Research map / Graph learning / Graph self-supervised learning

MolHIT: Advancing Molecular-Graph Generation with Hierarchical Discrete Diffusion Models

NeurIPS 2026 · Graph learning › Graph self-supervised learning · approach: Molecular graph pretraining · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MolHIT, a molecular graph diffusion model using hierarchical discrete diffusion (mid-level chemical-group states) and decoupled atom encoding (aromaticity/charge-aware tokens) plus a project-and-noise sampler, reaching near-perfect validity and SOTA quality on MOSES among graph diffusion models.

Problem

Graph diffusion models for 2D molecular generation have low chemical validity and poor property control compared with 1D SMILES-based models, and coarse atom encodings make reconstruction of charged and aromatic atoms ill-posed.

Approach

Augment discrete diffusion with intermediate hierarchical states that group atoms by chemical role (a hierarchical discrete diffusion model with a closed-form kernel and ELBO), combine it with a decoupled atom vocabulary that separates aromatic and charged atom types, and sample with a project-and-noise sampler.

Key results

• On MOSES, MolHIT reports Quality 94.2% and Validity 99.1%, with FCD 1.03 and scaffold novelty 0.39, the best among graph diffusion models.
• On multi-property guided generation (MOSES), MolHIT reaches average MAE 0.058 (52.4% lower than Marginal+DAE) and average Pearson r 0.807, with validity 96.31%.
• On scaffold extension, MolHIT achieves Hit@1 3.92 and Hit@5 9.79 versus DiGress 2.07 and 6.41.

Surprising findings

• DAE alone lowers FCD on MOSES to 0.89 but the full MolHIT FCD (1.03) and the PN sampler raises FCD to 1.65, showing a validity-FCD trade-off.
• On GuacaMol, original DiGress (trained on filtered data) retains the best FCD (68.0) over MolHIT (54.9 is better; DiGress org. 68.0), so DAE does not always help distribution matching.

Limitations

• Model size and architecture were not scaled or improved beyond the DiGress graph transformer.
• GuacaMol model was not trained to performance saturation (40 epochs versus 1,000 for DiGress).
• FCD on GuacaMol and MOSES after PN sampler is worse than some baselines, indicating trade-offs.

Experimental setup

Datasets: MOSES GuacaMol

Benchmarks: MOSES GuacaMol

Metrics: Validity Uniqueness Novelty Filters FCD SNN Scaf Quality Scaffold Novelty Scaffold Retrieval

Baselines: VAE CharRNN SAFE-GPT GenMol DiGress DisCo Cometh DeFoG

Builds on: DiGress D3PM SEDD classifier-free guidance

Closest papers by meaning

• NExT-Mol: 3D Diffusion Meets 1D Language Modeling for 3D Molecule Generation (ICLR 2025 · similarity 0.89)
• All-atom Diffusion Transformers: Unified generative modelling of molecules and materials (ICML 2025 · similarity 0.88)
• DMol: A Highly Efficient and Chemical Motif-Preserving Molecule Generation Platform (NeurIPS 2025 · similarity 0.88)
• Graph Diffusion that can Insert and Delete (NeurIPS 2025 · similarity 0.87)
• SubgDiff: A Subgraph Diffusion Model to Improve Molecular Representation Learning (NeurIPS 2024 · similarity 0.87)
• DiffMS: Diffusion Generation of Molecules Conditioned on Mass Spectra (ICML 2025 · similarity 0.87)