Configuration Reference

All scripts are driven by TOML configuration files located in bench_configs/.

Benchmarking Config

File: bench_configs/run_benchmarking.toml

Used by: scripts/run_benchmarking.py --config <path>

Required Keys

Key	Type	Valid Values	Description
data_path	str	Existing file or directory	Path to a single HDF5 file or a directory of HDF5 files
output_dir	str	Existing directory	Directory where results are saved
methods	list	["UMAP", "t-SNE", "GTM", "PCA"]	DR methods to benchmark (any subset)
n_components	int	Positive integer	Number of PCA components (typically 2)
k_neighbors	list[int]	List of positive integers	Values of k for neighborhood metrics, e.g. [2, 5, 10, 20, 50]
optimization_type	str	"insample", "outsample"	Whether to evaluate on training data or held-out validation
scaling	str	"standardize", "minmax", "none"	Feature scaling method
similarity_metric	str	"euclidean", "tanimoto"	Distance metric for ambient space
sample_size	int	Positive integer	Number of samples for metrics calculation on large datasets
test	bool	true, false	If true, uses a single parameter combination per method (for testing)
plot_data	bool	true, false	Whether to generate visualization plots

Optional Keys

Key	Type	Description
val_data_path	str	Path to validation HDF5 file (required when optimization_type = "outsample")
k_hit	int	Specific k value used as optimization target (must appear in k_neighbors)
log_level	str	Logging verbosity: "DEBUG", "INFO", "WARNING", "ERROR"

Example

data_path = "datasets/CHEMBL204.h5"
output_dir = "results/insample_eucl"
methods = ["UMAP", "t-SNE", "GTM", "PCA"]
n_components = 2
k_neighbors = [2, 5, 10, 20, 50]
k_hit = 20
optimization_type = "insample"
scaling = "standardize"
similarity_metric = "euclidean"
sample_size = 2500
test = false
plot_data = true
log_level = "INFO"

---

Feature Generation Config

File: bench_configs/features.toml

Used by: scripts/generate_descriptors.py <path>

Sections

[input]

Key	Type	Description
input_path	str	Directory containing input SMILES files
file_pattern	str	Glob pattern for input files (e.g. "*.smi")
output_path	str	Directory where output HDF5 files are saved

[chemdist]

Key	Type	Description
model_path	str	Path to trained PyTorch model (.pt file)
device	str	"cuda" or "cpu"

[chemdist.params]

Key	Type	Description
edge_in_feats	int	Number of edge features for GNN
embed_size	int	Embedding vector dimensionality
node_in_feats	int	Number of node features for GNN

[morgan]

Key	Type	Description
morgan_radius	int	Radius for Morgan fingerprints (default: 2)
morgan_fp_size	int	Fingerprint bit length (default: 1024)

[maccs]

No configuration needed. MACCS keys are 167-bit structural keys generated by RDKit.

[preprocess]

Controls whether to remove constant features after generation.

[logging]

Key	Type	Description
log_path	str	Directory for log files
log_level	str	Logging level

[parallel]

Key	Type	Description
num_workers	int	Number of parallel workers (1 = no parallelization)

Example

[input]
input_path = "datasets/"
file_pattern = "*.smi"
output_path = "features/"

[chemdist]
model_path = "models/model_trained.pt"
device = "cuda"

[chemdist.params]
edge_in_feats = 12
embed_size = 16
node_in_feats = 74

[morgan]
morgan_radius = 2
morgan_fp_size = 1024

[maccs]

[preprocess]

[logging]
log_path = "logs/"
log_level = "INFO"

[parallel]
num_workers = 4

---

Method Parameter Grids

Located in bench_configs/method_configs/. These define the hyperparameter search space for each DR method.

UMAP (umap_config.toml)

Parameter	Default Grid	Description
n_neighbors	[2, 4, 6, 8, 16, 32, 64, 128, 256]	Size of local neighborhood
min_dist	[0.0, 0.1, 0.2, 0.3, 0.4, 0.6, 0.8, 0.99]	Minimum distance between points in embedding
n_components	[2]	Output dimensionality

Total combinations: 72

t-SNE (tsne_config.toml)

Parameter	Default Grid	Description
perplexity	[1, 2, 4, 8, 16, 32, 64, 128]	Effective number of local neighbors
exaggeration	[1, 2, 3, 4, 5, 6, 8, 16, 32]	Early exaggeration factor
n_components	[2]	Output dimensionality

Total combinations: 72

GTM (gtm_config.toml)

Parameter	Default Grid	Description
num_nodes	[225, 625, 1600]	Number of grid nodes (square root gives grid dimension)
num_basis_functions	[100, 400, 1225]	Number of RBF basis functions
reg_coeff	[1, 10, 100]	Regularization coefficient
basis_width	[0.1, 0.4, 0.8, 1.2]	Width of RBF basis functions

Total combinations: 108