DeepChem

Transforms chemical structures into tensor-ready inputs. DeepChem provides a comprehensive suite of featurizers that convert SMILES strings, PDB files, and molecular graphs into fixed-length vectors or graph representations.

• ConvMolFeaturizer: Generates graph convolutional inputs capturing atom-level connectivity.
• WeaveFeaturizer: Produces pair-features for atoms and bonds, enabling more expressive graph convolutions.
• CoulombMatrix: Encodes 3D atomic coordinates and nuclear charges into a rotationally invariant matrix for quantum chemistry tasks.
• Circular Fingerprints: Implements Extended-Connectivity Fingerprints (ECFP) for baseline molecular similarity comparisons.

Implements state-of-the-art Message Passing Neural Networks (MPNNs) for learning directly on molecular graphs. These models treat atoms as nodes and bonds as edges, learning representations by iteratively aggregating information from neighboring atoms.

• GraphConvModel: A standard graph convolutional network for predicting molecular properties like solubility or toxicity.
• WeaveModel: A more expressive variant using pair-wise atom features for detailed chemical environment modeling.
• DAGModel: A directed acyclic graph network designed for structured output prediction, such as retrosynthetic pathway analysis.
• MPNNModel: A flexible implementation of the general message passing framework, configurable for custom message and update functions.

Abstracts the entire deep learning lifecycle into a consistent, scikit-learn-compatible API. A model can be instantiated, trained, and evaluated with minimal boilerplate, accelerating experimental iteration.

• model.fit(train_dataset): Handles batching, epoch iteration, and checkpointing automatically.
• model.predict(test_dataset): Returns NumPy arrays of predictions for downstream analysis.
• model.evaluate(test_dataset, metrics): Computes standard regression and classification metrics like ROC-AUC, RMSD, and Pearson's R.
• Hyperparameter Tuning: Integrates with Hyperopt for automated Bayesian optimization of model architecture and learning rates.

Manages scientific data of arbitrary scale with disk-memory hybrid storage. The DiskDataset and ImageDataset classes handle sharding, shuffling, and on-the-fly featurization, preventing memory exhaustion with large compound libraries.

• DiskDataset: Stores featurized data in sharded NumPy arrays on disk, enabling multi-GPU training on datasets too large for RAM.
• NumpyDataset: A simple in-memory dataset for rapid prototyping with small-to-medium datasets.
• ImageDataset: Specialized handling for microscopy and medical imaging data.
• Data Splitters: Implements domain-aware splitting strategies like scaffold splitting and butina clustering to ensure realistic evaluation that tests generalization to novel chemotypes.

Provides a curated collection of pre-trained models for transfer learning and benchmarking. These models capture generalizable chemical knowledge from massive datasets, reducing the data requirements for specialized tasks.

• ChemBERTa: A Transformer model pre-trained on millions of SMILES strings via masked language modeling, fine-tunable for property prediction.
• MolGAN: A generative adversarial network for de novo molecular design, capable of generating valid SMILES strings with desired properties.
• GROVER: A large-scale graph Transformer pre-trained on 10 million unlabeled molecules, achieving state-of-the-art results on molecular property prediction benchmarks.
• Tox21 & HIV Benchmarks: Includes standardized datasets and pre-trained weights for the MoleculeNet benchmark suite, enabling reproducible research.

Extends graph neural networks to bipartite interaction spaces. DeepChem's GraphConvModel can be adapted for Drug-Target Interaction (DTI) prediction by constructing heterogeneous graphs where drug and protein nodes are connected by interaction edges.

• Proteochemometric Modeling: Featurizes both the ligand (via graph convolutions) and the protein target (via sequence-based 1D convolutions) to predict binding affinity.
• Binding Affinity Regression: Trains models on datasets like PDBbind to predict the pKd or pKi of a protein-ligand complex directly from structure.
• Virtual Screening Pipelines: Chains a featurizer, a pre-trained affinity model, and a data loader to score millions of compounds against a target protein, ranking them by predicted potency.