Autoregressive graph generator that flattens molecules into token sequences, letting a decoder-only transformer sample valid structures in one pass.
No providers recorded yet. Browse all providers
Generative models for molecular graphs have converged on diffusion. Diffusion handles permutation symmetry gracefully, but it pays for that in compute: every sample requires hundreds of denoising steps over a dense node-pair representation, and the strongest variants lean on hand-engineered structural features that are expensive to compute. AutoGraph takes the opposite bet — that a graph can simply be written down as a sequence and modeled the way a language model models text.
Developed by Dexiong Chen, Markus Krimmel, and Karsten Borgwardt at the Max Planck Institute of Biochemistry (Borgwardt Lab) and presented at NeurIPS 2025, AutoGraph flattens an attributed graph into a random sequence of tokens through a reversible procedure, then trains a decoder-only transformer to predict the next token. The flattening is chosen so that every prefix of a sequence corresponds to an induced subgraph of the whole — the structural analogue of a sub-sentence. That property is what turns an implementation detail into a capability: conditioning generation on a required substructure means seeding the sequence with the tokens for that substructure and continuing, with no retraining and no auxiliary conditioning network.
The framework is domain-general — any attributed graph will do — but molecular generation is one of its two headline benchmark tracks, and the repository is tagged for it. The authors position the work explicitly as groundwork for graph foundation models.
The architecture is a standard decoder-only transformer trained with next-token prediction; the novelty lives in the tokenization rather than the network. Evaluation covers synthetic structural benchmarks — Planar graphs and stochastic block models, where validity, uniqueness, and novelty are measured directly — and molecular benchmarks on QM9, MOSES, and GuacaMol, following each suite's standard evaluation protocol. AutoGraph reaches state-of-the-art results across both tracks while generating up to 100 times faster and training 3 times faster than leading diffusion baselines, measured on a single NVIDIA H100. On the synthetic sets it is one of only two methods achieving strong structural validity, and prior autoregressive graph models fail these tests outright. Code and pretrained checkpoints are public under a BSD 3-Clause license.
For chemistry, AutoGraph is a de novo molecular generator whose speed makes large virtual library enumeration affordable and whose prefix conditioning maps cleanly onto scaffold hopping and R-group exploration: fix the core as a prefix, sample completions. Beyond molecules, the same model applies to any attributed graph a life scientist works with — biological networks, polymer topologies, reaction graphs — since nothing in the method is chemistry-specific. Groups already running transformer infrastructure benefit most, because AutoGraph reuses the standard autoregressive training and sampling stack rather than requiring a diffusion pipeline.
AutoGraph's argument is that the machinery that made language modeling scale — next-token prediction, prefix conditioning, transfer from pretraining — carries over to graphs once the serialization is chosen carefully, and that the diffusion detour was not necessary for quality. If graph foundation models materialize, this sequence-first framing is a plausible substrate for them. The honest caveats are scope and maturity: molecular chemistry is a benchmark domain here rather than the design target, so there is no built-in notion of synthesizability or 3D geometry, and adoption remains early relative to the established diffusion-based generators it competes with.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.