MolecularDiffusion.runmodes.generate.tasks_generate

Attributes

logger

Classes

GenerativeFactory

PharmacophoreConditionGenerator

Entry point for all ShEPhERD pharmacophore generation modes.

Module Contents

class MolecularDiffusion.runmodes.generate.tasks_generate.GenerativeFactory(task: MolecularDiffusion.modules.tasks.task.Task, task_type: str = 'unconditional', sampling_mode: str = 'ddpm', num_generate: int = 100, mol_size: List[int] = [0, 0], target_values: List[float] = [], property_names: List[str] = [], negative_target_values: List[float] = [], batch_size: int = 1, seed: int = 86, n_frames: int = 0, output_path: str = 'generated_mol', condition_configs={}, max_mol_size: int = 0)
conditional_generation()
hybrid_guidance()
preprocess_ref_structure(device)

Load and preprocess a reference molecular structure from an XYZ file.

This function reads an XYZ file, encodes atomic features, normalizes coordinates and features, and returns a tensor combining positions and processed features. When the model uses extra node features (ndim_extra > 0), they are either computed from the reference structure or zero-padded.

Returns:

Tensor of shape (1, n_atoms, 3 + n_features + ndim_extra + 1)

containing [normalized_coords | normalized_onehot | normalized_extra | normalized_charges].

Return type:

torch.Tensor

Raises:
property_guidance()
property_prediction(xh: torch.Tensor, t: int)
run()
structural_guidance()
unconditional_generation()
batch_size = 1
condition_configs
max_mol_size = 0
mol_size = [0, 0]
n_frames = 0
negative_target_values = []
num_generate = 100
output_path = 'generated_mol'
property_names = []
sampling_mode = 'ddpm'
seed = 86
target_values = []
task
task_type = 'unconditional'
class MolecularDiffusion.runmodes.generate.tasks_generate.PharmacophoreConditionGenerator(task, task_type: str = 'pharmacophore_condition', num_generate: int = 10, batch_size: int = 1, N_x1: list = [20], N_x4: int = 5, N_x1_sampling: str = 'uniform', distributions_path: str = None, distributions_key: str = 'gdb', num_steps: int = 400, prior_noise_scale: float = 1.0, denoising_noise_scale: float = 1.0, output_path: str = 'generated_pharmacophore', seed: int = 42, verbose: bool = True, reference_mol: str = None, mol_idx: int = 0, scaffold_smarts: str = None, inpaint_x1_pos: bool = True, inpaint_x1_x: bool = True, inpaint_x1_bonds: bool = True, stop_inpainting_at_time_x1_pos: float = 0.0, stop_inpainting_at_time_x1_x: float = 0.0, stop_inpainting_at_time_x1_bonds: float = 0.0, stop_inpainting_at_time_x4: float = 0.0, save_modalities: bool = False, compute_x1: bool = True, compute_x2: bool = True, compute_x3: bool = True, compute_x4: bool = True, start_time: float = 0.5, condition_configs: dict = {})

Entry point for all ShEPhERD pharmacophore generation modes.

task_type selects the generation mode:

  • unconditional : joint x1/x2/x3/x4 from pure noise

  • pharmacophore_condition: condition on pharmacophores extracted from reference_mol

  • shape_conditioned : condition on surface + electrostatics from reference_mol

  • pharmacophore_inpaint : scaffold inpainting — keep scaffold_smarts atoms, regenerate rest

  • from_intermediate_time : soft scaffold hopping starting from a noisy intermediate

For all conditional modes, provide:

reference_mol : path to .pkl (list of (molblock, charges) tuples) or .sdf file mol_idx : which molecule to use from the file (default 0)

For pharmacophore_inpaint:
scaffold_smarts: SMARTS pattern selecting atoms to keep (e.g. ‘c1ccccc1’)

If null, keeps all ring atoms and regenerates chains.

run()
N_x1 = [20]
N_x1_sampling = 'uniform'
N_x4 = 5
batch_size = 1
compute_x1 = True
compute_x2 = True
compute_x3 = True
compute_x4 = True
condition_configs
denoising_noise_scale = 1.0
inpaint_x1_bonds = True
inpaint_x1_pos = True
inpaint_x1_x = True
mol_idx = 0
num_generate = 10
num_steps = 400
output_path = 'generated_pharmacophore'
prior_noise_scale = 1.0
reference_mol = None
save_modalities = False
scaffold_smarts = None
seed = 42
start_time = 0.5
stop_inpainting_at_time_x1_bonds = 0.0
stop_inpainting_at_time_x1_pos = 0.0
stop_inpainting_at_time_x1_x = 0.0
stop_inpainting_at_time_x4 = 0.0
task
task_type = 'pharmacophore_condition'
verbose = True
MolecularDiffusion.runmodes.generate.tasks_generate.logger