DBMol

Structure-based drug design has largely relied on generative models trained directly on protein–ligand complexes, a strategy bounded by how few high-quality co-crystal structures exist. DBMol takes a different route: rather than learning a new conditional distribution over ligands, it treats an existing biomolecular structure predictor as a differentiable scoring function and optimizes molecules against it. The framework comes from researchers at EPFL, the MRC Laboratory of Molecular Biology, and the University of Cambridge, and asks whether the representations that let Boltz-2 and AlphaFold 3 predict interactions accurately can also serve as the objective for designing ligands from scratch.

The method alternates two stages. In the optimization stage, a molecule is held as a continuous relaxation of a molecular graph — atom types and bond types as categorical probability vectors — and gradients from Boltz-2 push it toward better pocket-specific contacts, higher complex confidence, and stronger predicted affinity. That relaxation quickly drifts away from anything chemically realizable, so a projection stage maps it back onto the discrete space of valid molecules using a pretrained discrete flow-matching graph generator, DeFoG. Iterating the two stages keeps designs both high-scoring and chemically valid.

DBMol is training-free in the sense that matters for generalization: neither backbone is updated, and nothing is fit per target. The paper is explicit that it "does not perform model retraining, reinforcement learning, or inference over learned molecule distributions." Only the molecule representation is optimized, so a previously unseen pocket needs no new data, no reference ligands, and no fine-tuning.

#Key Features

• Structure predictor as the design objective: Boltz-2 is used purely as a frozen differentiable oracle, supplying gradients for a composite loss over predicted affinity, a binder classification logit, pocket contacts, and ipTM confidence.
• Optimization–projection alternation: continuous gradient steps handle the scoring objective while the flow-matching projection enforces discreteness and valence rules, decoupling "what to optimize" from "what is a molecule."
• No reference-ligand supervision: designs are produced without any known binder for the target pocket, unlike ligand- or pharmacophore-conditioned generators.
• Zero-shot transfer to new pockets: the generic molecule-graph generator and the structure predictor are both reused unchanged, so new targets are handled without target-specific retraining.
• Held-out evaluation against self-confirmation bias: because Boltz-2 is the optimization signal, results are also scored with independent metrics including AlphaFold 3 confidence, interface PAE, and Vina docking.

#Technical Details

Each design run uses a single fixed configuration: 50 optimization iterations, after which the projection stage samples 2,048 candidate graphs, discards chemically invalid ones, and retains 100 valid molecules. The projection pass takes roughly 10 seconds, while a Boltz-2 backward pass costs about 30 seconds, making the structure predictor the dominant expense. On the seven-target LIT-PCBA protocol, DBMol raises the Boltz-2 success rate — favorable predicted affinity plus sufficient pocket specificity — from 0.13 for unconditional DeFoG generation to 0.47, and improves AF3-predicted pocket coverage from 0.39 to 0.51 while holding diversity at 0.89, the highest among the methods compared. On held-out metrics DBMol is competitive rather than dominant: its AF3 success rate of 0.92 sits with the pack, and its mean rank of 3.83 trails the ligand-conditioned SynFormer at 1.83, a gap the authors attribute to operating without reference-ligand supervision. Beyond the benchmark, the framework was applied to two de novo targets with no established small-molecule chemistry: the GPCR LGR4 and the immune checkpoint CD47.

#Applications

DBMol is aimed at the early hit-generation problem in structure-based drug discovery, particularly for pockets where no reference ligand exists to condition on — orphan GPCRs, protein–protein interaction interfaces, and other targets considered chemically underexplored. Because the approach requires only a pocket definition and a pretrained structure predictor, it can be pointed at targets as soon as a structure is available.

#Impact

The contribution is methodological: it reframes a co-folding model from a predictor into an optimizable objective, showing that gradients through a structure predictor carry enough signal to drive molecule generation rather than merely rank finished candidates. That reframing also sidesteps the data bottleneck the authors identify as the field's central constraint, the scarcity of high-quality protein–ligand binding datasets. The honest limits are clear in the paper: the objective remains a surrogate, gains are largest on the very metric being optimized, molecules are generated as 2D graphs rather than 3D poses, no experimental validation is reported, and the work is a preprint awaiting peer review. No code or weights have been released.