Model collection

Choose from 31 model classes for property prediction, molecular representations, and generation. Models accept SMILES strings and handle their own graph or sequence representations. Select a model below to see its parameters and method signatures. Example links open the code directly in each model’s reference page.

Predict properties

Fit on molecules and labels, then predict regression or classification targets. Start with GNN for a graph baseline or a SMILES model for a sequence baseline.

11 predictors

Model	Approach	Example
GNN	GIN or GCN message passing, with optional virtual nodes.	GNN example
GREA	Graph rationales and environment-based data augmentation.	GREA example
SGIR	Semi-supervised learning for imbalanced regression targets.	SGIR example
GRIN	Repetition-invariant graph representations for polymers.	GRIN example
BFGNN	Graph learning aligned with the Bellman–Ford algorithm.	BFGNN example
DIR	Learning invariant graph rationales.	DIR example
IRM	Invariant risk minimization with graph neural networks.	IRM example
RPGNN	Relational pooling over node permutations.	RPGNN example
SSR	Regularization for generalization across graph sizes.	SSR example
LSTM	Recurrent sequence learning from SMILES.	LSTM example
SMILES Transformer	Transformer sequence learning from SMILES.	Transformer example

Learn molecular representations

Encode molecules as vectors for downstream prediction, clustering, or similarity analysis. Self-supervised encoders train on unlabeled SMILES; supervised encoders need the labels or pairs described below.

9 encoders

Model	Training signal	Example
AttrMask	Reconstruct masked atom attributes.	AttrMask example
MoAmA	Motif-aware attribute masking and fingerprint reconstruction.	MoAmA example
GraphMAE	Graph masked autoencoding.	GraphMAE example
ContextPred	Predict molecular subgraph contexts.	ContextPred example
EdgePred	Predict whether pairs of atoms are connected.	EdgePred example
InfoGraph	Learn agreement between graph and substructure representations.	InfoGraph example
Supervised	Molecular property labels or computed molecular descriptors.	Supervised example
SAND	Pairwise similarity scores; dense or compressed molecular descriptors.	API and input format
Hugging Face	Load a pretrained sequence encoder, such as ChemBERTa or ChemGPT.	Pretrained encoder example

Generate molecules

Train a generator on molecules or load a pretrained model. Conditioning varies by method: property labels, molecular scaffolds, and scored demonstrations serve different purposes and have different input formats.

11 generators

Model	Approach and supported conditioning	Example
DemoDiff	Graph diffusion with optional molecular demonstrations and normalized score values.	DemoDiff example
GenMol	Masked diffusion over SAFE fragments, optionally retaining a supplied scaffold.	API and usage
GraphDiT	Graph diffusion transformer with optional property labels.	GraphDiT example
DeFoG	Discrete graph flow matching with optional property labels.	DeFoG example
DiGress	Unconditional discrete graph diffusion.	DiGress example
GDSS	Unconditional score-based graph diffusion.	GDSS example
JT-VAE	Unconditional junction-tree variational autoencoding.	JT-VAE example
Graph GA	Genetic search using property labels or a scoring oracle.	Graph GA example
MolGPT	Autoregressive SMILES transformer with optional properties and scaffold tokens.	MolGPT example
LSTM	Recurrent SMILES generation with optional property labels.	LSTM example
Hugging Face	NovoMolGen, MolGen, Molexar, or SAFE-GPT checkpoints; constraints depend on the model family.	SAFE-GPT example

See Installation for optional dependencies and Quickstart for complete workflows, including dataset preparation and checkpoint loading.