CByG

Generative models for structure-based drug design build 3D ligands directly inside a protein binding pocket, and the field has largely graded itself on a single number: predicted binding affinity. CByG (Controllable Bayesian Flow Network with Integrated Guidance) is built around the argument that this is the wrong objective for practical medicinal chemistry, where a candidate must also be synthesizable in a real laboratory and selective for its intended target over close homologs. It comes from researchers at Yonsei University, UBLBio, and Kangwon National University.

The technical contribution is a mechanism for steering generation toward several properties at once. Classifier-guidance techniques for diffusion models differentiate a property predictor with respect to the noisy intermediate sample. That works for continuous atom coordinates but degrades for categorical atom types, where the intermediate state is not naturally differentiable and the predictor is being asked to score a chemically meaningless partial molecule. Bayesian Flow Networks avoid this by running the generative trajectory through the parameter space of the distributions over atom types and positions, which stays continuous throughout. CByG establishes that the BFN Bayesian update is equivalent to a score-based gradient step via Tweedie's formula, then uses that equivalence to inject property gradients into both the continuous and the categorical channels of generation.

The paper pairs the model with a stricter evaluation protocol aimed at the failure modes it identifies: three independent docking engines rather than one, physical validity checks, retrosynthetic search in place of a heuristic synthesizability score, and a kinase-panel test of selectivity. It joins pocket-conditioned 3D generators such as Apo2Mol and shares the property-controllable framing of CoMole.

#Key Features

• Gradient guidance inside a Bayesian flow network: Property gradients are applied in the BFN's continuous parameter space, so discrete atom types receive the same well-behaved guidance signal as continuous coordinates.
• Uncertainty-aware guidance strength: A Bayesian neural network predictor returns both a property estimate and its variance, decomposed into aleatoric and epistemic terms, and the guidance signal is scaled by that confidence.
• Multi-property control at inference: Binding affinity, synthetic feasibility, and target selectivity are steered on one fixed generative checkpoint, without retraining a model per objective.
• Retrosynthesis-grounded evaluation: Generated molecules are scored by Monte Carlo tree search over retrosynthetic routes rather than by a heuristic synthetic accessibility score alone.

#Technical Details

The generative backbone inherits the architecture and pretrained weights of MolCRAFT, a Bayesian Flow Network for pocket-conditioned molecule generation. Training and evaluation use CrossDocked2020, which after RMSD filtering and a 30% sequence-identity split yields 100,000 training protein-ligand pairs and 100 held-out test pockets, with 100 molecules sampled per test pocket. The property predictor is trained separately on CrossDocked2020 complexes to map a 3D protein-ligand structure to a scalar score combining docking score and synthetic accessibility. It adapts a graph transformer by removing the equivariant head, yielding an SE(3)-invariant network of 16 attention heads with three invariant layers per block at hidden dimension 64.

On the 100 test pockets, CByG reaches a mean Vina docking score of -9.16 against -7.59 for MolCRAFT, -8.39 for DecompDiff, and -7.80 for TargetDiff, with the ranking preserved under SMINA (-9.61) and GNINA (-8.33). Its high-affinity rate (fraction of generated ligands beating the reference ligand) is 93.6% versus 63.4% for MolCRAFT, mean synthetic accessibility is 0.84, and PoseBusters validity is 94.9%. Retrosynthetic search solves 48.7% of generated molecules, above the 41.0% solved for the reference ligands. On a selectivity set assembled from binding data for 38 kinase inhibitors across a panel of 317 kinases, CByG succeeds on 78.3% of cases against 68.6% for a guided-diffusion baseline. Chemical diversity is 0.61, below several unguided baselines, a direct consequence of conditioning generation on narrower property targets.

#Applications

CByG targets hit generation in structure-based campaigns where a single objective is not enough: a project team that already has a pocket structure can bias the generator toward tractable chemistry and away from an off-target paralog in the same sampling run, rather than filtering a large unguided library afterwards. The kinase selectivity experiment is the clearest example, since kinase inhibitor programs routinely fail on cross-reactivity within the kinome rather than on potency against the intended target.

#Impact

CByG's broader contribution is diagnostic as much as generative. By replacing single-engine docking with a three-engine comparison and heuristic synthesizability with retrosynthetic search, it shows that models ranked highly on conventional SBDD leaderboards can produce molecules with no practical synthetic route, and that synthetic accessibility scores correlate poorly with retrosynthesis success. The authors report that roughly half of the molecules generated even by their own model remain synthetically infeasible, and treat that as the field's next open problem. The work is a preprint that has not completed peer review, and no code or weights have been publicly released, so reproduction depends on reimplementing the guidance scheme on top of MolCRAFT's public checkpoints.