Hypergraph Transformer pretrained on 0.6M chemical reactions, learning representations for both reaction outcomes and molecular property prediction.
No providers recorded yet. Browse all providers
A chemical reaction is defined by what changes: a few bonds break or form and electrons are reassigned among a few atoms, while the rest of each molecule and the surrounding solvent, catalyst and reagents shape how that happens. Pretraining on isolated molecules sees none of this, so its objectives carry little information about chemical transformation, which limits transfer to reaction-level tasks such as yield, selectivity or condition prediction.
LARK (Learning Atom-molecule-Reaction Knowledge) addresses this by using reactions themselves as the supervision signal. For each atom-mapped reaction it computes bond-electron matrices for the reactant and product sides, whose difference (ΔBE) records the net change in bond order and local electron assignment. The model is trained to reconstruct that change from either side of the reaction: given the reactants it predicts +ΔBE, and given the products it predicts −ΔBE, with shared weights and a consistency loss linking the two views. Because the changed atoms are only a small part of a multicomponent reaction, a role-aware reaction hypergraph Transformer lets each molecule exchange information with the others in the reaction, and gated feedback returns that context to the atom states used for reconstruction.
LARK was developed by Jianbo Qiao, Leyi Wei and colleagues at Shandong University and Macao Polytechnic University with collaborators in Beijing, Chengdu and Tianjin, and released as a bioRxiv preprint in September 2026.
Each molecule is encoded as a 2D attributed graph with a learned molecular super-token; atom features include element, charge, valence, hybridization, chirality, aromaticity and partial charge, and shortest-path distances and bond features supply attention biases. The released configuration uses 256-dimensional, six-layer encoders. Pretraining data came from the ORDerly release of the Open Reaction Database: about 600,000 reactions (roughly 514,000 train, 27,000 validation and 60,000 held-out test) covering 0.88 million unique molecular components. Before partitioning, every reaction matching a downstream validation or test item at the full-reaction, product-side or product-core level was removed from the corpus.
Downstream evaluation covered 42 tasks: 11 reaction and synthesis tasks, nine MoleculeNet tasks and 22 Therapeutics Data Commons ADMET tasks. On Li/Hong radical C–H functionalization, LARK reached an R² of 0.983 for activation free energy and ranked the true preferred site first in 95.98% of reaction groups. It reached 0.881 R² on held-out Suzuki yields, 0.964 accuracy on 46-class Schneider reaction classification, and 52.75% top-15 exact condition-set accuracy on USPTO-Condition. Across six MoleculeNet classification tasks it had the highest mean AUROC (0.779) among 26 methods reporting all six, ahead of M2UMol at 0.755, and it outperformed MolE on 13 of 22 ADMET task means. Patent-yield regression remained hard (R² 0.121 on USPTO Gram). An eight-variant ablation showed that adding unmapped reaction context improved 36 of 39 task means, and attention in the final molecular-encoder layer was enriched at the atoms that change.
LARK is aimed at computational and medicinal chemists who need one representation for both the molecules they make and the reactions that make them. Its reaction representations support yield prediction, site- and stereoselectivity estimation, reaction classification from few labelled examples, and recommending catalysts, solvents and reagents. Its molecular encoder, fine-tuned on single-molecule data, supports physicochemical, toxicity, metabolism and pharmacokinetic prediction for early drug discovery. The repository provides pretraining code and two fine-tuning examples but no pretrained checkpoints, so users must pretrain the model themselves.
LARK makes a case that chemical transformations are useful supervision for representation learning even when the downstream input is a single molecule with no reaction partners. The reaction-centre analyses give that claim a mechanistic basis: suppressing the most-attended atom state raised reconstruction error well beyond matched random suppression, which ties the learned atom representations to the bond and electron changes they were trained to predict. All results are from a preprint that has not been peer reviewed, and several comparisons rely on published baseline numbers with differing data partitions.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.