Version 3.0.2

Fixes

• Fix device kwarg colliding with the default in neural net constructors
• Fix qsar()/pcm() crashing on insufficient training data for non-year splits
• Fix missing source_path in from_dataframe test calls
• Cast Year to int32 when converting bioactivity data to Parquet
• Raise a clear ImportError instead of NameError when torch/skorch are missing
• Coerce Year column to numeric before temporal split comparison

Additions

• Add download_if_missing option to PapyrusDataset.from_dataframe

Docs

• Document Windows CUDA wheel install for the dnn extra

Notebook examples overhaul

• New protein_descriptors.ipynb: fetching ProDEC amino-acid-scale descriptors, the fixed-length AVG/ACC/PDT/FFT transforms, aligning short sequences with pyFAMSA for comparable raw descriptors, and pcm() modelling with both gradient boosting and the DNN classes (single-task, multi-task classifier)
• advanced_querying.ipynb: fixed the mold2 sentinel-NaN (~1e38) values not being unmasked before scaling, fixed DNN checkpoint paths colliding across MultiTaskNNRegressor runs, and added new sections covering the random/cluster/custom-cluster/custom split strategies and how to pull X_train/y_train/X_valid/y_valid arrays directly
• matchRCSB.ipynb: fixed the Colab setup cell's %%capture magic being placed after another comment line, which silently broke output suppression
• simple_examples.ipynb: added printed result counts to the similarity/substructure search examples and cleaned up stale imports/descriptor references