Configuration

MELITE reads defaults from melite/config_default.toml. A user TOML file can override only the settings that need to change.

Minimal Override

[paths]
output = "my_output/"

[datasets.morgan_r2_2048]
path = "data/morgan_r2_2048.npz"
label_path = "raw/labels.npy"
family = "fingerprints"
method = "Morgan"
variant = "r2_2048"

[models]
active = ["svc", "rf"]

Use the file from the CLI:

melite run --config my_config.toml
melite export --config my_config.toml --row 0

Input Layout

MELITE consumes pre-computed feature matrices and labels:

raw/labels.npy          <- target vector y, shape (n_samples,)
data/morgan_r2_2048.npz <- required key: X, optional key: y
data/maccs.npz
data/rdkit_descriptors.npz
data/PCA85.npz
data/UMAP90.npz

Each .npz file must contain an X array. If an embedded y array is present, MELITE validates it against the configured label_path to avoid silent feature-label mismatches.

MELITE is tabular at the modeling level. The learning algorithms only consume numeric X and y arrays, so the feature matrix may come from PCA, UMAP, fingerprints, descriptors, clinical variables, experimental measurements, industrial features, or manually selected numeric features.

Each concrete matrix candidate is registered under [datasets.<dataset_id>]. Required fields are path and label_path. Optional metadata fields are family, method, variant, level, and description; they are preserved in reports for traceability and do not control special-case execution logic.

[datasets.morgan_r2_2048]
path = "data/morgan_r2_2048.npz"
label_path = "raw/labels.npy"
family = "fingerprints"
method = "Morgan"
variant = "r2_2048"

[datasets.maccs]
path = "data/maccs.npz"
label_path = "raw/labels.npy"
family = "fingerprints"
method = "MACCS"

[datasets.rdkit_descriptors]
path = "data/rdkit_descriptors.npz"
label_path = "raw/labels.npy"
family = "descriptors"
method = "RDKit"

[datasets.pca85]
path = "data/PCA85.npz"
label_path = "raw/labels.npy"
family = "dimensionality"
method = "PCA"
level = 85

[datasets.umap90]
path = "data/UMAP90.npz"
label_path = "raw/labels.npy"
family = "dimensionality"
method = "UMAP"
level = 90

Registered datasets are loaded strictly. A missing dataset file, missing label_path, missing X, non-2D X, non-numeric X, length mismatch, or embedded y mismatch raises an error instead of silently skipping the entry.

Legacy [benchmark].reduction_types and levels remain supported for compatibility. When [datasets] is absent, MELITE synthesizes entries such as PCA70 and UMAP90 with dimensionality metadata.

Outputs

By default, MELITE writes results under output/:

output/
|-- results.txt
|-- results.csv
|-- Model_<model>_<dataset>.pkl
`-- figures/
    `-- <model>_<dataset>.png

Output	Purpose
results.txt	Human-readable benchmark report.
results.csv	Structured rows with model performance, parameters, and smoke marker.
.pkl artifact	Final selected model retrained on all available data.
PNG figure	F1, Accuracy, and AUC-ROC cross-validation distributions.

Model Families and Hyperparameter Grids

config_default.toml controls which model families are active:

[models]
active = ["svc", "rf", "xgb"]

Remove a key from active to skip that model family in a run. The detailed hyperparameter grids are defined in melite/config.py; they are developer-facing defaults rather than user TOML settings.

Config key	Model family	Benchmark coverage
svc	Support Vector Classifier	Full runs evaluate polynomial and RBF kernels. Smoke mode uses a linear SVC only.
rf	Random Forest Classifier	Evaluates tree count, tree depth, feature sampling, and split/leaf controls.
xgb	XGBoost Classifier	Evaluates tree count, learning rate, depth, sampling, gamma, and regularization.

SVC Kernels

Mode	Kernels	Parameters
Full benchmark	poly, rbf	C and gamma; plus coef0 and degree for poly.
Smoke mode	linear	C = 1.

Current full-run SVC grids:

Kernel	Values
poly	C = [0.01, 0.1, 1, 10]; degree = [3, 4, 5]; coef0 = [0.0, 0.1, 0.02, 0.6, 0.8, 1]; gamma = [0.001, 0.002, 0.004, 0.008, 0.01, 0.02, 0.04, 0.08, 0.1, 0.2]
rbf	C = [0.01, 0.02, 0.1, 0.2, 1, 2, 10, 20]; gamma = [0.001, 0.002, 0.004, 0.008, 0.01, 0.02, 0.04, 0.08, 0.1, 0.2]

Tree-Based Models

Model	Main parameters explored
Random Forest	n_estimators, max_depth, max_features, min_samples_split, and min_samples_leaf.
XGBoost	n_estimators, learning_rate, max_depth, subsample, colsample_bytree, gamma, reg_alpha, and reg_lambda.

These grids can be restricted at the family level through [models].active. Changing the individual hyperparameter values currently requires editing melite/config.py.