ChemDual

ChemDual is a chemistry-specialized large language model for forward reaction prediction and retrosynthesis, developed by researchers at Xiangtan University, Hunan University, and Hunan Normal University, and presented at IJCAI 2025. Given a set of reactants it predicts the product; given a target molecule it proposes the precursors needed to make it. Both tasks sit at the center of medicinal chemistry workflows, where a designed compound is only useful if somebody can actually synthesize it.

General-purpose LLMs handle these tasks poorly for two reasons the authors set out to address. First, there is no large instruction corpus that teaches a model how molecules come apart and go back together — reaction datasets are expensive to curate and comparatively small. Second, forward reaction and retrosynthesis are near-inverses of one another, but conventional fine-tuning treats them as unrelated tasks and discards that structure.

ChemDual's central idea is to recast the pair as recombination and fragmentation. Because a molecule can be split into fragments and reassembled without any experimental reaction record, this reframing turns cheap, unlabeled molecular structures into a very large source of supervision, which is then transferred to the scarce, expensive reaction data through joint training.

#Key Features

• Dual-task learning: Forward reaction and retrosynthesis are optimized jointly alongside the recombination/fragmentation surrogate, so gradients from each direction reinforce the other rather than competing.
• 4.4M-instruction synthesis corpus: Molecule–fragment pairs generated with the BRICS decomposition algorithm supply instruction data at a scale that curated reaction datasets cannot reach.
• Multi-scale tokenizer: The LLaMA tokenizer is extended with 16 dummy-atom types, 180 common functional groups, and fragment delimiter tokens, so the model reads chemistry at the group level instead of character by character.
• Docking-validated generation: Compounds generated by the model were evaluated by molecular docking and showed diverse, strong predicted binding affinity against protein targets, connecting the synthesis task to downstream design.
• Open code under Apache-2.0: The training and export pipeline is public, although the fine-tuned ChemDual checkpoint has not been released.

#Technical Details

ChemDual fine-tunes LLaMA-3.1-8B-Instruct. The instruction corpus is built from the roughly 20 million SMILES strings in ChEMBL-34, filtered to about 2.2 million valid molecules and then fragmented with BRICS to yield 4.4 million recombination and fragmentation instructions. On the Mol-Instruction benchmark ChemDual reaches an exact-match score of 0.869 for reaction prediction (BLEU 0.991, RDK fingerprint similarity 0.964) and 0.670 for retrosynthesis (BLEU 0.976, RDK FTS 0.901). On USPTO-50K it reaches 46.25% top-1, 66.95% top-5, and 77.42% top-10 accuracy; a variant coupling ChemDual with a Retroformer backbone raises top-1 accuracy to 49.95%.

#Applications

The model targets computer-aided synthesis planning: proposing synthetic routes for a hit compound, checking whether a generatively designed molecule is plausibly makeable, and filling in expected products for a planned reaction. Because both directions live in one instruction-following model, it slots naturally into agent-style drug discovery pipelines where a design step and a synthesizability check alternate. The docking analysis reported in the paper points at the same workflow from the other end — generating candidate compounds and scoring them against a protein target.

#Impact

ChemDual is an argument that the bottleneck in chemical LLMs is supervision, not scale: by inventing a self-supervised proxy task that shares structure with the real one, it extracts usable signal from unlabeled molecule libraries and beats both single-task specialist models and general open-source LLMs on standard benchmarks. The main practical limitation is availability — the repository publishes code under Apache-2.0, but the fine-tuned weights and the 4.4M-instruction dataset are both listed as forthcoming, so the results are not yet directly reproducible by downloading a checkpoint. Neither a model card nor a data card has been published.