Loading…
Protein-sequence-conditioned mixture-of-experts language model that designs small molecules and emits an explicit synthesis route for each one.
No providers recorded yet. Browse all providers
A target-aware generative model can propose a molecule with a beautiful docking score that no chemist can make. The usual remedy is to generate first and filter for synthesizability afterwards, which discards most of the output and leaves the survivors with no route to the bench. OmniSyn removes the filter by changing what the model emits: instead of a SMILES string, its decoder writes a sequence of synthesis actions — select a purchasable building block, apply a reaction template — so the output is a reaction trace whose terminal product is the designed compound. A molecule the model cannot build is one it cannot express.
The second constraint OmniSyn drops is structure. Pocket-conditioned designers such as TargetDiff need a resolved or predicted binding site with a defined pocket, which excludes much of the proteome and adds a structure-prediction step to every campaign. OmniSyn conditions on the target's amino-acid sequence alone — the premise shared with sequence-based generators like ProtoBind-Diff — while recovering the target-awareness that sequence-only methods have historically traded away.
The preprint was posted in September 2026 by a group led from Tongji University with collaborators at HKUST, East China Normal University, Monash University and the Zhongguancun Institute of Artificial Intelligence. One parameter set serves three design modes, selected by what is fed in: a protein sequence alone requests de novo ligand generation, a molecule alone requests synthesizability projection, and a protein plus a scaffold requests hit-to-lead (H2L) optimization.
OmniSyn couples a reaction-based autoregressive decoder to a Protein-Ligand Interaction Mixture-of-Experts (PLIM) conditioner. Protein sequences are embedded by a frozen ESM-2; seed molecules and scaffolds by a frozen atom-level molecular encoder. Four PLIM blocks maintain coupled single-token and pair tracks — a low-rank outer-product update, axial pair attention, a task-biased top-two sparse MoE layer, then pair-biased attention back to the single track — and masked mean pooling yields a 1,024-dimensional condition vector for the decoder. Training tables were built from Papyrus and ChEMBL bioactivity records mapped to reviewed UniProt sequences and standardized with RDKit, with the 35 MolGenBench evaluation targets excluded. Pre-training is route self-distillation against an ECFP4-conditioned teacher decoder; post-training freezes the decoder and applies multi-objective reinforcement learning to PLIM alone.
On those 35 unseen targets, OmniSyn reached a de novo SMILES-level hit fraction of 0.255% and hit rate of 0.480%, against 0.028% and 0.024% for the strongest sequence-based baseline, TamGen, plus a scaffold-level hit fraction of 2.517%. Only 26% of targets fell in the lowest Target-Aware Score band, where most baselines place over 92%. In H2L optimization it reached scaffold and SMILES hit rates of 10.011% and 0.163%. Independent AiZynthFinder retrosynthesis solved 68.47% of de novo molecules and 71.92% of optimized ones. Raising the sampling budget to one million molecules per target lifted scaffold and SMILES hit fractions to 13.42% and 1.67% with no saturation. MolGenBench was developed by the same laboratory, so these comparisons are not fully independent.
The three modes map onto consecutive stages of a small-molecule program. Sequence-only de novo generation suits targets that structure-based pipelines cannot reach, and the proteome-wide library gives a starting point for targets nobody has screened. H2L optimization takes a confirmed hit and its target and proposes elaborations that preserve the scaffold while naming the coupling chemistry, which a synthesis chemist can act on directly. Projection sits downstream of any other generator: candidates that fail retrosynthetic planning are redirected toward accessible chemical space with their pharmacophore largely intact.
OmniSyn's contribution is architectural rather than a leaderboard position: it argues that synthesis planning belongs inside the generative process rather than after it, and that a sequence-only conditioner can carry enough target information to compete with pocket-conditioned design. The 2.7-billion-molecule library is the largest reported target-specific generative collection, but it is not distributed with the preprint. The work awaits peer review and is released under a CC BY-NC-ND license; the authors state that code will be published on acceptance, and no weights, training data or generated molecules are currently downloadable, so the results cannot yet be reproduced independently.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.