DIDGen Parameters

Here is a list of parameters for DIDgen. They can be input as a single yaml file and as a python dictionary in the python API. You only need to specify the parameters that you want to change from default. The full list of parameters and their default value can be found in didgen/config.py.

Training and datasets

Parameter	Type	Default	Description
property	list of int	[4]	Index of the property to optimize in the dataset features. Multiple properties can be selected at once. e.g., 4 is the index of the HOMO-LUMO gap property in PyG QM9
type_list	list of str	['H', 'C', 'N', 'O', 'F']	Atomic ordering in the dataset one-hot encoding
max_size	int	25	Maximum molecular size to train on and to generate (including hydrogen atoms)
n_data	int	1000	Size of the dataset to use, if larger than the size of the dataset, all the data is used
datasets	list of str	['qm9']	Datasets to train on. 'qm9' will download PyG QM9 otherwise DIDgen will look in OUTDIR/dataset_name for a QM9like dataset (more on this coming soon, including scripts to use your own datasets)
num_epochs	int	400	Number of epochs
batch_size	int	30	Batch size
learning_rate	float	0.001	Learning rate (Adam optimizer)
noise_factor	float	0.05	Random noises added to input vectors to make the model less sensitive to non-integers (often not necessary)
use_pretrained	bool	False	Whether to reuse weights from previous run, useful when generating molecules without having to retrain. Weights are saved in OUTDIR/model_weights.pth.

Model

Parameter	Type	Default	Description
model	str	SimpleNet	Model to train and generate with. Current options are: "SimpleNet", "CrippenNet" and "CGCNN"
layer_list	list of int	[128]	List and size of MLP layers after the convolutions. e.g. [128,128,50] would be 3 layer MLP. Only available for "SimpleNet" and "CrippenNet"
atom_fea_len	int	20	Size of the atomic embedding before and during the convolutions
n_conv	int	3	Number of convolutions
weight_decay	int	0	Weight decay as implemented in Pytorch Adam optimizer
dropout	int	0	Dropout rate before each linear layer. Only available for "SimpleNet" and "CrippenNet"
show_train	bool	False	Whether to display and plot training statistics.
atom_class	bool	False	Whether to train on atomic classification. Only available for CrippenNet. Note that CrippenNet can be trained as a repressor.

Generation

Parameter	Type	Default	Description
target	list of float	[4.1]	Target value of the property. e.g. An energy gap of 4.1 eV
proportions	list of float	[0.411, 0.3026, 0.0602, 0.0735, 0.0015]	Target proportion between different elements following type_list. This is set as loose criteria (see stop_prop)
stop_loss	float	0.3	Stopping criteria for the property target. DIDgen stops when the value is within stop_loss of target
stop_prop	int	100	Stopping criteria for element proportions. Set such that criteria is always met. Could be useful to generate isomers.
bonding	list of int	[1, 4, 3, 2, 1]	Valence of each element in type_list
min_size	int	5	Minimum size of generated molecules
start_from	str	!	Molecular representation to start from. Options are: "random" to start from random noise, "saved" to reuse the previous starting point, "!" to start from a random entry in the dataset and id to start from a molecule with index id in the dataset.
starting_size	int	25	Maximum starting size when starting from random noise
n_iter	int	300	Maximum number of iterations to reach the target
inv_r	float	0.01	Rate of gradient descent (Adam optimizer)
l_loss	float	0.01	Relative importance of the target property loss (see equation 2 in the manuscript)
l_prop	int	100	Relative importance of the elemental proportion loss
show_losses	bool	False	Whether to display and plot optimization statistics (slow)
max_attempts	int	3	Maximum number of attempts (starts) to reach the target
bond_multiplier	float	0.3	Magnitude of the random noise when using start_from; "random". Larger magnitude creates more bonds.
embed	bool	True	Whether to find a conformer (3D atomic positions) for the generated molecule