Default parameters

This documents details the default parameters used in the ALMOS program.

CLUSTER

CLUSTER workflow.

This module prepares descriptors, applies cleanup, and selects representative molecules that cover the cleaned chemical descriptor space. Natural clustering diagnostics are optional and are not used by the default point-selection method.

Main user-facing parameters

General:

inputstr

Input CSV or SDF file used by the coverage-selection workflow.

namestr

Identifier column when descriptors are already present in the CSV.

ignorelist

Columns excluded from descriptor cleanup and coverage selection.

aqmebool

Generate descriptors with AQME before coverage selection.

ystr

Optional response column to ignore during coverage selection.

categoricalstr

Encoding mode for categorical descriptor columns.

Descriptor cleanup:

missing_thresholdfloat

Remove descriptor columns with too many missing values.

near_constant_thresholdfloat

Remove descriptors dominated by almost one single value.

iqr_thresholdfloat

Minimum absolute variability required for continuous descriptors.

rel_thresholdfloat

Minimum relative variability required for continuous descriptors.

binary_thresholdfloat

Minimum minority-class proportion required for binary descriptors.

correlation_thresholdfloat

If two descriptors are too correlated, one of them is removed.

min_descriptorsint

Minimum number of descriptors required after cleanup.

Coverage selection:

n_pointsint or None

Number of representative molecules to select. If not provided, ALMOS estimates an automatic coverage budget.

evaluatebool

If True, skip representative reselection and only evaluate the existing user-provided selection stored as batch = 0 in the input CSV.

modestr

Selection strategy. "representative" selects one real molecule nearest to each prototype centroid; diversity-focused selection keeps extra distant candidates.

cluster_auto_budget_candidateslist

Candidate budgets used as anchors for the automatic coverage scan.

cluster_auto_budget_marginal_gain_thresholdfloat

Marginal coverage-improvement threshold used to stop the budget scan.

cluster_auto_budget_min_umap_areafloat

Minimum UMAP area fraction required for automatic recommendation.

cluster_auto_budget_lookaheadint

Number of later tested budgets inspected for local-slowdown detection.

cluster_auto_budget_max_pointsint

Maximum automatic budget.

cluster_natural_reportbool

Run optional KMeans/GMM/HDBSCAN natural clustering diagnostics.

PCA safeguard:

enable_pcabool

Disable PCA with --no_pca.

cluster_high_dimensionality_thresholdint

Descriptor-count threshold above which PCA can be activated.

cluster_pca_explained_variance_thresholdfloat

Target explained variance retained by PCA.

cluster_pca_min_acceptable_variancefloat

Minimum variance required to accept PCA instead of raw descriptor space.

cluster_pca_min_components : int pca_max_components : int pca_max_components_fraction : float

Large dataset mode:

large_dataset_modebool

Disable with --no_large_dataset_mode.

cluster_standard_dataset_thresholdint

Upper limit of the standard regime.

cluster_very_large_dataset_thresholdint

Upper limit of the large regime.

cluster_ultra_large_dataset_thresholdint

Above this, the workflow enters ultra-large mode.

cluster_large_silhouette_sample_size : int cluster_very_large_silhouette_sample_size : int cluster_ultra_large_silhouette_sample_size : int cluster_large_dataset_stability_repeats : int cluster_fast_screening_top_candidates : int

Algorithm-specific search space:

cluster_kmeans_coarse_grid_size : int cluster_kmeans_top_refinement_candidates : int cluster_kmeans_refine_radius : int cluster_kmeans_bo_fraction : float cluster_kmeans_bo_max_evaluations : int cluster_gmm_dimensionality_threshold : int cluster_gmm_standard_coarse_grid_size : int cluster_gmm_standard_refine_radius : int cluster_gmm_large_coarse_grid_size : int cluster_gmm_large_refine_radius : int cluster_gmm_very_large_coarse_grid_size : int cluster_gmm_very_large_refine_radius : int cluster_gmm_bo_fraction : float cluster_gmm_bo_max_evaluations : int cluster_gmm_bic_shortlist_size : int cluster_hdbscan_standard_min_cluster_ratios : list cluster_hdbscan_large_min_cluster_ratios : list cluster_hdbscan_very_large_min_cluster_ratios : list cluster_hdbscan_standard_min_samples : list cluster_hdbscan_large_min_samples : list cluster_hdbscan_very_large_min_samples : list

Quality filters:

cluster_filter_max_noise_fractionfloat

Reject candidates with too many noise points, mainly for HDBSCAN.

cluster_filter_max_cluster_fractionfloat

Reject candidates dominated by one oversized cluster.

cluster_filter_max_imbalance_penaltyfloat

Reject candidates with extreme cluster-size imbalance.

cluster_quality_warning_silhouette_threshold : float cluster_quality_warning_stability_threshold : float cluster_quality_warning_noise_threshold : float cluster_quality_warning_imbalance_threshold : float cluster_quality_warning_final_score_threshold : float cluster_quality_good_silhouette_threshold : float cluster_quality_good_stability_threshold : float cluster_quality_good_noise_threshold : float cluster_quality_good_imbalance_threshold : float cluster_quality_good_final_score_threshold : float

AL

Parameters

albool

Indicates whether the active learning process is enabled and should be performed. Defaults to "False". This parameter is activated in command line (i.e. --al)

csv_namestr

Name of the CSV file containing the database. (i.e. 'FILE.csv').

ystr

Name of the column containing the response variable in the input CSV file (i.e. 'solubility').

namestr

Name of the column containing the molecule names in the input CSV file (i.e. 'names').

ignorelist, default=[]

List containing the columns of the input CSV file that will be ignored during the ROBERT process (i.e. --ignore "[name,SMILES]"). The descriptors will be included in the final CSV file. The y value, name column and batch column are automatically ignored by ROBERT.

n_expsint,

Number of experiments to be selected in the active learning process for the new batch. (i.e. '--n_exps 10') If not provided or invalid, the program will request the values in the proper format.

tolerancestr, default='medium'

Indicates the tolerance level for the convergence process, defining the percentage change threshold required for convergence. Options: 1. 'tight': Strictest level, convergence occurs if the metric improves by ≤1% (threshold = 0.01). 2. 'medium': Balanced level, convergence occurs if the metric improves by ≤5% (threshold = 0.05). 3. 'wide': Least strict, convergence occurs if the metric improves by ≤10% (threshold = 0.10). (i.e. '--tolerance tight')

robert_keywordsstr, default=""

Additional keywords to be passed to the ROBERT model generation (i.e. --robert_keywords "--model RF --train [70] --seed [0]")

objectivestr

Optimization direction for hit selection. Always required and must be 'max' or 'min'. (i.e. '--objective max')

modestr, optional

Optional manual override for the acquisition strategy. Use 'model' to rank by uncertainty or 'hit' to rank by prediction with uncertainty. If omitted, ALMOS selects the strategy automatically from the model score.

alphafloat, optional

Optional acquisition weight used in hit mode. It also overrides the automatic alpha when the strategy is auto and the selected score activates hit ranking. (i.e. '--alpha 0.5' or '--alfa 0.5')