[QUESTION]: generating MoleculeDataset with SMILES and additional features (NOT CSV INPUT)

What are you trying to do? I have a GCNN model built using chemprop that takes SMILES and an additional (binary) feature as inputs. When predicting test compounds, I am not loading data from a CSV file (in this case I could simply provide path to the file containing additional features for the test compounds as an argument to chemprop_predict method). But I have the SMILES available in a list object in Python. So, I went ahead and tried to generate a MoleculeDataset object using the utility method get_data_from_smiles available in chemprop/data/utils.py. But I see that the method takes a list of feature generators as argument but not the additional features. Therefore, I modified the method as follows:

Previous attempts

def get_data_from_smiles_with_additional_features(smiles: List[str],
                         skip_invalid_smiles: bool = True,
                         logger: Logger = None,
                         features: List = None,
                         features_generator: List[str] = None) -> MoleculeDataset:
    """
    Converts SMILES to a MoleculeDataset.

    :param smiles: A list of SMILES strings.
    :param skip_invalid_smiles: Whether to skip and filter out invalid smiles.
    :param logger: Logger.
    :param features: List of additional features
    :param features_generator: List of features generators.
    :return: A MoleculeDataset with all of the provided SMILES.
    """
    debug = logger.debug if logger is not None else print

    data = MoleculeDataset([
        MoleculeDatapoint(
            smiles=smile,
            features=feature,
            row=OrderedDict({'smiles': smile}),
            features_generator=features_generator
        ) for smile, feature in zip(smiles, features)
    ])

    # Filter out invalid SMILES
    if skip_invalid_smiles:
        original_data_len = len(data)
        data = filter_invalid_smiles(data)

        if len(data) < original_data_len:
            debug(f'Warning: {original_data_len - len(data)} SMILES are invalid.')

    return data

Question The predictions I obtain using this method match the predictions from command line when using an additional file with the same features. However, I have only used a single feature (i.e., one column in a file). Is there a better way to deal with this when I have multiple additional features?

Thank you, Vishal

[shihchengli]

import torch
from chemprop.args import TrainArgs
path = '/path/to/checkpoint'
state = torch.load(path, map_location=lambda storage, loc: storage)
args = TrainArgs()
args.from_dict(vars(state["args"]), skip_unsettable=True)
args.features_size  # expected shape of the additional features

@iwwwish You can use the above code to get the size of molecular features. I will close this issue now. If you have other questions about Chemprop, feel free to open a new issue.

[iwwwish]

That worked! Thank you.

Is there a way to obtain information from the model file itself whether it takes additional features as input and if yes, the expected shape of the additional features?

Best, Vishal