YuelBond

A diffusion model that builds a molecule in three dimensions hands back a cloud of atomic coordinates and nothing else. Bond orders have to be inferred afterward, and the standard tool for that, RDKit's xyz2mol routine, infers them by measuring bond lengths and angles against covalent radii and valence rules. Those rules presume the geometry is already close to right; on generative-model output it frequently is not. Given 1,000 test molecules with perturbed coordinates, RDKit could process only 217, recovering bond orders at a 64.2% F1-score.

YuelBond replaces that rule-based step with a learned one. Each molecule becomes a graph whose nodes are atoms and whose edges join atom pairs within 3 Å, carrying the interatomic distance as an edge feature. Conventional graph networks predict node labels; YuelBond redirects the architecture at edges, updating each edge from the two node embeddings, the distance, and its own prior features, so message passing sharpens the description of a bond rather than an atom. A final linear layer maps each edge embedding to one of four classes: single, double, aromatic, or triple. In the 2D setting the edges come from the molecule's own connectivity instead of a distance cutoff, so the same design covers graphs carrying no coordinates at all.

Jian Wang and Nikolay Dokholyan built YuelBond at Penn State College of Medicine, posting it to bioRxiv in May 2025; the peer-reviewed version appeared in the Journal of Chemical Information and Modeling in January 2026, by which point the lab had moved to the University of Virginia. It belongs to the same Yuel line as Yuel 2, a binding-affinity predictor, and YuelPocket, a binding-site predictor, though architecturally independent of both. YuelBond does not design molecules; it is the chemistry-validation stage that runs after something else has.

#Key Features

• Edge-centric message passing: Edge features are refined layer by layer from node embeddings, interatomic distance and the prior edge state, rather than read off atom embeddings at the end.
• Three input regimes: Separate checkpoints handle exact 3D coordinates, crude de novo generated compounds (CDGs) with perturbed geometry, and 2D graphs carrying connectivity but no bond orders.
• Sanitized and kekulized families: A parallel set of checkpoints is trained on a kekulized corpus, covering pipelines that expand aromatic rings into alternating single and double bonds.
• Per-bond probabilities: Softmax over the four bond classes gives a confidence distribution per edge, surfacing cases where 2D connectivity alone admits several valid assignments.
• Fixed checkpoints at inference: Prediction runs from a downloaded .ckpt file against new molecules; training is a separate script the user never has to touch.

#Technical Details

Training used the GEOM corpus of over 450,000 drug-like molecules, taking each one's lowest-energy conformer and splitting randomly into train, validation and test sets; molecules under 2 or over 150 atoms were dropped. The CDG regime adds Gaussian noise of σ = 0.2 Å to atomic positions before rebuilding the 3 Å neighbor graph. The released configuration stacks 16 message-passing layers of 64 hidden features with SiLU activations, layer normalization and residual connections, optimized with AdamW under a masked cross-entropy loss counting only real, non-padded edges.

Reconstructing bonds from exact coordinates reaches 98.2% accuracy and 98.4% F1; on noisy CDG geometry the model holds 92.7% accuracy and 92.7% F1; reassigning bond orders from 2D connectivity alone gives 80.1% accuracy and 78.3% F1, the hardest setting because bond lengths and angles are unavailable. The kekulized checkpoints track closely at 96.4%, 90.9%, and 76.6% F1 across the same three scenarios. The journal version adds out-of-distribution tests on 2,000 molecules each from the DiGress 2D and DecompDiff 3D generators.

#Applications

YuelBond's natural home is inside a generative drug design pipeline, after sampling and before anything that needs a valid molecule: RDKit sanitization, SMILES export, validity statistics, force-field parameterization for molecular dynamics, or docking. It also covers cheminformatics chores where connectivity is known but bond orders were never recorded, such as reading molecules out of coordinate-only file formats. The Dokholyan lab uses it in exactly this role in its own YuelDesign generator, and an unaffiliated diffusion design repository has adopted the released 3D and CDG checkpoints for the same step.

#Impact

Bond perception is usually treated as plumbing, and its failures get recorded as generator failures — a model marked down for invalid molecules when what failed was the rule-based reader downstream of it. By quantifying that gap and showing that a learned edge classifier stays accurate where the rules break, YuelBond makes the perception step an explicit, benchmarkable component rather than an assumed one. Its scope is correspondingly narrow: it assigns bond orders and nothing else, its evaluation is entirely in silico, and the 2D-only setting leaves real room for improvement on double and triple bonds, which are rare and lack the geometric cues the 3D modes use.