Models

LocalRetro is a research implementation for retrosynthesis prediction using local reaction templates, with workflows for template extraction, model training, testing and decoding predicted reactants.

Retrosynthesis

RxnMapper uses attention from an unsupervised ALBERT model to assign atom mappings to valid reaction SMILES, returning mapped reactions and confidence scores through a Python interface.

Cheminformatics · Reaction Prediction

DECIMER converts chemical structure images into predicted SMILES using EfficientNet-V2 and a transformer, with a Python prediction interface and a separately described hand-drawn model.

Chemical Structure Recognition

MolScribe is an image-to-graph model for converting molecular diagrams into chemical structures, with SMILES and molfile output plus optional atom, bond and confidence information.

Chemical Structure Recognition

TANKBind is a research model for predicting protein–ligand binding structures and affinity, with repository notebooks for prediction, dataset preparation, self-docking evaluation and virtual screening.

Drug Discovery

EquiBind is an SE(3)-equivariant model for predicting receptor binding locations and ligand poses directly from protein–ligand structures, with workflows for inference and training.

Drug Discovery

DiffDock predicts protein–small-molecule binding poses using a diffusion-based model, with single-complex and batch inference, sequence-based protein inputs, and pose confidence scores.

Drug Discovery

Official research code for E(3)-equivariant diffusion models that generate 3D molecules, with QM9 and GEOM-Drugs training workflows, sample analysis, and property-conditioned generation.

Molecular Generation

GeoDiff is a geometric diffusion model for molecular conformation generation, with official code for GEOM-based training, checkpoint sampling, and conformation and property evaluation.

Molecular Generation · Computational Chemistry

GraphAF is a flow-based autoregressive model for molecular graph generation, with a reference-code link and a README update pointing to an implementation in TorchDrug.

Molecular Generation

JT-VAE is the official Junction Tree Variational Autoencoder implementation for molecular graph generation, with VAE training code and scripts for Bayesian optimization and joint property-predictor training.

Molecular Generation

MoFlow is an invertible molecular graph generation model with documented workflows for QM9 and zinc250k training, reconstruction, latent-space sampling, interpolation and property optimization.

Molecular Generation

MolGPT trains a small custom GPT with next-token prediction on MOSES and Guacamol for unconditional and conditional molecular generation, with trained weights and Ecco-based saliency analysis linked.

Molecular Generation

ChemBERTa provides BERT-like models for chemical SMILES, with RoBERTa masked-language modelling checkpoints and notebooks for pre-training, fine-tuning and molecular property prediction research.

Molecular Property Prediction

GraphMVP is a molecular representation-learning research implementation that combines 2D topology and 3D geometry during pre-training, then uses 2D graphs for downstream classification and regression.

Molecular Property Prediction

GROVER provides pretrained molecular graph transformers and a PyTorch workflow for self-supervised pretraining, property-prediction finetuning, inference, evaluation and molecular fingerprint generation.

Molecular Property Prediction

MolCLR is a molecular contrastive-learning framework that pre-trains graph neural networks on unlabelled molecules and supports fine-tuning for downstream molecular property prediction.

Molecular Property Prediction

Uni-Mol is a 3D molecular representation learning framework with molecular and protein-pocket models, supported by related tools for property prediction, conformation modeling and docking.

Molecular Property Prediction · Molecular Generation

GemNet is a PyTorch reference implementation of a geometric message-passing model for molecular energies and atomic forces, with notebooks for training, ASE-based prediction and molecular dynamics.

Molecular Property Prediction · Computational Chemistry

DimeNet provides reference implementations of DimeNet and DimeNet++ for directional message passing on molecular graphs, with training notebooks, test-set prediction workflows and pretrained models.

Molecular Property Prediction · Quantum Chemistry