Graph autoencoder pairing a GNN encoder with a hypergraph-grammar decoder, so every embedding decodes to a structurally valid molecule.
No providers recorded yet. Browse all providers
Decode a point in a SMILES-based latent space and the output can be a string with an unclosed ring or a five-bonded carbon. Grammar-based decoders fix that, but earlier ones such as MHG-VAE had the opposite problem: their encoders worked by parsing a molecule into production rules, and molecules the grammar could not parse had no embedding at all. The MHG-GNN authors report that MHG-VAE failed to encode roughly half the molecules in one of their preliminary datasets.
MHG-GNN, which IBM's packaging calls MHG-GED (Molecular Hypergraph Grammar with Graph-based Encoder Decoder), separates the two jobs. The encoder is a graph neural network that reads the molecular graph directly, so any valid molecular graph gets a latent vector. The decoder does not emit atoms, bonds or characters. It emits a sequence of production rules from a Molecular Hypergraph Grammar extracted from the training molecules. Every rule application keeps each hypernode attached to exactly two hyperedges, so any complete derivation is a chemically valid structure by construction.
The model was developed at IBM Research Tokyo with collaborators from JSR Corporation. It was posted to arXiv in September 2023 and presented at the NeurIPS 2023 AI4Mat workshop. The authors present it as a first step toward a foundation model for materials science, not as a finished large-scale one. It became the molecular-graph arm of IBM's FM4M (Foundation Models for Materials) suite, alongside the SMILES-based SMI-TED.
encode() and decode() for SMILES input, and downstream models train on its embeddings without refitting the autoencoder.Atoms (9 features) and bonds (3 features) are embedded into 256 dimensions and updated by GIN-style message passing. The graph representation concatenates summed node vectors across iterations, giving 256 × (r + 1) dimensions for r iterations. For decoding, that vector passes through mean and log-variance heads into a 256-dimensional Gaussian latent. A 3-layer GRU with 384 hidden units and 128-dimensional rule embeddings then generates the production-rule sequence. Training used a β-VAE objective (β = 0.01) on A100 GPUs. The base dataset in the paper is 1,381,747 PubChem molecules, from which 16,362 production rules were extracted. The released checkpoint's model card describes its pretraining set as about 1.34 million curated PubChem molecules.
The paper evaluated frozen embeddings with LightGBM regressors on six property-prediction tasks from three proprietary industrial datasets. These were 5,000 simulated polymers (density, bulk modulus, refractive index), 7,884 simulated photoresist molecules (HOMO, LUMO) and 346 experimentally measured chromophores (near-infrared λmax). The radius was selected on validation data, and MHG-GNN reached test R² of 0.578, 0.516, 0.865, 0.896, 0.845 and 0.845. That beat 1024-bit ECFP6 fingerprints on every task and edged out 1,613 two-dimensional Mordred descriptors on every task, although the margin over Mordred is small on most. Performance peaked at 6 or 7 message-passing iterations and declined at 8.
The main use is as a graph-based molecular featurizer for small-data property prediction. Polymer, photoresist and dye datasets often have only hundreds of labels. Users encode a library once and fit a light regressor on the embeddings. Because the decoder always returns a valid structure, the same latent space can also back optimization loops that search embeddings and read out molecules. Within FM4M, the graph embeddings are combined with SMILES and SELFIES embeddings in IBM's multi-view mixture-of-experts property predictor, and IBM's FM4M demo Space loads the checkpoint for interactive use.
MHG-GNN's contribution is architectural. Pairing a graph encoder with a grammar decoder gives a molecular autoencoder that is both total, meaning every molecule gets an embedding, and sound, meaning every embedding decodes to a real molecule. The evidence behind it is modest. The paper is a short workshop contribution evaluated on six tasks drawn from proprietary datasets that are not public, and the only baselines are two classical descriptor sets rather than other pretrained molecular models. The code and a pretrained checkpoint are public under Apache 2.0 through IBM's materials repository and HuggingFace. The curated pretraining set is not distributed as a dataset, and no generation or optimization benchmarks are reported. Its lasting role is as the graph modality in IBM's later multi-view work.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.