Chemprop

Property prediction plays a key role in drug discovery and materials science, helping researchers estimate characteristics like toxicity, bioactivity, and solubility.

We have now integrated Chemprop, a PyTorch-based framework, into our platform, enabling easy access with just a few clicks.

Model parameters

Parameter	Description	Default
dataset_type	Type of dataset (e.g., classification or regression). Determines the training loss function.	Regression
metric	Evaluation metric. Does not impact training loss. Defaults to "AUC" for classification, "RMSE" for regression.	None
multiclass_num_classes	Number of classes when running multiclass classification.	3
num_folds	Number of folds when performing cross validation.	1
data_seed	Seed for data splitting. For multiple folds, increments by 1 for each fold.	0
split_sizes	Proportions for train/validation/test splits.	0.8, 0.1, 0.1
split_type	Data splitting method (e.g., random, cross-validation).	Random
activation	Activation function used in the model (e.g., ReLU, tanh).	ReLU
atom_messages	Uses messages on atoms rather than bonds.	False
message_bias	Adds bias to linear layers.	False
ensemble_size	Number of models in the ensemble.	1
message_hidden_dim	Hidden layer dimensionality in the message-passing network.	300
depth	Number of message-passing steps.	3
dropout	Dropout probability for training.	0.0
undirected	Sums bond vectors to treat edges as undirected.	False
ffn_hidden_dim	Hidden dimension size for the feed-forward network.	300
ffn_num_layers	Number of layers in the feed-forward network after message-passing encoding.	2
epochs	Total number of training epochs.	50
batch_size	Batch size for training.	64
warmup_epochs	Linear increase of learning rate from initial to max, then exponential decay to final.	2.0
init_lr	Starting learning rate for training.	0.0001
max_lr	Peak learning rate for training.	0.001
final_lr	Ending learning rate after decay.	0.0001
no_descriptor_scaling	Disables feature scaling for descriptors.	False

Train

To train a model with Chemprop:

1. Prepare dataset
   • Ensure it includes a column with SMILES or molblocks.
2. Set up the model
   • Go to ML > Models > Train Model….
   • Select dataset, specify the target column (values to predict), and the feature column (SMILES or molblocks).
3. Choose Chemprop as the model engine
   • From available engines, select Chemprop.
4. Adjust model settings
   • Configure training parameters (epochs, batch size, learning rate) and evaluation metrics as needed.
5. Train model
   • Click Train to start the training process and generate the model.

Predict

Once training is complete, follow these steps to apply a trained model for predictions:

1. Select the dataset
   • Choose the table containing the features for prediction.
2. Apply the model
   • Go to ML > Models > Apply Model.
   • Select the trained model you wish to use from the available options.
3. Generate predictions
   • Click Apply Model to make predictions on the selected dataset.