Version 3.2.1

Change Log

From v3.2.0 to v3.2.1

Fixes

• Add variable version to papyrus_filter for consistent version use.

Changes

None.

New Features

None.

Removed Features

None.

Version 3.2.0

Change Log

From v3.1.1 to v3.2.0

Fixes

• Fixed a bug in ChempropModel that caused it not to work with missing values in the
  target column.

Changes

• calibration_score is now implemented under the Metric class as CalibrationScore.

New Features

• Added a range of new
  metrics: BEDROC, EnrichmentFactor, RobustInitialEnhancement,
  Prevalence, Sensitivity, Specificity, PositivePredictivity, NegativePredictivity,
  CohenKappa, BalancedPositivePredictivity, BalancedNegativePredictivity,
  BalancedMatthewsCorrcoeff, BalancedCohenKappa, KSlope, R20, KPrimeSlope,
  RPrime20, Pearson, Spearman, Kendall, AverageFoldError,
  AbsoluteAverageFoldError, PercentageWithinFoldError
• Added MaskedMetric which can be wrapped around any metric to mask datapoints
  when a target value is missing.
• Added a tutorial on model and data serialization.
• ApplicabilityDomain now has a transform method that can be used to transform
  a dataset to a continuous applicability domain score, such as the distance to the
  nearest neighbor in the training set (an example was added to the
  tutorials).

Removed Features

None.

Version 3.1.1

Change Log

From v3.0.2 to v3.1.1

Fixes

• Fixed a bug in QSPRDataset where property transformations were not applied.
• Fixed a bug where an attached standardizer would be refit when calling
  QSPRModel.predictMols with use_applicability_domain=True.
• Fixed random seed not set in FoldsFromDataSplit.iterFolds for ClusterSplit.

Changes

• renamed PandasDataTable.transform to PandasDataTable.transformProperties
• moved imputeProperties, dropEmptyProperties and hasProperty from MoleculeTable
  to PandasDataTable.
• removed getProperties, addProperty, removeProperty, now use PandasDataTable
  methods directly.
• Since the way descriptors are saved has changed, this release is incompatible with
  previous data sets and models. However, these can be easily converted to the new
  format by adding
  a prefix with descriptor set name to the old descriptor tables. Feel free to contact
  us if you require assistance with this.
• Due to some changes in rdkit-2023.9.6, the add_rdkit
  option for molecule tables temporarily might not work.
  This also affects the current ChemProp integration, which was not adapted to 2.0.0 yet.
  In order to prevent these issues, QSPRpred now forces rdkit version rdkit-2023.9.5,
  but we will be working on resolving these.

New Features

• Descriptors are now saved with prefixes to indicate the descriptor sets. This reduces
  the chance of name collisions when using multiple descriptor sets.
• Added new methods to MoleculeTable and QSARDataset for more fine-grained control
  of clearing, dropping and restoring of descriptor sets calculated for the dataset.
  • dropDescriptorSets will drop descriptors associated with the given descriptor
    sets.
  • dropDescriptors will drop individual descriptors associated with the given
    descriptor sets and properties.
  • All drop actions are restorable with restoreDescriptorSets unless explicitly
    cleared from the data set with the clear parameter of dropDescriptorSets.
• Added a proper API for parallelization backend selection and configuration (see
  documentation of ParallelGenerator and JITParallelGenerator for more information).
• Clusters can now be added to a MoleculeTable with addClusters and retrieved with
  getClusters, similar to scaffolds.

Removed Features

• removed support for PyBoost since the project was abandoned by the original developers and is no longer maintained

Version 3.0.2

Change Log

From v3.0.1 to v3.0.2

Fixes

• Fixed a bug where an attached standardizer would be refit when calling
  QSPRModel.predictMols with use_applicability_domain=True.
• Fixed a bug with use_applicability_domain=True in QSPRModel.predictMols
  where an error would be raised if there were invalid molecules in the input.
• Fixed a bug where dataset type was not properly set to numeric
  in MlChemADWrapper.contains
• Fixed a bug in QSPRDataset where property transformations were not applied.
• Fixed a bug where an attached standardizer would be refit when calling
  QSPRModel.predictMols with use_applicability_domain=True.
• Fixed random seed not set in FoldsFromDataSplit.iterFolds for ClusterSplit.
• Fixed a bug where class ratios were shuffled in the RatioDistributionAlgorithm.

Changes

• The module containing the sole model base class (QSPRModel) was renamed
  from models to model.
• Restrictions on numpy versions were removed to allow for more flexibility in
  package installations. However, the BorutaFilter feature selection method does not
  function with numpy versions 1.24.0 and above. Therefore, this functionality now
  requires a downgrade to numpy version 1.23.0 or lower. This was reflected in the
  documentation and numpy itself outputs a reasonable error message if the version is
  incompatible.
• Data type in MlChemADWrapper is now set to float64 by default, instead
  of float32.
• Saving of models after hyperparameter optimization was improved to ensure parameters
  are always propagated to the underlying estimator as well.

New Features

• The DataFrameDescriptorSet class was extended to allow more flexibility when joining
  custom descriptor sets.
• Added the prepMols method to DescriptorSet to allow separated customization of
  molecule preparation before descriptor calculation.
• The package can now be installed from the PyPI repository 🐍📦.
• New argument (refit_optimal) was added to HyperparameterOptimization.optimize()
  method to make refitting of the model with optimal parameters easier.

Removed Features

None.

v3.0.1

Change Log

From v3.0.0 to v3.0.1

Fixes

• Fixed a bug in QSPRDataset where property transformations were not applied.

Changes

• renamed PandasDataTable.transform to PandasDataTable.transformProperties
• moved imputeProperties, dropEmptyProperties and hasProperty from MoleculeTable
  to PandasDataTable.
• removed getProperties, addProperty, removeProperty, now use PandasDataTable
  methods directly.

New Features

Removed Features

v3.0.0

Change Log

From v2.1.1 to v3.0.0

Fixes

• Fixed random seeds to give reproducible results. Each dataset is initialized with a
  single random state (either from the constructor or a random number generator) which
  is used in all subsequent random operations. Each model is initialized with a single
  random state as well: it uses the random state from the dataset, unless it's overriden
  in the constructor. When a dataset is saved to a file so is its random state, which is
  used by the dataset when the dataset is reloaded.
• fixed error with serialization of the DNNModel.params attribute, when no parameters
  are set.
• Fix bug with saving predictions from classification model
  when ModelAssessor.useProba set to False.
• Add missing implementation of QSPRDataset.removeProperty
• Improved behavior of the Papyrus data source (does not attempt to connect to the
  internet if the data set already exists).
• It is now possible to define new descriptor sets outside the package without errors.
• Basic consistency of models is also checked in the unit test suite, including in
  the qsprpred.extra package.
• Fixed a problem with feature standardizer being retrained on prediction data when a
  prediction from SMILES was invoked. This affected all versions of the package higher
  or equal to v2.1.0.
• Fixes to the fromMolTable method in various data set implementations, in particular
  in copying of the feature standardizer and other settings.
• Fixed not working cluster split and --imputation from data_CLI.py.
• Fixed a problem with ProteinDescriptorSet.getDescriptors returning descriptors in
  wrong order with Pandas <v2.2.0.

Changes

• The model is now independent of data sets. This means that the model no longer
  contains a reference to the data set it was trained on.
  • The fitAttached method was replaced with fitDataset, which takes the data set
    as
    an argument.
  • Assessors now also accept a data set as a second argument. Therefore, the same
    assessor
    can be used to assess different data sets with the same model settings.
  • The monitoring API was also slightly modified to reflect this change.
  • If a model requires initialization of some settings from data, this can be done in
    its initFromDataset method, which takes the data set as an argument. This method
    is called automatically before fitting, model assessment, and hyperparameter
    optimization.
• The whole package was refactored to simplify certain commonly used imports. The
  tutorial code was adjusted to reflect that.
• The jupyter notebooks in the tutorial now pass a random state to ensure consistent
  results.
• The default parameter values for STFullyConnected have changed from n_epochs =
  1000 to n_epochs = 100, from neurons_h1 = 4000 to neurons_h1 = 256
  and neurons_hx = 1000 to neurons_hx = 128.
• Rename HyperParameterOptimization to HyperparameterOptimization.
• TargetProperty.fromList and TargetProperty.fromDict now accept a both a string and
  a TargetTask as the task argument,
  without having to set the task_from_str argument, which is now deprecated.
• Make EarlyStopping.mode flexible for QSPRModel.fitDataset.
• save_params argument added to OptunaOptimization to save the best hyperparameters
  to the model (default: True).
• We now use jsonpickle for object serialization, which is more flexible than the
  non-standard approach before, but it also means previous models will not be compatible
  with this version.
• SklearnMetric was renamed to SklearnMetrics, it now also accepts an scikit-learn
  scorer name as input.
• QSPRModel.fitDataset now accepts a save_model (default: True)
  and save_dataset (default: False) argument to save the model and dataset to a file
  after fitting.
• Tutorials were completely rewritten and expanded. They can now be found in
  the tutorials folder instead of the tutorial folder.
• MetricsPlot now supports multi-class and multi-task classification models.
• CorrelationPlot now supports multi-task regression models.
• The behaviour of QSPRDataset was changed with regards to target properties. It now
  remembers the original state of any target property and all changes are performed in
  place on the original property column (i.e. conversion to multi-class classification).
  This is to always maintain the same property name and always have the option to reset
  it to the raw original state (i.e. if we switch to regression or want to repeat a
  transformation).
• The default log level for the package was changed from INFO to WARNING. A new
  tutorial
  was added to explain how to change the log level.
• RepeatsFilter argument year_name renamed to time_col and
  arugment additional_cols added.
• The perc argument of BorutaPy can now be set from the CLI.
• Descriptor calculators (previously used to aggregate and manage descriptor sets) were
  completely removed from the API and descriptor sets can now be added directly to the
  molecule tables.
• The rdkit-like descriptor and fingerprint retrieval functions were removed from the
  API because they complicated implementation of customized descriptors.
• The apply method was simplified and a new API was clearly defined for parallel
  processing of properties over data sets. To improve molecule processing,
  a processMols method was added to MoleculeTable.

New Features

• The qsprpred.benchmarks module was added, which contains functions to easily
  benchmark
  models on datasets.
• Most unit tests now have a variant that checks whether using a fixed random seed gives
  reproducible results.
• The build pipeline now contains a check that the jupyter notebooks give the same
  results as ones that were observed before.
• Added FitMonitor, AssessorMonitor, and HyperparameterOptimizationMonitor base
  classes to monitor the progress of fitting, assessing, and hyperparameter
  optimization, respectively.
• Added BaseMonitor class to internally keep track of the progress of a fitting,
  assessing, or hyperparameter optimization process.
• Added FileMonitor class to save the progress of a fitting, assessing, or
  hyperparameter optimization process to files.
• Added WandBMonitor class to save the progress of a fitting, assessing, or
  hyperparameter optimization process to Weights & Biases.
• Added NullMonitor class to ignore the progress of a fitting, assessing, or
  hyperparameter optimization process.
• Added ListMonitor class to combine multiple monitors.
• Cross-validation, testing, hyperparameter optimization and early-stopping were made
  more flexible by allowing custom splitting and fold generation strategies. A tutorial
  showcasing these features was created.
• Added a reset method to QSPRDataset, which resets splits and loads all descriptors
  into the training set matrix again.
• Added ConfusionMatrixPlot to plot confusion matrices.
• Added the searchWithIndex, searchOnProperty, searchWithSMARTS and sample
  to MoleculeTable to facilitate more advanced sampling from data.
• Assessors now have the split_multitask_scores flag that can be used to evaluate each
  task seperately with single-task metrics.
• MoleculeDataSets now has the smiles property to easily get smiles.
• A Docker-based runner in testing/runner can now be used to test GPU-enabled features
  and run the full CI pipeline.
• It is now possible to save PandasDataTables to a CSV file instead of the default
  pickle format (slower, but more human-readable).
• New RegressionPlot class WilliamsPlot added to plot Williams plots.
• Data sets can now be optionally stored in the csv format and not just as a pickle
  file. This makes it easier to debug and share data sets, but it is slower to load and
  save.
• Added ApplicabilityDomain class to calculate applicability domain and filter
  outliers from test sets.

Removed Features

• The Metric interface has been simplified in order to make it easier to implement
  custom metrics. The Metric interface now only requires the implementation of
  the __call__ method, which takes predictions and returns a float. The Metric
  interface no longer requires the implementation
  of needsDiscreteToScore, needsProbaToScore and supportsTask. However, this means
  the base functionality of checkMetricCompatibility, isClassificationMetric
  and isRegressionMetric are no longer available.
• Default hyperparameter search space file, no longer available.

v2.1.1

Change Log

From v2.1.0 to v2.1.1

Fixes

• ⚠️ Important! ⚠️ Fixed bug in predictMols where the feature_standardizer was
  not being applied to the calculated features. This bug was introduced in v2.1.0.
  Models trained with v2.1.0 are compatible with v2.1.1, make sure to update
  QSPRpred to v2.1.1 to ensure that the feature_standardizer is applied when
  predicting on new molecules.

Changes

New Features

Removed Features

v2.1.0

Change Log

From v2.0.1 to v2.1.0.a2

Fixes

• fixed error with serialization of the DataFrameDescriptorSet (#63)
• Papyrus descriptors are not fetched by default anymore from the Papyrus adapter, which caused fetching of unnecessary data.
• A potential bug in new version of pandas broke scaffold generation so a workaround was implemented.

Changes

• QSPRModel.evaluate moved to a separate class EvaluationMethod in qsprpred.models.interfaces, with subclasses for cross-validation and making predictions on a test set in qsprpred.models.evaluation_methods (CrossValidation and EvaluateTestSetPerformance respectively).
• QSPRModel attribute scoreFunc is removed.
• 'qspr/models' is no longer added to the output path of QSPRModel.save, allowing for complete control over the output path.
• SKlearnMetrics.supportsTask now uses a dictionary like dict[ModelTasks, list[str]] to map tasks to supported metric names. (#53)
• GBMTRandomSplit and ScaffoldSplit now use the GBMTDataSplit to create balanced splits. RandomSplit still functions the same way as a completely random test split.
• PCMSplit replaces StratifiedPerTarget and is compatible with RandomSplit, ScaffoldSplit and ClusterSplit.
• DuplicatesFilter refactored toRepeatsFilter, as it also captures scenarios where triplicates/quadruplicates are found in the dataset. These scenarios are now also covered by the respective UnitTest.
• The versioning scheme of development snapshots has changed from devX to alphaX/betaX, where X is an integer that increments with each release.
• The following model class have been renamed and moved:
  • models.models.QSPRsklearn > models.sklearn.SklearnModel
  • deep.models.QSPRDNN > extra.gpu.models.dnn.DNNModel
  • extra.models.pcm.ModelPCM > extra.models.pcm.PCMModel
  • extra.models.pcm.QSPRsklearnPCM > extra.models.pcm.SklearnPCMModel
• The command line interface modules now use input and output file paths instead
  of automatically placing all files in a subfolder qspr, allowing for more
  control over the output and input paths.

New Features

• GBMTDataSplit - parent class to create globally balanced splits with the gbmt-split package.
• ClusterSplit - splits data based clustering of molecular fingerprints (uses GBMTDataSplit).
• Raise error if search space for optuna optimization is missing search space type annotation or if type not in list.
• When installing package with pip, the commit hash and date of the installation is saved into qsprpred._version
• HyperParameterOptimization classes now accept a evaluation_method argument, which is an instance of EvaluationMethod (see above). This allows for hyperparameter optimization to be performed on a test set, or on a cross-validation set. (#11)
• HyperParameterOptimization now accepts score_aggregation argument, which is a function that takes a list of scores and returns a single score. This allows for the use of different aggregation functions, such as np.mean or np.median to combine scores from different folds. (#45)
• A new tutorial adding_new_components.ipynb has been added to the tutorials folder, which demonstrates how to add new model to QSPRpred.
• A new function Metrics.checkMetricCompatibility has been added, which checks if a metric is compatible with a given task and a given prediction methods (i.e. predict or predictProba)
• In EvaluationMethod (see above), an attribute use_proba has been added, which determines whether the predict or predictProba method is used to make predictions (#56).
• Add new descriptorset SmilesDesc to use the smiles strings as a descriptor.
• New module early_stopping with classes EarlyStopping and EarlyStoppingMode has been added. This module allows for more control over early stopping in models that support it.
• Add new descriptorset SmilesDesc to use the smiles strings as a descriptor.
• Refactoring of the test suite under qsprpred.data and improvement of temporary file handling (!114).
• PyBoostModel - QSPRpred wrapper for py-boost models. Requires optional pyboost dependencies.
• ChempropModel - QSPRpred wrapper for Chemprop models. Requires optional deep dependencies.
• The data_CLI argument --log_transform (-lt) has been changed to --transform_data (-t), which now accepts a number of transformations to apply to the target data. Available transformations are log, log10, log2, sqrt, cbrt, exp, exp2, exp10, square, cube, reciprocal.
• New data_CLI, model_CLI and predict_CLI argument --skip_backup (-sb) to skip the backup of the output files. WARNING: This will overwrite existing files.

Removed Features

• StratifiedPerTarget is replaced by PCMSplit.

v2.0.1

Change Log

From v2.0.0 to v2.0.1

Fixes

• Requirement python version in pyproject.toml updated to 3.10, as older version of python don't support the type hinting used in the code.
• Corrected type hinting for QSPRModel.handleInvalidsInPredictions, which resulted in an error when importing the package in google colab.
• The predictMols method returned random predictions in v2.0.0 due to unpatched shuffling code. This has now been fixed.

Changes

New Features

• raise error if search space for optuna optimization is missing search space type annotation or if type not in list

v2.0.0

Change Log

From v1.3.1 to v2.0.0

Fixes

• more robust error handling of invalid molecules in MoleculeTable
• Not all scorers in supported_scoring were actually working in the multi-class case, the scorer support is now
  divided by single and multiclass support (moved to metrics.py, see also New Features).
• Instead of all smiles, only invalid smiles are now printed to the log when they are removed.
• problems with PaDEL descriptors and fingerprints on Linux were fixed
• fixed serialization issues with DataFrameDescriptorSet and saving and loading of MSA for PCM descriptor calculations
• the Papyrus adapter was fixed so that the quality and data set filtering options work properly (before only high quality Papyrus++ data was fetched no matter the options)
• previously, in some cases cross-validation splits might not have been shuffled during hyperparameter optimization and evaluation on cross-validation folds (this might have resulted in suboptimal cross-validation performance and bad choices of hyperparameters), a fix was made in b029e78
• score_func can now be set in QSPRModel.

Changes

• Hyperparameter optimization moved to a separate class from QSPRModel.bayesOptimization and QSPRModel.gridSearch to OptunaOptimization and GridSearchOptimization in the new module qsprpred.models.param_optimzation with a base clase HyperParameterOptimization in qsprpred.models.interfaces.
• ⚠️ Important! ⚠️ QSPRModel attribute model now called estimator, which is always an instance of alg, while alg may no longer be an instance but only a Type.
• Converting input data for qsprpred.models.neural_network.Base to dataloaders now executed in the fit and predict functions instead of in the qspred.deep.models.QSPRDNN class.
• MoleculeTable now uses a custom index. When a MoleculeTable is created a new column (QSPRID) is added (overwritten if already present), which is then used as the index of the underlying data frame.
  • It is possible to override this with a custom index by passing index_cols to the MoleculeTable constructor. These columns will be then used as index or a multi-index if more than one column is passed.
  • Due to this change, scaffoldsplit now uses these IDs instead of unreliable SMILES strings (see documentation for the new API).
• If there are invalid molecules in MoleculeTable, addDescriptors now fails by default. You can disable this by passing fail_on_invalid=False to the method.
• To support multitask modelling, the representation of the target in the QSPRdataset has changed to a list of
  TargetPropertys (see New Features). These can be automatically initizalid from dictionaries in the QSPRdataset
  init.
• A fill_value argument was also added to the predict_CLI script to allow for filling missing values in the
  prediction data set as well.
• ⚠️ Important! ⚠️ setup.py and setup.cfg were substituted with pyproject.toml and MANIFEST.in. A lighter version of the package is now the default installation option!!!
  • Installation options for the optional dependencies are described in README.md
  • CI scripts were modified to test the package on the full version. See changes in .gitlab-ci.yml.
  • Features using the extra dependencies were moved to qsprpred.extra and qsprpred.deep subpackages. The structure of the subpackages is the same as of the main package, so you just need to remember to use qsprpred.extra or qsprpred.deep instead of just qsprpred in your imports if you were using these features from the main package before.
• The way descriptors are stored in MoleculeTable was changed. They now reside in their own DescriptorTable instances that are linked to the orginal MoleculeTable
  • This change was made to allow several types of descriptors to be calculated and used efficiently (facilitated by a the DescriptorsCalculators interface)
  • Unfortunately, this change is not backwards compatible, so previously pickled MoleculeTable instances will not work with this version. There were also changes to how models handle multiple descriptor types, which also makes them incompatible with previous versions. However, this can be fixed by modifying the old JSON files as illustrated in commits 7d3f863 and 6564f02.
• 'LowVarianceFilter` now includes boundary in the filtered features, e.g. if threshold is 0.1, also features that
  have a variance of 0.1 will be removed.
• Added the ExtendedValenceSignature molecular descriptor based on Jean-Loup Faulon's work.
• removed default parameter setting scikit-learn SVC and SVR max_iter 10000.
• added matthews_corrcoef to the supported metrics for binary classification.

New Features

• New feature split ManualSplit for splitting data by a user-defined column
• The index of the MoleculeTable can now be used to relate cross-validation and test outputs to the original molecules. Therefore, the index is now also saved in the model training outputs.
• the Papyrus.getData() method now accepts activity_types parameter to select a list of activity types to get.
• Added the checkMols method to MoleculeTable to use for indication of invalid molecules in the data.
• Support for Sklearn Multitask modelling
• New class abstract class Metric, which is an abstract base class that allows for creating custom scorers.
• Subclass SklearnMetric of the Metric class, this class wraps the sklearn metrics, to allow for checking
  the compatibility of each Sklearn scoring function with the QSPRSklearn model type.
• New class TargetProperty, to allow for multitask modelling, a QSPRdataset has to have the option of multiple
  targetproperties. To support this a targer property is now defined seperatly from the dataset as a TargetProperty
  instance, which holds the information on name, TargetTask (see also Changes) and threshold of the property.
• Support for protein descriptors and PCM modeling was added.
  • The PCMDataSet class was introduced that extends QSPRDataset and adds the addProteinDescriptors method, which can be used to calculate protein descriptors by linking information from the table with sequencing data.
• Support for precalculated descriptors was added with addCustomDescriptors method of MoleculeTable.
  • It allows for adding precalculated descriptors to the MoleculeTable by linking the information from the table with external precalculated descriptors.
• The tutorial was improved with more detailed sections on data preparation and PCM modelling added.
• We agreed on and adopted a style guide for contributions to the package. This is described and exemplified in docs/style_guide.py. This is also supported by several development tools that were configured to check and automatically format the code. Instructions are included in the example file as well.
• Style guide implemented. As a consequence, some classes, methods, and attributes were renamed to comply with the style guide. Some changes were:
  • Fingerprint functions from RDKit are now also implemented. For checking the available fingerprints in qsprpred, the user can now access AVAIL_FPS through from qsprpred.data.utils.descriptor_utils.fingerprints import AVAIL_FPS.
  • Fingerprint abstract base class now moved to qsprpred.data.utils.descriptor_utils.interfaces.
  • Instance attributes are now written in camelCase, and method arguments are snake_case. As an example of this, the old parameter descsets from MoleculeDescriptorsCalculator is now renamed as desc_sets, stored as the attribute self.descSets. Functions are still written in snake_case.