Molfeat

If you're doing molecular machine learning, this gives you 100+ featurizers in one library. It handles the annoying stuff: converting SMILES to ECFP fingerprints, RDKit descriptors, or embeddings from pretrained models like ChemBERTa. The transformer API is scikit-learn compatible with built-in parallelization, and you can save/reload configs for reproducibility. One catch: it only works on Python 3.9-3.10, not 3.11+, which is frustrating if you're on a newer stack. The optional dependencies are a maze (DGL, graphormer, transformers) so expect some conda-forge wrestling for GNN models. But for QSAR pipelines or virtual screening, it beats rolling your own featurization every time.