v1.3.1

Change Log

From v1.3.0 to v1.3.1

Fixes

• Fix not re-initiating model weights during DNN training
• Feature values converted to np.float32 and then np.inf are converted to nan on DescriptorsCalculator.__call__.

Changes

• QSPRDataset.prepareDataset changed attributes from standardize and sanitize to only standardizer.
  • Accepted parameters are either chembl, old, or a function that reads and standardizes smiles.
  • None is now also supported to allow skipping smiles standardization.
  • SMILES standardization now runs in parallel, but if the input function is not pickable, will just run on a single core.
• QSPRModel.predictMols now accepts parameters smiles_standardizer, n_jobs and fill_value.

v1.3.0

Change Log

From v1.2.0 to v1.3.0

Fixes

• problems with PaDEL descriptors and fingerprints on Linux were fixed

Changes

• QSPRModel metadata now contains two extra entries:
  1. model_class - the fully qualified class name of the model
  2. version - the version of QSPRPred used to save the model
  • this change is not compatible with older files, but you can manually add these two entries and it should work fine in the newer version

New Features

• The QSPRModel.fromFile() method can now instantiate a model from a file directly without knowing the underlying model type. It simply uses the class path stored in the model metadata file now.

v1.2.0

Change Log

From v1.1.0 to v1.2.0

Fixes

• Fix issue with Mordred descriptor
• Descriptor sets now process a list of molecules instead of just one at a time (prevents performance issues if multiple sets are calculated in parallel)
• Empty values of descriptors are now not imputed with 0 automatically, but are left as NaN or None instead

Changes

• Some features not specific to machine learning were extracted from QSPRDataset to a new class called MoleculeTable
  • MoleculeTable is mainly to hold data about molecules, including their descriptors, scaffolds, bioactivities and other data
    • this class also now manages settings for parallelization and chunking in the constructor rather than on per method basis
    • this class will be used as the base class for other data set classes that need molecule data, but have to perform their own transformations to do their job
  • QSPRDataset derives from MoleculeTable an object describing the training and test set for modelling and also handles data preparation
• QSPRDataset now handles saving of its metadata and other related files (i.e. standardizers and other data transformers) with one method (save) -> names of the files start with a chosen prefix, which is a name given to the data set
• The SKLearnStandardizer was added for scaler fitting, applying, saving and loading
  • The standardization of features is now possible with the feature_standardizer argument of QSPRDataset.prepareDataset by supplying an instance of SKLearnStandardizer or directly a StandardScaler or any other standardizer from sklearn.preprocessing with BaseEstimator interface
    • standardization is now also done separately for training and test sets in cross-validation as well
• The DescriptorSet interface was updated and all built-in descriptors were adapted to reflect this change.
  • The presence of descriptors property getter and setter is now enforced.
  • When called the DescriptorSet implementations now strictly return lists.
  • Conversion to descriptor data frame is now handled exclusively in DescriptorsCalculator
• The datasplit interface was changed to mimic the sklearn.model_selection.BaseCrossValidator interface so all sklearn cross-validation methods can be used with QSPRPred out of the box to either generate train/test split or cross-validation splits (see the new features below)
• Default chunk_size for MoleculeTable was set to 50 so that smaller data sets can take advantage of more CPUs as well.
• The number of CPUs to use for parallel operations by MoleculeTable is now set in the __init__ of the class and is 1 by default so that the default behaviour is to not use parallelism.
• DescriptorSets are now initialized with the specific arguments instead of args and kwargs.
• MorganFP was replaced by a more general class FingerprintSet which uses an object from the Fingerprint class as its fingerprint type
• The Predictor class was replaced, its features are now accessible with the models directly:

  • from qsprpred import QSPRsklearn # QSPRDNN can be used the same way
    from qsprpred import QSPRDataset
    
    # creation and loading 
    model = QSPRsklearn( # or QSPRDNN
        name="any_name",
        base_dir="/some/path"
    )
    
    # loading directly from meta file also possible 
    model = QSPRsklearn.fromFile("/path/to/any_name_meta.json")
    
    # predictions can be done directly on a list of SMILES
    model.predictMols([
        'CC(=C)C1CC2=C(O1)C=CC3=C2OC4COC5=CC(=C(C=C5C4C3=O)OC)OC',
        'CCOC(=O)C1=C2CN(C(=O)C3=C(N2C=N1)C=CC(=C3)F)C'
    ])
    
    # classifiers can also use predict_probas=True to get probablities
    model.predictMols([
        'CC(=C)C1CC2=C(O1)C=CC3=C2OC4COC5=CC(=C(C=C5C4C3=O)OC)OC',
        'CCOC(=O)C1=C2CN(C(=O)C3=C(N2C=N1)C=CC(=C3)F)C'
    ], use_probas=True)
    
    # it is also possible to give a QSPRDataset directly:
    dataset = QSPRDataset(name="data")
    model.predict(dataset)

  • Calls to predict, predictProba or predictMols with use_probas=True will now return a score of None for invalid molecules.

New Features

• Tutorials for training and using the QSPR models
• Depiction of results for classification models (see qsprpred.plotting.classification)
• The precomputed flag was added to QSPRDataset
• Added an option to directly fetch QSPRDataset from Papyrus with accession IDs (see qsprpred.data.sources.papyrus)
• The datasplit interface is now used to both generate train/test split and also the cross-validation splits
• Train/test split of the data set is now saved in the matrix itself and is reloaded upon deserialization
• MoleculeTable was updated with new features to generate scaffolds of molecules
• TanimotoDistances was added as descriptortype.
• Balanced class weighing was added as an option to the CLI
• PredictorDesc was added as a new DescriptorSet type. It uses a QSPRpred model as descriptor.
• New submodule for evaluation metric custom (qsprpred.metrics) with calibration_error function to estimate the calibration of a classifier
• Added the Mold2 and PaDEL molecular descriptors