Uni-Mol3

Most machine learning models for chemistry describe one molecule at a time. Chemical reactions, however, are inherently multi-molecular events: reactants, catalysts, solvents, reagents, and products interact as a system, and the outcome depends on spatial features such as steric hindrance and chiral induction that a SMILES string cannot express. Uni-Mol3 targets that gap, treating an entire reaction — not a single compound — as the modeling unit.

Developed by DP Technology with the AI for Science Institute and Peking University, and released as a preprint in July 2025, Uni-Mol3 extends the Uni-Mol series from single-molecular representation learning to multi-molecular reaction modeling. Its central idea is a molecular tokenizer that quantizes 1D atom types, 2D bond graphs, and 3D conformer coordinates into a discrete vocabulary, producing a "3D-aware molecular language" that a sequence-to-sequence model can read and generate in the same way a language model reads text.

That representation is paired with a two-stage pretraining curriculum — first molecular, then reactional — followed by prompt-conditioned fine-tuning. A single pretrained backbone therefore serves forward product prediction, retrosynthesis, reaction condition generation, and yield regression, with the task selected by a prefix token rather than by a separate architecture.

#Key Features

• 3D-aware molecular tokenizer: Uni-Tokenizer encodes atom types, molecular graphs, and ETKDG-generated 3D coordinates into discrete tokens, trained by atom-type reconstruction and coordinate denoising so that a token reflects an atom's local geometric environment.
• Two-stage pretraining: Molecular pretraining learns valid-molecule "grammar" through atom-level masking and next-token prediction; reaction pretraining then applies molecule-level masking across full reaction records to learn how reactant, condition, and product sets relate.
• Prompt-aware multi-task fine-tuning: A task-specific prefix token routes the same backbone to four reaction tasks with little or no change to the output layers, and mixed-task fine-tuning improves retrosynthesis and condition generation over single-task training.
• Robust generation: Across benchmarks the model produces the lowest rate of RDKit-invalid outputs among the compared sequence models, including 0.15% on USPTO-MIT product prediction and 0.06% on USPTO-50k retrosynthesis.
• Released weights across the pipeline: The tokenizer, both pretraining checkpoints, and task-specific fine-tuned checkpoints are distributed publicly under an MIT-licensed codebase.

#Technical Details

Uni-Mol3 uses a T5-style encoder-decoder backbone with 8 encoder layers, 8 decoder layers, a hidden dimension of 768, and 8 attention heads. The tokenizer is initialized from the 84M Uni-Mol2 checkpoint and uses finite scalar quantization over a two-track transformer encoder and decoder. Molecular pretraining runs for 1,000,000 steps on the ~19M-molecule Uni-Mol dataset drawn largely from ZINC and ChEMBL; reaction pretraining runs for 1,500,000 steps on Pistachio-full, 11,973,789 patent-derived reactions filtered from NextMove Software's 16.7M-reaction Pistachio database. Training used 8 NVIDIA H100 GPUs.

Evaluation spans 10 datasets and 4 task families. On USPTO-MIT product prediction Uni-Mol3 reaches 90.8% top-1 accuracy in the reactant-condition separated setting and 89.6% in the harder mixed setting, ahead of T5Chem (90.2% / 88.9%) and Chemformer. On Pistachio-RS retrosynthesis it reaches 76.9% top-1; on the smaller USPTO-50k it reaches 49.0%, below Chemformer's 52.3% while leading on edit distance, fingerprint similarity, and validity. For condition generation it leads on USPTO-Condition (30.5% top-1) and is close behind T5Chem on USPTO-500-MT. On the Buchwald-Hartwig yield dataset it takes the best score on 11 of 12 metrics across four out-of-sample splits.

#Applications

The model suits synthesis planning and process chemistry workflows: proposing plausible products for a set of reactants, generating retrosynthetic disconnections for a target, suggesting catalysts, solvents, and reagents when conditions are unspecified, and estimating yield before committing bench time. Medicinal and materials chemists benefit most directly, particularly in the condition-generation setting, where the model is not restricted to a fixed catalog of catalyst and solvent slots. Its cross-dataset transfer results make it a reasonable starting point for fine-tuning on proprietary in-house reaction collections.

#Impact

Uni-Mol3 shows that discrete 3D tokenization can substitute for SMILES as the input language of reaction models, and that reaction-level pretraining transfers to out-of-distribution chemistry — the clearest evidence being the Buchwald-Hartwig splits, where test additives are absent from training. Its gains over strong baselines are consistent but often modest in top-1 accuracy, with the larger margins appearing in edit distance, validity, and generalization. The work is a preprint awaiting peer review, and the Pistachio corpus used for reaction pretraining is commercially licensed and cannot be redistributed, so full reproduction of the pretraining stage requires a Pistachio license even though preprocessing scripts and all downstream checkpoints are public.