pg-sui 0.2.0__py3-none-any.whl → 1.6.14.dev9__py3-none-any.whl

pgsui/data_processing/containers.py

@@ -0,0 +1,1424 @@
+from __future__ import annotations
+
+from dataclasses import asdict, dataclass, field
+from typing import Any, Dict, Literal, Optional, Sequence
+
+from pgsui.data_processing.config import apply_dot_overrides, load_yaml_to_dataclass
+
+
+@dataclass
+class _SimParams:
+    """Container for simulation hyperparameters.
+
+    Attributes:
+        prop_missing (float): Proportion of missing values to simulate.
+        strategy (Literal["random", "random_inv_genotype"]): Strategy for simulating missing values.
+        missing_val (int | float): Value to represent missing data.
+        het_boost (float): Boost factor for heterozygous genotypes.
+        seed (int | None): Random seed for reproducibility.
+    """
+
+    prop_missing: float = 0.3
+    strategy: Literal["random", "random_inv_genotype"] = "random_inv_genotype"
+    missing_val: int | float = -1
+    het_boost: float = 2.0
+    seed: int | None = None
+
+    def to_dict(self) -> dict:
+        return asdict(self)
+
+
+@dataclass
+class _ImputerParams:
+    """Container for imputer hyperparameters.
+
+    Attributes:
+        n_nearest_features (int | None): Number of nearest features to consider for imputation.
+        max_iter (int): Maximum number of iterations for the imputation algorithm.
+        initial_strategy (Literal["mean", "median", "most_frequent", "constant"]): Strategy for initial imputation.
+        keep_empty_features (bool): Whether to keep features that are entirely missing.
+        random_state (int | None): Random seed for reproducibility.
+        verbose (bool): If True, enables verbose logging during imputation.
+    """
+
+    n_nearest_features: int | None = 10
+    max_iter: int = 10
+    initial_strategy: Literal["mean", "median", "most_frequent", "constant"] = (
+        "most_frequent"
+    )
+    keep_empty_features: bool = True
+    random_state: int | None = None
+    verbose: bool = False
+
+    def to_dict(self) -> dict:
+        return asdict(self)
+
+
+@dataclass
+class _RFParams:
+    """Container for RandomForest hyperparameters.
+
+    Attributes:
+        n_estimators (int): Number of trees in the forest.
+        max_depth (int | None): Maximum depth of the trees.
+        min_samples_split (int): Minimum number of samples required to split an internal node.
+        min_samples_leaf (int): Minimum number of samples required to be at a leaf node.
+        max_features (Literal["sqrt", "log2"] | float | None): Number of features to consider for split.
+        criterion (Literal["gini", "entropy", "log_loss"]): Function to measure the quality of a split.
+        class_weight (Literal["balanced", "balanced_subsample", None]): Weights associated with classes.
+    """
+
+    n_estimators: int = 300
+    max_depth: int | None = None
+    min_samples_split: int = 2
+    min_samples_leaf: int = 1
+    max_features: Literal["sqrt", "log2"] | float | None = "sqrt"
+    criterion: Literal["gini", "entropy", "log_loss"] = "gini"
+    class_weight: Literal["balanced", "balanced_subsample", None] = "balanced"
+
+    def to_dict(self) -> dict:
+        return asdict(self)
+
+
+@dataclass
+class _HGBParams:
+    """Container for HistGradientBoosting hyperparameters.
+
+    Attributes:
+        max_iter (int): Maximum number of iterations.
+        learning_rate (float): Learning rate shrinks the contribution of each tree.
+        max_depth (int | None): Maximum depth of the individual regression estimators.
+        min_samples_leaf (int): Minimum number of samples required to be at a leaf node.
+        n_iter_no_change (int): Number of iterations with no improvement to wait before early stopping.
+        tol (float): Tolerance for the early stopping.
+        max_features (float | None): The fraction of features to consider when looking for the best split.
+        class_weight (Literal["balanced", "balanced_subsample", None]): Weights associated with classes.
+        random_state (int | None): Random seed for reproducibility.
+        verbose (bool): If True, enables verbose logging during training.
+    """
+
+    max_iter: int = 100
+    learning_rate: float = 0.1
+    max_depth: int | None = None
+    min_samples_leaf: int = 1
+    n_iter_no_change: int = 10
+    tol: float = 1e-7
+    max_features: float | None = 1.0
+    class_weight: Literal["balanced", "balanced_subsample", None] = "balanced"
+    random_state: int | None = None
+    verbose: bool = False
+
+    def to_dict(self) -> dict:
+        return asdict(self)
+
+
+@dataclass
+class ModelConfig:
+    """Model architecture configuration.
+
+    Attributes:
+        latent_init (Literal["random", "pca"]): Method for initializing the latent space.
+        latent_dim (int): Dimensionality of the latent space.
+        dropout_rate (float): Dropout rate for regularization.
+        num_hidden_layers (int): Number of hidden layers in the neural network.
+        hidden_activation (Literal["relu", "elu", "selu", "leaky_relu"]): Activation function.
+        layer_scaling_factor (float): Scaling factor for the number of neurons in hidden layers.
+        layer_schedule (Literal["pyramid", "constant", "linear"]): Schedule for scaling hidden layer sizes.
+        gamma (float): Parameter for the focal loss function.
+    """
+
+    latent_init: Literal["random", "pca"] = "random"
+    latent_dim: int = 2
+    dropout_rate: float = 0.2
+    num_hidden_layers: int = 2
+    hidden_activation: Literal["relu", "elu", "selu", "leaky_relu"] = "relu"
+    layer_scaling_factor: float = 5.0
+    layer_schedule: Literal["pyramid", "constant", "linear"] = "pyramid"
+    gamma: float = 2.0
+
+
+@dataclass
+class TrainConfig:
+    """Training procedure configuration.
+
+    Attributes:
+        batch_size (int): Number of samples per training batch.
+        learning_rate (float): Learning rate for the optimizer.
+        lr_input_factor (float): Factor to scale the learning rate for input layer.
+        l1_penalty (float): L1 regularization penalty.
+        early_stop_gen (int): Number of generations with no improvement to wait before early stopping.
+        min_epochs (int): Minimum number of epochs to train.
+        max_epochs (int): Maximum number of epochs to train.
+        validation_split (float): Proportion of data to use for validation.
+        weights_beta (float): Smoothing factor for class weights.
+        weights_max_ratio (float): Maximum ratio for class weights to prevent extreme values.
+        device (Literal["gpu", "cpu", "mps"]): Device to use for computation.
+    """
+
+    batch_size: int = 32
+    learning_rate: float = 1e-3
+    lr_input_factor: float = 1.0
+    l1_penalty: float = 0.0
+    early_stop_gen: int = 20
+    min_epochs: int = 100
+    max_epochs: int = 5000
+    validation_split: float = 0.2
+    weights_beta: float = 0.9999
+    weights_max_ratio: float = 1.0
+    device: Literal["gpu", "cpu", "mps"] = "cpu"
+
+
+@dataclass
+class TuneConfig:
+    """Hyperparameter tuning configuration.
+
+    Attributes:
+        enabled (bool): If True, enables hyperparameter tuning.
+        metric (Literal["f1", "accuracy", "pr_macro"]): Metric to optimize during tuning.
+        n_trials (int): Number of hyperparameter trials to run.
+        resume (bool): If True, resumes tuning from a previous state.
+        save_db (bool): If True, saves the tuning results to a database.
+        fast (bool): If True, uses a faster but less thorough tuning approach.
+        max_samples (int): Maximum number of samples to use for tuning. 0 means all samples.
+        max_loci (int): Maximum number of loci to use for tuning. 0 means all loci.
+        epochs (int): Number of epochs to train each trial.
+        batch_size (int): Batch size for training during tuning.
+        eval_interval (int): Interval (in epochs) at which to evaluate the model during tuning.
+        infer_epochs (int): Number of epochs for inference during tuning.
+        patience (int): Number of evaluations with no improvement before stopping early.
+        proxy_metric_batch (int): If > 0, uses a subset of data for proxy metric evaluation.
+    """
+
+    enabled: bool = False
+    metric: Literal[
+        "f1",
+        "accuracy",
+        "pr_macro",
+        "average_precision",
+        "roc_auc",
+        "precision",
+        "recall",
+    ] = "f1"
+    n_trials: int = 100
+    resume: bool = False
+    save_db: bool = False
+    fast: bool = True
+    max_samples: int = 512
+    max_loci: int = 0  # 0 = all
+    epochs: int = 500
+    batch_size: int = 64
+    eval_interval: int = 20
+    infer_epochs: int = 100
+    patience: int = 10
+    proxy_metric_batch: int = 0
+
+
+@dataclass
+class EvalConfig:
+    """Evaluation configuration.
+
+    Attributes:
+        eval_latent_steps (int): Number of optimization steps for latent space evaluation.
+        eval_latent_lr (float): Learning rate for latent space optimization.
+        eval_latent_weight_decay (float): Weight decay for latent space optimization.
+    """
+
+    eval_latent_steps: int = 50
+    eval_latent_lr: float = 1e-2
+    eval_latent_weight_decay: float = 0.0
+
+
+@dataclass
+class PlotConfig:
+    """Plotting configuration.
+
+    Attributes:
+        fmt (Literal["pdf", "png", "jpg", "jpeg", "svg"]): Output file format.
+        dpi (int): Dots per inch for the output figure.
+        fontsize (int): Font size for text in the plots.
+        despine (bool): If True, removes the top and right spines from plots.
+        show (bool): If True, displays the plot interactively.
+    """
+
+    fmt: Literal["pdf", "png", "jpg", "jpeg", "svg"] = "pdf"
+    dpi: int = 300
+    fontsize: int = 18
+    despine: bool = True
+    show: bool = False
+
+
+@dataclass
+class IOConfig:
+    """I/O configuration.
+
+    Attributes:
+        prefix (str): Prefix for output files. Default is "pgsui".
+        verbose (bool): If True, enables verbose logging. Default is False.
+        debug (bool): If True, enables debug mode. Default is False.
+        seed (int | None): Random seed for reproducibility. Default is None.
+        n_jobs (int): Number of parallel jobs to run. Default is 1.
+        scoring_averaging (Literal["macro", "micro", "weighted"]): Averaging method.
+    """
+
+    prefix: str = "pgsui"
+    verbose: bool = False
+    debug: bool = False
+    seed: int | None = None
+    n_jobs: int = 1
+    scoring_averaging: Literal["macro", "micro", "weighted"] = "macro"
+
+
+@dataclass
+class SimConfig:
+    """Top-level configuration for data simulation and imputation.
+
+    Attributes:
+        simulate_missing (bool): If True, simulates missing data.
+        sim_strategy (Literal["random", ...]): Strategy for simulating missing data.
+        sim_prop (float): Proportion of data to simulate as missing.
+        sim_kwargs (dict | None): Additional keyword arguments for simulation.
+    """
+
+    simulate_missing: bool = False
+    sim_strategy: Literal[
+        "random",
+        "random_weighted",
+        "random_weighted_inv",
+        "nonrandom",
+        "nonrandom_weighted",
+    ] = "random"
+    sim_prop: float = 0.10
+    sim_kwargs: dict | None = None
+
+
+@dataclass
+class NLPCAConfig:
+    """Top-level configuration for ImputeNLPCA.
+
+    Attributes:
+        io (IOConfig): I/O configuration.
+        model (ModelConfig): Model architecture configuration.
+        train (TrainConfig): Training procedure configuration.
+        tune (TuneConfig): Hyperparameter tuning configuration.
+        evaluate (EvalConfig): Evaluation configuration.
+        plot (PlotConfig): Plotting configuration.
+        sim (SimConfig): Simulation configuration.
+    """
+
+    io: IOConfig = field(default_factory=IOConfig)
+    model: ModelConfig = field(default_factory=ModelConfig)
+    train: TrainConfig = field(default_factory=TrainConfig)
+    tune: TuneConfig = field(default_factory=TuneConfig)
+    evaluate: EvalConfig = field(default_factory=EvalConfig)
+    plot: PlotConfig = field(default_factory=PlotConfig)
+    sim: SimConfig = field(default_factory=SimConfig)
+
+    @classmethod
+    def from_preset(
+        cls, preset: Literal["fast", "balanced", "thorough"] = "balanced"
+    ) -> "NLPCAConfig":
+        """Build a NLPCAConfig from a named preset."""
+        if preset not in {"fast", "balanced", "thorough"}:
+            raise ValueError(f"Unknown preset: {preset}")
+
+        cfg = cls()
+
+        # Common baselines
+        cfg.io.verbose = False
+        cfg.train.validation_split = 0.20
+        cfg.model.hidden_activation = "relu"
+        cfg.model.layer_schedule = "pyramid"
+        cfg.model.latent_init = "random"
+        cfg.evaluate.eval_latent_lr = 1e-2
+        cfg.evaluate.eval_latent_weight_decay = 0.0
+        cfg.sim.simulate_missing = True
+        cfg.sim.sim_strategy = "random"
+        cfg.sim.sim_prop = 0.2
+
+        if preset == "fast":
+            # Model
+            cfg.model.latent_dim = 4
+            cfg.model.num_hidden_layers = 1
+            cfg.model.layer_scaling_factor = 2.0
+            cfg.model.dropout_rate = 0.10
+            cfg.model.gamma = 1.5
+            # Train
+            cfg.train.batch_size = 128
+            cfg.train.learning_rate = 1e-3
+            cfg.train.early_stop_gen = 5
+            cfg.train.min_epochs = 10
+            cfg.train.max_epochs = 120
+            cfg.train.weights_beta = 0.9999
+            cfg.train.weights_max_ratio = 2.0
+            # Tuning
+            cfg.tune.enabled = True
+            cfg.tune.fast = True
+            cfg.tune.n_trials = 25
+            cfg.tune.epochs = 120
+            cfg.tune.batch_size = 128
+            cfg.tune.max_samples = 512
+            cfg.tune.max_loci = 0
+            cfg.tune.eval_interval = 20
+            cfg.tune.infer_epochs = 20
+            cfg.tune.patience = 5
+            cfg.tune.proxy_metric_batch = 0
+            # Eval
+            cfg.evaluate.eval_latent_steps = 20
+
+        elif preset == "balanced":
+            # Model
+            cfg.model.latent_dim = 8
+            cfg.model.num_hidden_layers = 2
+            cfg.model.layer_scaling_factor = 3.0
+            cfg.model.dropout_rate = 0.20
+            cfg.model.gamma = 2.0
+            # Train
+            cfg.train.batch_size = 128
+            cfg.train.learning_rate = 8e-4
+            cfg.train.early_stop_gen = 15
+            cfg.train.min_epochs = 50
+            cfg.train.max_epochs = 600
+            cfg.train.weights_beta = 0.9999
+            cfg.train.weights_max_ratio = 2.0
+            # Tuning
+            cfg.tune.enabled = True
+            cfg.tune.fast = True
+            cfg.tune.n_trials = 75
+            cfg.tune.epochs = 300
+            cfg.tune.batch_size = 128
+            cfg.tune.max_samples = 2048
+            cfg.tune.max_loci = 0
+            cfg.tune.eval_interval = 20
+            cfg.tune.infer_epochs = 40
+            cfg.tune.patience = 10
+            cfg.tune.proxy_metric_batch = 0
+            # Eval
+            cfg.evaluate.eval_latent_steps = 30
+
+        else:  # thorough
+            # Model
+            cfg.model.latent_dim = 16
+            cfg.model.num_hidden_layers = 3
+            cfg.model.layer_scaling_factor = 5.0
+            cfg.model.dropout_rate = 0.30
+            cfg.model.gamma = 2.5
+            # Train
+            cfg.train.batch_size = 64
+            cfg.train.learning_rate = 6e-4
+            cfg.train.early_stop_gen = 20  # Reduced from 30
+            cfg.train.min_epochs = 100
+            cfg.train.max_epochs = 800  # Reduced from 1200
+            cfg.train.weights_beta = 0.9999
+            cfg.train.weights_max_ratio = 2.0
+            # Tuning
+            cfg.tune.enabled = True
+            cfg.tune.fast = False
+            cfg.tune.n_trials = 150
+            cfg.tune.epochs = 600
+            cfg.tune.batch_size = 64
+            cfg.tune.max_samples = 5000  # Capped from 0
+            cfg.tune.max_loci = 0
+            cfg.tune.eval_interval = 10
+            cfg.tune.infer_epochs = 80
+            cfg.tune.patience = 15  # Reduced from 20
+            cfg.tune.proxy_metric_batch = 0
+            # Eval
+            cfg.evaluate.eval_latent_steps = 50
+
+        return cfg
+
+    def apply_overrides(self, overrides: Dict[str, Any] | None) -> "NLPCAConfig":
+        """Apply flat dot-key overrides."""
+        if not overrides:
+            return self
+        for k, v in overrides.items():
+            node = self
+            parts = k.split(".")
+            for p in parts[:-1]:
+                node = getattr(node, p)
+            last = parts[-1]
+            if hasattr(node, last):
+                setattr(node, last, v)
+            else:
+                raise KeyError(f"Unknown config key: {k}")
+        return self
+
+    def to_dict(self) -> Dict[str, Any]:
+        return asdict(self)
+
+
+@dataclass
+class UBPConfig:
+    """Top-level configuration for ImputeUBP.
+
+    Attributes:
+        io (IOConfig): I/O configuration.
+        model (ModelConfig): Model architecture configuration.
+        train (TrainConfig): Training procedure configuration.
+        tune (TuneConfig): Hyperparameter tuning configuration.
+        evaluate (EvalConfig): Evaluation configuration.
+        plot (PlotConfig): Plotting configuration.
+        sim (SimConfig): Simulated-missing configuration.
+    """
+
+    io: IOConfig = field(default_factory=IOConfig)
+    model: ModelConfig = field(default_factory=ModelConfig)
+    train: TrainConfig = field(default_factory=TrainConfig)
+    tune: TuneConfig = field(default_factory=TuneConfig)
+    evaluate: EvalConfig = field(default_factory=EvalConfig)
+    plot: PlotConfig = field(default_factory=PlotConfig)
+    sim: SimConfig = field(default_factory=SimConfig)
+
+    @classmethod
+    def from_preset(
+        cls, preset: Literal["fast", "balanced", "thorough"] = "balanced"
+    ) -> "UBPConfig":
+        """Build a UBPConfig from a named preset."""
+        if preset not in {"fast", "balanced", "thorough"}:
+            raise ValueError(f"Unknown preset: {preset}")
+
+        cfg = cls()
+
+        # Common baselines
+        cfg.io.verbose = False
+        cfg.model.hidden_activation = "relu"
+        cfg.model.layer_schedule = "pyramid"
+        cfg.model.latent_init = "random"
+        cfg.sim.simulate_missing = True
+        cfg.sim.sim_strategy = "random"
+        cfg.sim.sim_prop = 0.2
+
+        if preset == "fast":
+            # Model
+            cfg.model.latent_dim = 4
+            cfg.model.num_hidden_layers = 1
+            cfg.model.layer_scaling_factor = 2.0
+            cfg.model.dropout_rate = 0.10
+            cfg.model.gamma = 1.5
+            # Train
+            cfg.train.batch_size = 128
+            cfg.train.learning_rate = 1e-3
+            cfg.train.early_stop_gen = 5
+            cfg.train.min_epochs = 10
+            cfg.train.max_epochs = 120
+            cfg.train.weights_beta = 0.9999
+            cfg.train.weights_max_ratio = 2.0
+            # Tuning
+            cfg.tune.enabled = True
+            cfg.tune.fast = True
+            cfg.tune.n_trials = 25
+            cfg.tune.epochs = 120
+            cfg.tune.batch_size = 128
+            cfg.tune.max_samples = 512
+            cfg.tune.max_loci = 0
+            cfg.tune.eval_interval = 20
+            cfg.tune.infer_epochs = 20
+            cfg.tune.patience = 5
+            cfg.tune.proxy_metric_batch = 0
+            # Eval
+            cfg.evaluate.eval_latent_steps = 20
+            cfg.evaluate.eval_latent_lr = 1e-2
+            cfg.evaluate.eval_latent_weight_decay = 0.0
+
+        elif preset == "balanced":
+            # Model
+            cfg.model.latent_dim = 8
+            cfg.model.num_hidden_layers = 2
+            cfg.model.layer_scaling_factor = 3.0
+            cfg.model.dropout_rate = 0.20
+            cfg.model.gamma = 2.0
+            # Train
+            cfg.train.batch_size = 128
+            cfg.train.learning_rate = 8e-4
+            cfg.train.early_stop_gen = 15
+            cfg.train.min_epochs = 50
+            cfg.train.max_epochs = 600
+            cfg.train.weights_beta = 0.9999
+            cfg.train.weights_max_ratio = 2.0
+            # Tuning
+            cfg.tune.enabled = True
+            cfg.tune.fast = True
+            cfg.tune.n_trials = 75
+            cfg.tune.epochs = 300
+            cfg.tune.batch_size = 128
+            cfg.tune.max_samples = 2048
+            cfg.tune.max_loci = 0
+            cfg.tune.eval_interval = 20
+            cfg.tune.infer_epochs = 40
+            cfg.tune.patience = 10
+            cfg.tune.proxy_metric_batch = 0
+            # Eval
+            cfg.evaluate.eval_latent_steps = 30
+            cfg.evaluate.eval_latent_lr = 1e-2
+            cfg.evaluate.eval_latent_weight_decay = 0.0
+
+        else:  # thorough
+            # Model
+            cfg.model.latent_dim = 16
+            cfg.model.num_hidden_layers = 3
+            cfg.model.layer_scaling_factor = 5.0
+            cfg.model.dropout_rate = 0.30
+            cfg.model.gamma = 2.5
+            # Train
+            cfg.train.batch_size = 64
+            cfg.train.learning_rate = 6e-4
+            cfg.train.early_stop_gen = 20  # Reduced from 30
+            cfg.train.min_epochs = 100
+            cfg.train.max_epochs = 800  # Reduced from 1200
+            cfg.train.weights_beta = 0.9999
+            cfg.train.weights_max_ratio = 2.0
+            # Tuning
+            cfg.tune.enabled = True
+            cfg.tune.fast = False
+            cfg.tune.n_trials = 150
+            cfg.tune.epochs = 600
+            cfg.tune.batch_size = 64
+            cfg.tune.max_samples = 5000  # Capped from 0
+            cfg.tune.max_loci = 0
+            cfg.tune.eval_interval = 10
+            cfg.tune.infer_epochs = 80
+            cfg.tune.patience = 15  # Reduced from 20
+            cfg.tune.proxy_metric_batch = 0
+            # Eval
+            cfg.evaluate.eval_latent_steps = 50
+            cfg.evaluate.eval_latent_lr = 1e-2
+            cfg.evaluate.eval_latent_weight_decay = 0.0
+
+        return cfg
+
+    def apply_overrides(self, overrides: Dict[str, Any] | None) -> "UBPConfig":
+        """Apply flat dot-key overrides."""
+        if not overrides:
+            return self
+
+        for k, v in overrides.items():
+            node = self
+            parts = k.split(".")
+            for p in parts[:-1]:
+                node = getattr(node, p)
+            last = parts[-1]
+            if hasattr(node, last):
+                setattr(node, last, v)
+            else:
+                raise KeyError(f"Unknown config key: {k}")
+        return self
+
+    def to_dict(self) -> Dict[str, Any]:
+        return asdict(self)
+
+
+@dataclass
+class AutoencoderConfig:
+    """Top-level configuration for ImputeAutoencoder.
+
+    Attributes:
+        io (IOConfig): I/O configuration.
+        model (ModelConfig): Model architecture configuration.
+        train (TrainConfig): Training procedure configuration.
+        tune (TuneConfig): Hyperparameter tuning configuration.
+        evaluate (EvalConfig): Evaluation configuration.
+        plot (PlotConfig): Plotting configuration.
+        sim (SimConfig): Simulated-missing configuration.
+    """
+
+    io: IOConfig = field(default_factory=IOConfig)
+    model: ModelConfig = field(default_factory=ModelConfig)
+    train: TrainConfig = field(default_factory=TrainConfig)
+    tune: TuneConfig = field(default_factory=TuneConfig)
+    evaluate: EvalConfig = field(default_factory=EvalConfig)
+    plot: PlotConfig = field(default_factory=PlotConfig)
+    sim: SimConfig = field(default_factory=SimConfig)
+
+    @classmethod
+    def from_preset(
+        cls, preset: Literal["fast", "balanced", "thorough"] = "balanced"
+    ) -> "AutoencoderConfig":
+        """Build a AutoencoderConfig from a named preset."""
+        if preset not in {"fast", "balanced", "thorough"}:
+            raise ValueError(f"Unknown preset: {preset}")
+
+        cfg = cls()
+
+        # Common baselines (no latent refinement at eval)
+        cfg.io.verbose = False
+        cfg.train.validation_split = 0.20
+        cfg.model.hidden_activation = "relu"
+        cfg.model.layer_schedule = "pyramid"
+        cfg.evaluate.eval_latent_steps = 0
+        cfg.evaluate.eval_latent_lr = 0.0
+        cfg.evaluate.eval_latent_weight_decay = 0.0
+        cfg.sim.simulate_missing = True
+        cfg.sim.sim_strategy = "random"
+        cfg.sim.sim_prop = 0.2
+
+        if preset == "fast":
+            cfg.model.latent_dim = 4
+            cfg.model.num_hidden_layers = 1
+            cfg.model.layer_scaling_factor = 2.0
+            cfg.model.dropout_rate = 0.10
+            cfg.model.gamma = 1.5
+            cfg.train.batch_size = 128
+            cfg.train.learning_rate = 1e-3
+            cfg.train.early_stop_gen = 5
+            cfg.train.min_epochs = 10
+            cfg.train.max_epochs = 120
+            cfg.train.weights_beta = 0.9999
+            cfg.train.weights_max_ratio = 2.0
+            cfg.tune.enabled = True
+            cfg.tune.fast = True
+            cfg.tune.n_trials = 25
+            cfg.tune.epochs = 120
+            cfg.tune.batch_size = 128
+            cfg.tune.max_samples = 512
+            cfg.tune.max_loci = 0
+            cfg.tune.eval_interval = 20
+            cfg.tune.patience = 5
+            cfg.tune.proxy_metric_batch = 0
+            if hasattr(cfg.tune, "infer_epochs"):
+                cfg.tune.infer_epochs = 0
+
+        elif preset == "balanced":
+            cfg.model.latent_dim = 8
+            cfg.model.num_hidden_layers = 2
+            cfg.model.layer_scaling_factor = 3.0
+            cfg.model.dropout_rate = 0.20
+            cfg.model.gamma = 2.0
+            cfg.train.batch_size = 128
+            cfg.train.learning_rate = 8e-4
+            cfg.train.early_stop_gen = 15
+            cfg.train.min_epochs = 50
+            cfg.train.max_epochs = 600
+            cfg.train.weights_beta = 0.9999
+            cfg.train.weights_max_ratio = 2.0
+            cfg.tune.enabled = True
+            cfg.tune.fast = True
+            cfg.tune.n_trials = 75
+            cfg.tune.epochs = 300
+            cfg.tune.batch_size = 128
+            cfg.tune.max_samples = 2048
+            cfg.tune.max_loci = 0
+            cfg.tune.eval_interval = 20
+            cfg.tune.patience = 10
+            cfg.tune.proxy_metric_batch = 0
+            if hasattr(cfg.tune, "infer_epochs"):
+                cfg.tune.infer_epochs = 0
+
+        else:  # thorough
+            cfg.model.latent_dim = 16
+            cfg.model.num_hidden_layers = 3
+            cfg.model.layer_scaling_factor = 5.0
+            cfg.model.dropout_rate = 0.30
+            cfg.model.gamma = 2.5
+            cfg.train.batch_size = 64
+            cfg.train.learning_rate = 6e-4
+            cfg.train.early_stop_gen = 20  # Reduced from 30
+            cfg.train.min_epochs = 100
+            cfg.train.max_epochs = 800  # Reduced from 1200
+            cfg.train.weights_beta = 0.9999
+            cfg.train.weights_max_ratio = 2.0
+            cfg.tune.enabled = True
+            cfg.tune.fast = False
+            cfg.tune.n_trials = 150
+            cfg.tune.epochs = 600
+            cfg.tune.batch_size = 64
+            cfg.tune.max_samples = 5000  # Capped from 0
+            cfg.tune.max_loci = 0
+            cfg.tune.eval_interval = 10
+            cfg.tune.patience = 15  # Reduced from 20
+            cfg.tune.proxy_metric_batch = 0
+            if hasattr(cfg.tune, "infer_epochs"):
+                cfg.tune.infer_epochs = 0
+
+        return cfg
+
+    def apply_overrides(self, overrides: Dict[str, Any] | None) -> "AutoencoderConfig":
+        """Apply flat dot-key overrides."""
+        if not overrides:
+            return self
+        for k, v in overrides.items():
+            node = self
+            parts = k.split(".")
+            for p in parts[:-1]:
+                node = getattr(node, p)
+            last = parts[-1]
+            if hasattr(node, last):
+                setattr(node, last, v)
+            else:
+                raise KeyError(f"Unknown config key: {k}")
+        return self
+
+    def to_dict(self) -> Dict[str, Any]:
+        return asdict(self)
+
+
+@dataclass
+class VAEExtraConfig:
+    """VAE-specific knobs.
+
+    Attributes:
+        kl_beta (float): Final β for KL divergence term.
+        kl_warmup (int): Number of epochs with β=0 (warm-up period).
+        kl_ramp (int): Number of epochs for linear ramp to final β.
+    """
+
+    kl_beta: float = 1.0
+    kl_warmup: int = 50
+    kl_ramp: int = 200
+
+
+@dataclass
+class VAEConfig:
+    """Top-level configuration for ImputeVAE (AE-parity + VAE extras).
+
+    Attributes:
+        io (IOConfig): I/O configuration.
+        model (ModelConfig): Model architecture configuration.
+        train (TrainConfig): Training procedure configuration.
+        tune (TuneConfig): Hyperparameter tuning configuration.
+        evaluate (EvalConfig): Evaluation configuration.
+        plot (PlotConfig): Plotting configuration.
+        vae (VAEExtraConfig): VAE-specific configuration.
+        sim (SimConfig): Simulated-missing configuration.
+    """
+
+    io: IOConfig = field(default_factory=IOConfig)
+    model: ModelConfig = field(default_factory=ModelConfig)
+    train: TrainConfig = field(default_factory=TrainConfig)
+    tune: TuneConfig = field(default_factory=TuneConfig)
+    evaluate: EvalConfig = field(default_factory=EvalConfig)
+    plot: PlotConfig = field(default_factory=PlotConfig)
+    vae: VAEExtraConfig = field(default_factory=VAEExtraConfig)
+    sim: SimConfig = field(default_factory=SimConfig)
+
+    @classmethod
+    def from_preset(
+        cls, preset: Literal["fast", "balanced", "thorough"] = "balanced"
+    ) -> "VAEConfig":
+        """Build a VAEConfig from a named preset."""
+        if preset not in {"fast", "balanced", "thorough"}:
+            raise ValueError(f"Unknown preset: {preset}")
+
+        cfg = cls()
+
+        # Common baselines (match AE; no latent refinement at eval)
+        cfg.io.verbose = False
+        cfg.train.validation_split = 0.20
+        cfg.model.hidden_activation = "relu"
+        cfg.model.layer_schedule = "pyramid"
+        cfg.evaluate.eval_latent_steps = 0
+        cfg.evaluate.eval_latent_lr = 0.0
+        cfg.evaluate.eval_latent_weight_decay = 0.0
+        cfg.sim.simulate_missing = True
+        cfg.sim.sim_strategy = "random"
+        cfg.sim.sim_prop = 0.2
+
+        # VAE KL schedules, shortened for speed
+        cfg.vae.kl_beta = 1.0
+        cfg.vae.kl_warmup = 25
+        cfg.vae.kl_ramp = 100
+
+        if preset == "fast":
+            cfg.model.latent_dim = 4
+            cfg.model.num_hidden_layers = 1
+            cfg.model.layer_scaling_factor = 2.0
+            cfg.model.dropout_rate = 0.10
+            cfg.model.gamma = 1.5
+            cfg.vae.kl_beta = 0.5  # Lower beta for fast training
+            cfg.train.batch_size = 128
+            cfg.train.learning_rate = 1e-3
+            cfg.train.early_stop_gen = 5
+            cfg.train.min_epochs = 10
+            cfg.train.max_epochs = 120
+            cfg.train.weights_beta = 0.9999
+            cfg.train.weights_max_ratio = 2.0
+            cfg.tune.enabled = True
+            cfg.tune.fast = True
+            cfg.tune.n_trials = 25
+            cfg.tune.epochs = 120
+            cfg.tune.batch_size = 128
+            cfg.tune.max_samples = 512
+            cfg.tune.max_loci = 0
+            cfg.tune.eval_interval = 20
+            cfg.tune.patience = 5
+            cfg.tune.proxy_metric_batch = 0
+            if hasattr(cfg.tune, "infer_epochs"):
+                cfg.tune.infer_epochs = 0
+
+        elif preset == "balanced":
+            cfg.model.latent_dim = 8
+            cfg.model.num_hidden_layers = 2
+            cfg.model.layer_scaling_factor = 3.0
+            cfg.model.dropout_rate = 0.20
+            cfg.model.gamma = 2.0
+            cfg.train.batch_size = 128
+            cfg.train.learning_rate = 8e-4
+            cfg.train.early_stop_gen = 15
+            cfg.train.min_epochs = 50
+            cfg.train.max_epochs = 600
+            cfg.train.weights_beta = 0.9999
+            cfg.train.weights_max_ratio = 2.0
+            cfg.tune.enabled = True
+            cfg.tune.fast = True
+            cfg.tune.n_trials = 75
+            cfg.tune.epochs = 300
+            cfg.tune.batch_size = 128
+            cfg.tune.max_samples = 2048
+            cfg.tune.max_loci = 0
+            cfg.tune.eval_interval = 20
+            cfg.tune.patience = 10
+            cfg.tune.proxy_metric_batch = 0
+            if hasattr(cfg.tune, "infer_epochs"):
+                cfg.tune.infer_epochs = 0
+
+        else:  # thorough
+            cfg.model.latent_dim = 16
+            cfg.model.num_hidden_layers = 3
+            cfg.model.layer_scaling_factor = 5.0
+            cfg.model.dropout_rate = 0.30
+            cfg.model.gamma = 2.5
+            cfg.train.batch_size = 64
+            cfg.train.learning_rate = 6e-4
+            cfg.train.early_stop_gen = 20  # Reduced from 30
+            cfg.train.min_epochs = 100
+            cfg.train.max_epochs = 800  # Reduced from 1200
+            cfg.train.weights_beta = 0.9999
+            cfg.train.weights_max_ratio = 2.0
+            cfg.tune.enabled = True
+            cfg.tune.fast = False
+            cfg.tune.n_trials = 150
+            cfg.tune.epochs = 600
+            cfg.tune.batch_size = 64
+            cfg.tune.max_samples = 5000  # Capped from 0
+            cfg.tune.max_loci = 0
+            cfg.tune.eval_interval = 10
+            cfg.tune.patience = 15  # Reduced from 20
+            cfg.tune.proxy_metric_batch = 0
+            if hasattr(cfg.tune, "infer_epochs"):
+                cfg.tune.infer_epochs = 0
+
+        return cfg
+
+    def apply_overrides(self, overrides: Dict[str, Any] | None) -> "VAEConfig":
+        """Apply flat dot-key overrides."""
+        if not overrides:
+            return self
+        for k, v in overrides.items():
+            node = self
+            parts = k.split(".")
+            for p in parts[:-1]:
+                node = getattr(node, p)
+            last = parts[-1]
+            if hasattr(node, last):
+                setattr(node, last, v)
+            else:
+                raise KeyError(f"Unknown config key: {k}")
+        return self
+
+    def to_dict(self) -> Dict[str, Any]:
+        return asdict(self)
+
+
+@dataclass
+class MostFrequentAlgoConfig:
+    """Algorithmic knobs for ImputeMostFrequent.
+
+    Attributes:
+        by_populations (bool): Whether to compute per-population modes.
+        default (int): Fallback mode if no valid entries in a locus.
+        missing (int): Code for missing genotypes in 0/1/2.
+    """
+
+    by_populations: bool = False
+    default: int = 0
+    missing: int = -1
+
+
+@dataclass
+class DeterministicSplitConfig:
+    """Evaluation split configuration shared by deterministic imputers.
+
+    Attributes:
+        test_size (float): Proportion of data to use as the test set.
+        test_indices (Optional[Sequence[int]]): Specific indices to use as the test set.
+    """
+
+    test_size: float = 0.2
+    test_indices: Optional[Sequence[int]] = None
+
+
+@dataclass
+class MostFrequentConfig:
+    """Top-level configuration for ImputeMostFrequent.
+
+    Attributes:
+        io (IOConfig): I/O configuration.
+        plot (PlotConfig): Plotting configuration.
+        split (DeterministicSplitConfig): Data splitting configuration.
+        algo (MostFrequentAlgoConfig): Algorithmic configuration.
+        sim (SimConfig): Simulation configuration.
+        tune (TuneConfig): Hyperparameter tuning configuration.
+        train (TrainConfig): Training configuration.
+    """
+
+    io: IOConfig = field(default_factory=IOConfig)
+    plot: PlotConfig = field(default_factory=PlotConfig)
+    split: DeterministicSplitConfig = field(default_factory=DeterministicSplitConfig)
+    algo: MostFrequentAlgoConfig = field(default_factory=MostFrequentAlgoConfig)
+    sim: SimConfig = field(default_factory=SimConfig)
+    tune: TuneConfig = field(default_factory=TuneConfig)
+    train: TrainConfig = field(default_factory=TrainConfig)
+
+    @classmethod
+    def from_preset(
+        cls,
+        preset: Literal["fast", "balanced", "thorough"] = "balanced",
+    ) -> "MostFrequentConfig":
+        """Construct a preset configuration."""
+        if preset not in {"fast", "balanced", "thorough"}:
+            raise ValueError(f"Unknown preset: {preset}")
+
+        cfg = cls()
+        cfg.io.verbose = False
+        cfg.split.test_size = 0.2
+        cfg.sim.simulate_missing = True
+        cfg.sim.sim_strategy = "random"
+        cfg.sim.sim_prop = 0.2
+
+        return cfg
+
+    def apply_overrides(self, overrides: Dict[str, Any] | None) -> "MostFrequentConfig":
+        """Apply dot-key overrides."""
+        if not overrides:
+            return self
+        for k, v in overrides.items():
+            node = self
+            parts = k.split(".")
+            for p in parts[:-1]:
+                node = getattr(node, p)
+            last = parts[-1]
+            if hasattr(node, last):
+                setattr(node, last, v)
+            else:
+                pass
+        return self
+
+    def to_dict(self) -> Dict[str, Any]:
+        return asdict(self)
+
+
+@dataclass
+class RefAlleleAlgoConfig:
+    """Algorithmic knobs for ImputeRefAllele.
+
+    Attributes:
+        missing (int): Code for missing genotypes in 0/1/2.
+    """
+
+    missing: int = -1
+
+
+@dataclass
+class RefAlleleConfig:
+    """Top-level configuration for ImputeRefAllele.
+
+    Attributes:
+        io (IOConfig): I/O configuration.
+        plot (PlotConfig): Plotting configuration.
+        split (DeterministicSplitConfig): Data splitting configuration.
+        algo (RefAlleleAlgoConfig): Algorithmic configuration.
+        sim (SimConfig): Simulation configuration.
+        tune (TuneConfig): Hyperparameter tuning configuration.
+        train (TrainConfig): Training configuration.
+    """
+
+    io: IOConfig = field(default_factory=IOConfig)
+    plot: PlotConfig = field(default_factory=PlotConfig)
+    split: DeterministicSplitConfig = field(default_factory=DeterministicSplitConfig)
+    algo: RefAlleleAlgoConfig = field(default_factory=RefAlleleAlgoConfig)
+    sim: SimConfig = field(default_factory=SimConfig)
+    tune: TuneConfig = field(default_factory=TuneConfig)
+    train: TrainConfig = field(default_factory=TrainConfig)
+
+    @classmethod
+    def from_preset(
+        cls, preset: Literal["fast", "balanced", "thorough"] = "balanced"
+    ) -> "RefAlleleConfig":
+        """Presets mainly keep parity with logging/IO and split test_size."""
+        if preset not in {"fast", "balanced", "thorough"}:
+            raise ValueError(f"Unknown preset: {preset}")
+
+        cfg = cls()
+        cfg.io.verbose = False
+        cfg.split.test_size = 0.2
+        cfg.sim.simulate_missing = True
+        cfg.sim.sim_strategy = "random"
+        cfg.sim.sim_prop = 0.2
+        return cfg
+
+    def apply_overrides(self, overrides: Dict[str, Any] | None) -> "RefAlleleConfig":
+        """Apply dot-key overrides."""
+        if not overrides:
+            return self
+        for k, v in overrides.items():
+            node = self
+            parts = k.split(".")
+            for p in parts[:-1]:
+                node = getattr(node, p)
+            last = parts[-1]
+            if hasattr(node, last):
+                setattr(node, last, v)
+            else:
+                pass
+        return self
+
+    def to_dict(self) -> Dict[str, Any]:
+        return asdict(self)
+
+
+def _flatten_dict(
+    d: Dict[str, Any], prefix: str = "", out: Optional[Dict[str, Any]] = None
+) -> Dict[str, Any]:
+    """Flatten a nested dictionary into dot-key format."""
+    out = out or {}
+    for k, v in d.items():
+        kk = f"{prefix}.{k}" if prefix else k
+        if isinstance(v, dict):
+            _flatten_dict(v, kk, out)
+        else:
+            out[kk] = v
+    return out
+
+
+@dataclass
+class IOConfigSupervised:
+    """I/O, logging, and run identity.
+
+    Attributes:
+        prefix (str): Prefix for output files and logs.
+        seed (Optional[int]): Random seed for reproducibility.
+        n_jobs (int): Number of parallel jobs to use.
+        verbose (bool): Whether to enable verbose logging.
+        debug (bool): Whether to enable debug mode.
+    """
+
+    prefix: str = "pgsui"
+    seed: Optional[int] = None
+    n_jobs: int = 1
+    verbose: bool = False
+    debug: bool = False
+
+
+@dataclass
+class PlotConfigSupervised:
+    """Plot/figure styling.
+
+    Attributes:
+        fmt (Literal["pdf", "png", "jpg", "jpeg"]): File format.
+        dpi (int): Resolution in dots per inch.
+        fontsize (int): Base font size for plot text.
+        despine (bool): Whether to remove top/right spines.
+        show (bool): Whether to display plots interactively.
+    """
+
+    fmt: Literal["pdf", "png", "jpg", "jpeg"] = "pdf"
+    dpi: int = 300
+    fontsize: int = 18
+    despine: bool = True
+    show: bool = False
+
+
+@dataclass
+class TrainConfigSupervised:
+    """Training/evaluation split (by samples).
+
+    Attributes:
+        validation_split (float): Proportion of data to use for validation.
+    """
+
+    validation_split: float = 0.20
+
+    def __post_init__(self):
+        if not (0.0 < self.validation_split < 1.0):
+            raise ValueError("validation_split must be between 0.0 and 1.0")
+
+
+@dataclass
+class ImputerConfigSupervised:
+    """IterativeImputer-like scaffolding used by current supervised wrappers.
+
+    Attributes:
+        n_nearest_features (Optional[int]): Number of nearest features to use.
+        max_iter (int): Maximum number of imputation iterations to perform.
+    """
+
+    n_nearest_features: Optional[int] = 10
+    max_iter: int = 10
+
+
+@dataclass
+class SimConfigSupervised:
+    """Simulation of missingness for evaluation.
+
+    Attributes:
+        prop_missing (float): Proportion of features to set as missing.
+        strategy (Literal["random", "random_inv_genotype"]): Strategy.
+        het_boost (float): Boosting factor for heterogeneity.
+        missing_val (int): Internal code for missing genotypes.
+    """
+
+    prop_missing: float = 0.5
+    strategy: Literal["random", "random_inv_genotype"] = "random_inv_genotype"
+    het_boost: float = 2.0
+    missing_val: int = -1
+
+
+@dataclass
+class TuningConfigSupervised:
+    """Optuna tuning envelope."""
+
+    enabled: bool = True
+    n_trials: int = 100
+    metric: str = "pr_macro"
+    n_jobs: int = 8
+    fast: bool = True
+
+
+@dataclass
+class RFModelConfig:
+    """Random Forest hyperparameters.
+
+    Attributes:
+        n_estimators (int): Number of trees in the forest.
+        max_depth (Optional[int]): Maximum depth of the trees.
+        min_samples_split (int): Minimum number of samples required to split.
+        min_samples_leaf (int): Minimum number of samples required at a leaf.
+        max_features (Literal["sqrt", "log2"] | float | None): Features to consider.
+        criterion (Literal["gini", "entropy", "log_loss"]): Split quality metric.
+        class_weight (Literal["balanced", "balanced_subsample", None]): Class weights.
+    """
+
+    n_estimators: int = 100
+    max_depth: Optional[int] = None
+    min_samples_split: int = 2
+    min_samples_leaf: int = 1
+    max_features: Literal["sqrt", "log2"] | float | None = "sqrt"
+    criterion: Literal["gini", "entropy", "log_loss"] = "gini"
+    class_weight: Literal["balanced", "balanced_subsample", None] = "balanced"
+
+
+@dataclass
+class HGBModelConfig:
+    """Histogram-based Gradient Boosting hyperparameters.
+
+    Attributes:
+        n_estimators (int): Number of boosting iterations (max_iter).
+        learning_rate (float): Step size for each boosting iteration.
+        max_depth (Optional[int]): Maximum depth of each tree.
+        min_samples_leaf (int): Minimum number of samples required at a leaf.
+        max_features (float | None): Proportion of features to consider.
+        n_iter_no_change (int): Iterations to wait for early stopping.
+        tol (float): Minimum improvement in the loss.
+    """
+
+    n_estimators: int = 100  # maps to max_iter
+    learning_rate: float = 0.1
+    max_depth: Optional[int] = None
+    min_samples_leaf: int = 1
+    max_features: float | None = 1.0
+    n_iter_no_change: int = 10
+    tol: float = 1e-7
+
+    def __post_init__(self) -> None:
+        if isinstance(self.max_features, float):
+            if not (0.0 < self.max_features <= 1.0):
+                raise ValueError("max_features as float must be in (0.0, 1.0]")
+
+        if self.n_estimators <= 0:
+            raise ValueError("n_estimators must be a positive integer")
+
+
+@dataclass
+class RFConfig:
+    """Configuration for ImputeRandomForest.
+
+    Attributes:
+        io (IOConfigSupervised): Run identity, logging, and seeds.
+        model (RFModelConfig): RandomForest hyperparameters.
+        train (TrainConfigSupervised): Sample split for validation.
+        imputer (ImputerConfigSupervised): IterativeImputer scaffolding.
+        sim (SimConfigSupervised): Simulated missingness.
+        plot (PlotConfigSupervised): Plot styling.
+        tune (TuningConfigSupervised): Optuna knobs.
+    """
+
+    io: IOConfigSupervised = field(default_factory=IOConfigSupervised)
+    model: RFModelConfig = field(default_factory=RFModelConfig)
+    train: TrainConfigSupervised = field(default_factory=TrainConfigSupervised)
+    imputer: ImputerConfigSupervised = field(default_factory=ImputerConfigSupervised)
+    sim: SimConfigSupervised = field(default_factory=SimConfigSupervised)
+    plot: PlotConfigSupervised = field(default_factory=PlotConfigSupervised)
+    tune: TuningConfigSupervised = field(default_factory=TuningConfigSupervised)
+
+    @classmethod
+    def from_preset(cls, preset: str = "balanced") -> "RFConfig":
+        """Build a config from a named preset."""
+        cfg = cls()
+        if preset == "fast":
+            cfg.model.n_estimators = 100  # Increased from 50
+            cfg.model.max_depth = None
+            cfg.imputer.max_iter = 5
+            cfg.io.n_jobs = 1
+            cfg.tune.enabled = False
+        elif preset == "balanced":
+            cfg.model.n_estimators = 200  # Increased from 100
+            cfg.model.max_depth = None
+            cfg.imputer.max_iter = 10
+            cfg.io.n_jobs = 1
+            cfg.tune.enabled = False
+            cfg.tune.n_trials = 100
+        elif preset == "thorough":
+            cfg.model.n_estimators = 500
+            cfg.model.max_depth = 50  # Added safety cap
+            cfg.imputer.max_iter = 15
+            cfg.io.n_jobs = 1
+            cfg.tune.enabled = False
+            cfg.tune.n_trials = 250
+        else:
+            raise ValueError(f"Unknown preset: {preset}")
+
+        return cfg
+
+    @classmethod
+    def from_yaml(cls, path: str) -> "RFConfig":
+        """Load from YAML; honors optional top-level 'preset'."""
+        return load_yaml_to_dataclass(path, cls)
+
+    def apply_overrides(self, overrides: Dict[str, Any] | None) -> "RFConfig":
+        """Apply flat dot-key overrides."""
+        if overrides:
+            apply_dot_overrides(self, overrides)
+        return self
+
+    def to_dict(self) -> Dict[str, Any]:
+        return asdict(self)
+
+    def to_imputer_kwargs(self) -> Dict[str, Any]:
+        return {
+            "prefix": self.io.prefix,
+            "seed": self.io.seed,
+            "n_jobs": self.io.n_jobs,
+            "verbose": self.io.verbose,
+            "debug": self.io.debug,
+            "model_n_estimators": self.model.n_estimators,
+            "model_max_depth": self.model.max_depth,
+            "model_min_samples_split": self.model.min_samples_split,
+            "model_min_samples_leaf": self.model.min_samples_leaf,
+            "model_max_features": self.model.max_features,
+            "model_criterion": self.model.criterion,
+            "model_validation_split": self.train.validation_split,
+            "model_n_nearest_features": self.imputer.n_nearest_features,
+            "model_max_iter": self.imputer.max_iter,
+            "sim_prop_missing": self.sim.prop_missing,
+            "sim_strategy": self.sim.strategy,
+            "sim_het_boost": self.sim.het_boost,
+            "plot_format": self.plot.fmt,
+            "plot_fontsize": self.plot.fontsize,
+            "plot_despine": self.plot.despine,
+            "plot_dpi": self.plot.dpi,
+            "plot_show_plots": self.plot.show,
+        }
+
+
+@dataclass
+class HGBConfig:
+    """Configuration for ImputeHistGradientBoosting.
+
+    Attributes:
+        io (IOConfigSupervised): Run identity, logging, and seeds.
+        model (HGBModelConfig): HistGradientBoosting hyperparameters.
+        train (TrainConfigSupervised): Sample split for validation.
+        imputer (ImputerConfigSupervised): IterativeImputer scaffolding.
+        sim (SimConfigSupervised): Simulated missingness.
+        plot (PlotConfigSupervised): Plot styling.
+        tune (TuningConfigSupervised): Optuna knobs.
+    """
+
+    io: IOConfigSupervised = field(default_factory=IOConfigSupervised)
+    model: HGBModelConfig = field(default_factory=HGBModelConfig)
+    train: TrainConfigSupervised = field(default_factory=TrainConfigSupervised)
+    imputer: ImputerConfigSupervised = field(default_factory=ImputerConfigSupervised)
+    sim: SimConfigSupervised = field(default_factory=SimConfigSupervised)
+    plot: PlotConfigSupervised = field(default_factory=PlotConfigSupervised)
+    tune: TuningConfigSupervised = field(default_factory=TuningConfigSupervised)
+
+    @classmethod
+    def from_preset(cls, preset: str = "balanced") -> "HGBConfig":
+        """Build a config from a named preset."""
+        cfg = cls()
+        if preset == "fast":
+            cfg.model.n_estimators = 50
+            cfg.model.learning_rate = 0.15
+            cfg.model.max_depth = None
+            cfg.imputer.max_iter = 5
+            cfg.io.n_jobs = 1
+            cfg.tune.enabled = False
+            cfg.tune.n_trials = 50
+        elif preset == "balanced":
+            cfg.model.n_estimators = 100
+            cfg.model.learning_rate = 0.1
+            cfg.model.max_depth = None
+            cfg.imputer.max_iter = 10
+            cfg.io.n_jobs = 1
+            cfg.tune.enabled = False
+            cfg.tune.n_trials = 100
+        elif preset == "thorough":
+            cfg.model.n_estimators = 500
+            cfg.model.learning_rate = 0.05  # Reduced from 0.08
+            cfg.model.n_iter_no_change = 20  # Increased patience
+            cfg.model.max_depth = None
+            cfg.imputer.max_iter = 15
+            cfg.io.n_jobs = 1
+            cfg.tune.enabled = False
+            cfg.tune.n_trials = 250
+        else:
+            raise ValueError(f"Unknown preset: {preset}")
+        return cfg
+
+    @classmethod
+    def from_yaml(cls, path: str) -> "HGBConfig":
+        return load_yaml_to_dataclass(path, cls)
+
+    def apply_overrides(self, overrides: Dict[str, Any] | None) -> "HGBConfig":
+        if overrides:
+            apply_dot_overrides(self, overrides)
+        return self
+
+    def to_dict(self) -> Dict[str, Any]:
+        return asdict(self)
+
+    def to_imputer_kwargs(self) -> Dict[str, Any]:
+        return {
+            "prefix": self.io.prefix,
+            "seed": self.io.seed,
+            "n_jobs": self.io.n_jobs,
+            "verbose": self.io.verbose,
+            "debug": self.io.debug,
+            "model_n_estimators": self.model.n_estimators,
+            "model_learning_rate": self.model.learning_rate,
+            "model_n_iter_no_change": self.model.n_iter_no_change,
+            "model_tol": self.model.tol,
+            "model_max_depth": self.model.max_depth,
+            "model_min_samples_leaf": self.model.min_samples_leaf,
+            "model_max_features": self.model.max_features,
+            "model_validation_split": self.train.validation_split,
+            "model_n_nearest_features": self.imputer.n_nearest_features,
+            "model_max_iter": self.imputer.max_iter,
+            "sim_prop_missing": self.sim.prop_missing,
+            "sim_strategy": self.sim.strategy,
+            "sim_het_boost": self.sim.het_boost,
+            "plot_format": self.plot.fmt,
+            "plot_fontsize": self.plot.fontsize,
+            "plot_despine": self.plot.despine,
+            "plot_dpi": self.plot.dpi,
+            "plot_show_plots": self.plot.show,
+        }