in stable order.

Column	Type	Required	Nullable	Unique	Description	Allowed / range / example
feature	str	yes	no	yes	Opaque PART-SPLIT-SCALE feature id; split with split_feat_id, never parse by hand.	e.g. TMD_C_JMD_C-Segment(3,4)-KLEP840101
category	str	yes	no	no	AAontology category of the feature’s scale.	allowed: ASA/Volume, Composition, Conformation, Energy, Others, Polarity, …; e.g. Energy
subcategory	str	yes	no	no	AAontology subcategory.	e.g. Charge
scale_name	str	yes	no	no	Human-readable scale name.	e.g. Charge
scale_description	str	yes	no	no	One-sentence scale description.	e.g. Net charge (Klein et al., 1984).
abs_auc	float	yes	no	no	Absolute adjusted AUC; primary ranking statistic.	range: [-0.5, 0.5]; e.g. 0.244
abs_mean_dif	float	yes	no	no	Absolute mean difference between test and reference group.	range: [0, 1]; e.g. 0.104
mean_dif	float	yes	no	no	Signed mean difference (test - reference); the sign gives the direction.	range: [-1, 1]; e.g. 0.104
std_test	float	yes	no	no	Standard deviation in the test group.	range: [0, inf]; e.g. 0.107
std_ref	float	yes	no	no	Standard deviation in the reference group.	range: [0, inf]; e.g. 0.11
p_val_mann_whitney	float	yes	no	no	Mann-Whitney U p-value (default). Named ‘p_val_ttest_indep’ when parametric=True.	range: [0, 1]; e.g. 0.0
p_val_fdr_bh	float	yes	no	no	Benjamini-Hochberg FDR-corrected p-value.	range: [0, 1]; e.g. 0.0
positions	str	yes	no	no	Comma-separated 1-based residue positions the feature spans.	e.g. 31,32,33,34,35
p_val_ttest_indep	float	no	no	no	Independent t-test p-value; replaces ‘p_val_mann_whitney’ when parametric=True.	range: [0, 1]; e.g. 0.01
amino_acids_test	str	no	no	no	Amino acids at the feature positions in the test group (diagnostic).	e.g. K,R,K
amino_acids_ref	str	no	no	no	Amino acids at the feature positions in the reference group (diagnostic).	e.g. A,L,G
feature_description	str	no	yes	no	Optional readable one-sentence feature description.	e.g. Charge of the C-terminal TMD segment.
selection_frequency	float	no	no	no	Bootstrap selection frequency (fraction of resampling rounds a feature was selected); present only when CPP(bootstrap=True).	range: [0, 1]; e.g. 0.9
abs_auc_ci_low	float	no	yes	no	Lower bound of the bootstrap confidence interval of ‘abs_auc’; present only when CPP(bootstrap=True) runs with a confidence level (bootstrap_kws={‘ci’: <level>}).	range: [-0.5, 0.5]; e.g. 0.18
abs_auc_ci_high	float	no	yes	no	Upper bound of the bootstrap confidence interval of ‘abs_auc’; present only when CPP(bootstrap=True) runs with a confidence level (bootstrap_kws={‘ci’: <level>}).	range: [-0.5, 0.5]; e.g. 0.31
mean_dif_ci_low	float	no	yes	no	Lower bound of the bootstrap confidence interval of ‘mean_dif’; present only when CPP(bootstrap=True) runs with a confidence level (bootstrap_kws={‘ci’: <level>}).	range: [-1, 1]; e.g. 0.06
mean_dif_ci_high	float	no	yes	no	Upper bound of the bootstrap confidence interval of ‘mean_dif’; present only when CPP(bootstrap=True) runs with a confidence level (bootstrap_kws={‘ci’: <level>}).	range: [-1, 1]; e.g. 0.14
feat_importance	float	no	no	no	Feature importance from TreeModel.fit (post-fit).	range: [0, inf]; e.g. 0.97
feat_importance_std	float	no	no	no	Std of the feature importance across CV rounds (post-fit).	range: [0, inf]; e.g. 1.44
feat_impact	float	no	no	no	SHAP-based signed feature impact from ShapModel (post-fit, pro).	e.g. 0.12
feat_impact_std	float	no	no	no	Std of the feature impact (post-fit, pro).	range: [0, inf]; e.g. 0.03

df_eval

Evaluation table; one row per evaluated set. Columns depend on the evaluating class (AAclust vs dPULearn); ‘name’ labels the row when present.

Column	Type	Required	Nullable	Unique	Description	Allowed / range / example
name	str	no	no	no	Label of the evaluated set.	e.g. Set 1
n_clusters	int	no	no	no	AAclust: number of clusters.	range: [1, inf]; e.g. 8
BIC	float	no	no	no	AAclust: Bayesian Information Criterion.	e.g. -1234.5
CH	float	no	no	no	AAclust: Calinski-Harabasz score.	range: [0, inf]; e.g. 42.1
SC	float	no	no	no	AAclust: Silhouette Coefficient.	range: [-1, 1]; e.g. 0.34
n_rel_neg	int	no	no	no	dPULearn: number of reliable negatives identified.	range: [0, inf]; e.g. 50
avg_STD	float	no	no	no	dPULearn: average within-group standard deviation.	range: [0, inf]; e.g. 0.21
avg_KLD_pos	float	no	no	no	dPULearn: average KL divergence vs the positive set.	e.g. 0.12

df_mut

AAMut per-scale mutation effects (one row per scale for a substitution).

Column	Type	Required	Nullable	Unique	Description	Allowed / range / example
from_aa	str	yes	no	no	Substituted-from amino acid (one letter).	allowed: A, C, D, E, F, G, …; e.g. M
to_aa	str	yes	no	no	Substituted-to amino acid (one letter).	allowed: A, C, D, E, F, G, …; e.g. V
scale_id	str	yes	no	no	Amino acid scale identifier.	e.g. KLEP840101
category	str	yes	no	no	AAontology category.	allowed: ASA/Volume, Composition, Conformation, Energy, Others, Polarity, …; e.g. Energy
subcategory	str	yes	no	no	AAontology subcategory.	e.g. Charge
delta	float	yes	no	no	Signed per-scale substitution delta (to_aa - from_aa).	e.g. -0.31
abs_delta	float	yes	no	no	Magnitude of the per-scale delta.	range: [0, inf]; e.g. 0.31

df_seqmut_scan

SeqMut.scan per-mutation scan (one row per scanned single mutation).

Column	Type	Required	Nullable	Unique	Description	Allowed / range / example
entry	str	yes	no	no	Protein/sequence identifier.	e.g. P05067
pos	int	yes	no	no	1-based mutated position.	range: [1, inf]; e.g. 123
from_aa	str	yes	no	no	Original amino acid.	allowed: A, C, D, E, F, G, …; e.g. M
to_aa	str	yes	no	no	Substituted amino acid.	allowed: A, C, D, E, F, G, …; e.g. V
mutation	str	yes	no	no	HGVS-like label ‘<from><pos><to>’.	e.g. M123V
region	str	yes	no	no	Sequence part the position falls in.	allowed: jmd_n, tmd, jmd_c; e.g. tmd
delta_cpp	float	yes	no	no	Summed absolute (L1) feature-space magnitude of the mutation.	range: [0, inf]; e.g. 2.4
shift_score	float	yes	no	no	Signed shift toward the test-class profile.	e.g. -0.8
delta_pred	float	no	no	no	Change of the model prediction score (percentage points) the mutation induces; present only when a fitted model is bound to SeqMut.	e.g. 12.5
wt_pred	float	no	no	no	Wild-type prediction score (%) for the target class (per-entry constant); present only with a model.	range: [0, 100]; e.g. 41.0
wt_pred_std	float	no	yes	no	Standard deviation of the wild-type prediction score (%); NaN when the model gives no std. Present only with a model.	range: [0, 100]; e.g. 49.2

df_seqmut_variant

SeqMut.combine combined-variant table (one row per multi-mutation variant; all of a variant’s point mutations are applied to the same sequence).

Column	Type	Required	Nullable	Unique	Description	Allowed / range / example
entry	str	yes	no	no	Protein/sequence identifier.	e.g. P05067
variant	str	yes	no	no	Combined-variant label, ‘+’-joined single mutations.	e.g. R20K+K27P
n_mut	int	yes	no	no	Number of point mutations in the variant.	range: [1, inf]; e.g. 2
sequence_mut	str	yes	no	no	Full sequence with all the variant’s mutations applied.	e.g. …K…P…
delta_cpp	float	yes	no	no	Summed absolute (L1) feature-space magnitude of the combined variant.	range: [0, inf]; e.g. 4.1
shift_score	float	yes	no	no	Signed shift toward the test-class profile for the combined variant.	e.g. 1.3
delta_pred	float	no	no	no	Change of the model prediction score (percentage points) for the combined variant; present only with a model.	e.g. 18.7
candidate_id	str	no	no	no	Lineage identifier: sha256 content hash over the source sequence and the candidate’s ordered mutations; present only when a lineage record was requested.	e.g. sha256:9f86d0818…

df_seqmut_eval

SeqMut.eval per-protein/region summary (one row per region).

Column	Type	Required	Nullable	Unique	Description	Allowed / range / example
entry	str	yes	no	no	Protein/sequence identifier.	e.g. P05067
region	str	yes	no	no	Sequence part.	allowed: jmd_n, tmd, jmd_c; e.g. tmd
n_mut	int	yes	no	no	Number of scanned mutations.	range: [0, inf]; e.g. 20
n_disruptive	int	yes	no	no	Number flagged disruptive.	range: [0, inf]; e.g. 3
frac_disruptive	float	yes	no	no	n_disruptive / n_mut.	range: [0, 1]; e.g. 0.15
mean_delta_cpp	float	yes	no	no	Mean delta_cpp (summed absolute feature-space magnitude) over scanned mutations.	range: [0, inf]; e.g. 1.1

df_pareto

SeqOpt.run Pareto front (one row per non-dominated variant of the single wild-type). One column per objective is inserted between ‘sequence_mut’ and ‘rank’ at run time (e.g. delta_pred, n_mut), so the objective columns below are documented as optional/dynamic.

Column	Type	Required	Nullable	Unique	Description	Allowed / range / example
entry	str	yes	no	no	Wild-type protein/sequence identifier (constant).	e.g. P05067
variant	str	yes	no	no	Variant label, ‘+’-joined single mutations.	e.g. R20K+K27P
n_mut	int	yes	no	no	Number of point mutations in the variant.	range: [0, inf]; e.g. 2
sequence_mut	str	yes	no	no	Full sequence with all the variant’s mutations applied.	e.g. …K…P…
rank	int	yes	no	no	Non-dominated front index from fast non-dominated sorting (0 = best/first front).	range: [0, inf]; e.g. 0
crowding	float	yes	yes	no	NSGA-II crowding distance within the front (larger = more isolated = preferred); inf at the boundary points.	range: [0, inf]; e.g. 1.7
delta_pred	float	no	no	no	Objective: change of the model prediction score (percentage points). Present when an objective uses it.	e.g. 18.7
delta_cpp	float	no	no	no	Objective: summed absolute (L1) feature-space magnitude. Present when an objective uses it.	range: [0, inf]; e.g. 4.1
shift_score	float	no	no	no	Objective: signed shift toward the test-class profile. Present when an objective uses it.	e.g. 1.3
candidate_id	str	no	no	no	Lineage identifier: sha256 content hash over the source sequence and the candidate’s ordered mutations; present only when a lineage record was requested.	e.g. sha256:9f86d0818…

df_seqopt_eval

SeqOpt.eval Pareto-quality summary (one row per run / front).

Column	Type	Required	Nullable	Unique	Description	Allowed / range / example
hypervolume	float	yes	no	no	Objective-space volume dominated by the first front relative to the reference point.	range: [0, inf]; e.g. 0.62
n_front	int	yes	no	no	Number of variants on the first (rank=0) front.	range: [1, inf]; e.g. 12
spread	float	yes	yes	no	Objective-space diversity / spread of the front (0 = degenerate).	range: [0, inf]; e.g. 0.41

df_annot

Canonical per-residue annotation table (AnnotationPreprocessor); one row per annotated residue/feature.

Column	Type	Required	Nullable	Unique	Description	Allowed / range / example
protein_id	str	yes	no	no	UniProt accession (mirrors ‘entry’).	e.g. P05067
start	int	yes	no	no	1-based start (UniProt-canonical frame).	range: [1, inf]; e.g. 672
end	int	yes	no	no	1-based stop, inclusive (single residue: start==end).	range: [1, inf]; e.g. 714
aa	str	yes	yes	no	Expected residue identity (encode-time guard).	e.g. K
feature_type	str	yes	no	no	Registry key (e.g. ‘phospho’, ‘binding’).	e.g. binding
category	str	no	no	no	Annotation category.	e.g. PTM
source	str	yes	no	no	‘UniProt’ or a user source name.	e.g. UniProt
evidence	str	no	yes	no	ECO evidence code.	e.g. ECO:0000269
score	float	no	yes	no	Optional annotation score.	range: [0, 1]; e.g. 0.9
bond_id	str	no	yes	no	Pairing id for DISULFID / CROSSLNK endpoints.	e.g. b1

df_logo

AALogo composition/probability matrix consumed by sequence-logo plots.

Matrix / array contract:

• index: 1-based sequence position

• columns: amino acid (one-letter)

• values: per-position composition / probability / information value (dtype: float)

df_pred

AAPred.predict output in long format; the columns depend on ‘level’: ‘sequence’ = entry, score, score_std (one row per protein); ‘domain’ = entry, offset, score, is_best (one row per protein and boundary shift); ‘window’ = entry, position, score, score_std (one row per protein and residue anchor). ‘predicted_label’ is appended when a threshold is given. Both score columns carry the scale chosen by score_range: ‘proba’ (the default, [0, 1]) or ‘percent’ (the same values multiplied by 100, [0, 100]); the score row below contracts one range per scale.

Column	Type	Required	Nullable	Unique	Description	Allowed / range / example
entry	str	yes	no	no	Protein identifier from df_seq; unique only at level=’sequence’.	e.g. P05067
offset	int	no	no	no	level=’domain’: boundary shift applied to tmd_start and tmd_stop, in residues.	e.g. -2
position	int	no	no	no	level=’window’: 1-based anchor position of the scored window.	range: [1, inf]; e.g. 31
score	float	yes	no	no	Positive-class score averaged over the fitted models. Its range depends on the score_range argument: [0, 1] on the default ‘proba’ scale and [0, 100] on ‘percent’, which holds the same values times 100.	range per score_range: proba: [0, 1]; percent: [0, 100]; e.g. 0.83
score_std	float	no	no	no	Standard deviation of the score across the fitted models (level=’sequence’ and ‘window’), on the same scale as score.	range: [0, inf]; e.g. 0.04
is_best	bool	no	no	no	level=’domain’: True for the highest-scoring offset of each protein.	e.g. True
predicted_label	int	no	no	no	Class label from the score when a threshold is given (score >= threshold is the positive label).	e.g. 1

df_rel

ReliabilityModel.predict output (and ReliabilityModel.predict_candidates, which returns the same columns for a rebuilt candidate matrix); one row per sample, one column per reliability axis: stability (score_std, ci_*), applicability domain (ood_score, in_domain, ad_*), score ambiguity (margin, entropy), validity (conformal_set) and the headline flag (reliable).

Column	Type	Required	Nullable	Unique	Description	Allowed / range / example
score	float	yes	no	no	Positive-class probability averaged over the ensemble members.	range: [0, 1]; e.g. 0.78
score_std	float	yes	no	no	Standard deviation of the score across the ensemble members.	range: [0, inf]; e.g. 0.06
ci_low	float	yes	no	no	Lower bound of the score’s confidence interval.	range: [0, 1]; e.g. 0.66
ci_high	float	yes	no	no	Upper bound of the score’s confidence interval.	range: [0, 1]; e.g. 0.88
ood_score	float	yes	yes	no	k-NN distance relative to the training threshold (1.0 = threshold); NaN when the training reference is degenerate.	range: [0, inf]; e.g. 0.72
in_domain	bool	yes	no	no	True if the sample lies inside the training applicability domain (ood_score <= 1).	e.g. True
ad_knn	float	yes	no	no	Mean distance to the k nearest training samples in the standardized feature space.	range: [0, inf]; e.g. 1.9
ad_mahalanobis	float	yes	yes	no	Mahalanobis distance to the training center; NaN when the training reference is degenerate (n_features >= n_samples), where the covariance is rank-deficient and the distance is not identifiable.	range: [0, inf]; e.g. 2.4
ad_leverage	float	yes	yes	no	Leverage (hat value) relative to the training feature space; NaN on the same degenerate training reference as ad_mahalanobis.	range: [0, inf]; e.g. 0.05
score_calibrated	float	yes	yes	no	Calibrated positive-class probability; NaN when no calibrator is available (calibration was disabled or could not be fitted).	range: [0, 1]; e.g. 0.74
margin	float	yes	no	no	Sharpness |p - 0.5| * 2 (1 = decisive, 0 = coin-flip), from the calibrated score when available and otherwise the ensemble score.	range: [0, 1]; e.g. 0.48
entropy	float	yes	no	no	Binary entropy of the calibrated score when available, otherwise the ensemble score (0 = decisive, 1 = coin-flip).	range: [0, 1]; e.g. 0.83
conformal_set	str	yes	no	no	Split-conformal prediction set.	allowed: neg, pos, both, none; e.g. pos
reliable	bool	yes	no	no	Headline flag: in the applicability domain and a conformal singleton, or margin >= 0.5 when no conformal reference is available.	e.g. True
ad_status	str	yes	no	no	Banded applicability-domain verdict derived from ood_score and the fitted ad_borderline band; ‘unknown’ when the training reference is degenerate.	allowed: inside, borderline, outside, unknown; e.g. inside
ad_nearest_train	int	yes	no	no	0-based row index of the closest training sample in the X passed to fit.	range: [0, inf]; e.g. 12

df_eval_reliability

ReliabilityModel.eval output: one row per equal-width score bin (the calibration curve) plus a summary row (bin=’summary’) holding the in-domain fraction in mean_score, the empirical conformal coverage in empirical_pos, and the number of evaluated samples in n_samples. With add_metrics=True, rows bin=’brier’ and bin=’ece’ follow, holding the Brier score and the expected calibration error in mean_score (empirical_pos is NaN).

Column	Type	Required	Nullable	Unique	Description	Allowed / range / example
bin	str	yes	no	yes	Score-bin label, ‘summary’ for the summary row, or ‘brier’ / ‘ece’ for the optional calibration-metric rows.	e.g. 0.00-0.20
mean_score	float	yes	yes	no	Mean predicted score in the bin (summary row: in-domain fraction); NaN for an empty bin.	range: [0, 1]; e.g. 0.12
empirical_pos	float	yes	yes	no	Empirical positive rate in the bin (summary row: conformal coverage); NaN for an empty bin and for the metric rows.	range: [0, 1]; e.g. 0.1
n_samples	int	yes	no	no	Number of samples in the bin (summary row: all evaluated samples).	range: [0, inf]; e.g. 18

df_eval_selective

AAPred.eval_selective output: the selective-prediction (risk-coverage) measurement in long format, one row per (metric, coverage level). Samples are ranked by a per-sample confidence signal and each level retains the leading ceil(target * n_samples) of them, so the reported coverage is the fraction actually retained (it can exceed the requested target after rounding up) and the coverage=1.0 row is the ordinary out-of-fold score of that metric. score_aurc repeats, on each of a metric’s rows, the area under that metric’s own coverage-performance curve.

Column	Type	Required	Nullable	Unique	Description	Allowed / range / example
metric	str	yes	no	no	Performance metric scored on the retained subset.	allowed: accuracy, balanced_accuracy, precision, recall, f1, roc_auc, …; e.g. balanced_accuracy
coverage	float	yes	no	no	Fraction of samples actually retained at this level (n_retained / n_samples), which can exceed the requested target after rounding up.	range: [0, 1]; e.g. 0.6
n_retained	int	yes	no	no	Samples retained at this level: ceil(target * n_samples), at least 1.	range: [1, inf]; e.g. 18
score	float	yes	yes	no	Metric value on the retained subset; NaN where the metric is undefined there (balanced_accuracy and roc_auc need both classes present).	e.g. 0.91
score_aurc	float	yes	yes	no	Area under this metric’s coverage-performance curve divided by the actual coverage span (a flat curve at 0.8 has an area of 0.8); NaN for a single-level grid or a curve with a NaN point.	e.g. 0.88

X

ML-ready numeric feature matrix produced by SequenceFeature.feature_matrix; consumed by TreeModel / sklearn.

Matrix / array contract:

• index: sample (row order matches df_seq / labels)

• columns: feature (column order matches df_feat[‘feature’])

• values: scale value of the feature for the sample (dtype: float)

prediction

TreeModel.predict_proba output: two 1-D arrays of length n_samples, (pred, pred_std) – the Monte-Carlo mean and standard deviation of the positive-class probability per sample.

Matrix / array contract:

• index: sample (row order matches X)

• columns: (pred, pred_std)

• values: positive-class probability in [0, 1] (pred) and its std (dtype: float)