BridgeDock uses unbound protein structures to guide AI molecular docking

Molecular docking predicts the three-dimensional arrangement of a protein bound to a small molecule, or ligand. Those structures can help researchers understand molecular interactions and guide structure-based drug design. In their BridgeDock preprint, Shikun Feng and colleagues ask how pretrained models can make better use of an available protein structure when predicting that arrangement.

The distinction matters because the authors’ target task is flexible docking: both the ligand and protein may need to change conformation. The starting protein is unbound, or “apo”; the desired endpoint is the bound, or “holo,” complex. BridgeDock makes that starting structure part of the generative trajectory.

A bridge rather than a blank slate

Co-folding models such as AlphaFold 3 generate atomic coordinates through iterative denoising, beginning with Gaussian noise and using sequence-derived information. The authors argue that this formulation does not directly exploit an apo structure as the starting point for docking.

BridgeDock instead constructs a diffusion bridge between an initial protein–ligand structure and the target bound complex. Intermediate training states combine the starting structure, the target structure and noise. At inference, the implemented sampler begins near the apo endpoint, with added noise, rather than from pure Gaussian coordinates.

A second component makes this bridge compatible with a pretrained denoiser. BridgeDock subtracts the apo contribution from each intermediate coordinate state, rescales the remainder and supplies a corresponding noise level. This presents the network with a denoising problem resembling its original training task. The apo structure also remains available through a separate conditioning input.

The practical implementation uses Protenix-Mini, an open-source AlphaFold 3–style model—not AlphaFold 3 itself. The researchers report continued training on 40,000 protein–ligand complexes for 80,000 steps, using eight A100 GPUs for approximately five days.

What the benchmark gains show

For evaluation, protein starting conformations were predicted from amino-acid sequences, and ligand conformations were generated with RDKit. Each ligand was initially placed at the geometric center of the full apo protein. BridgeDock received no binding-pocket annotations or known pocket center at test time.

On PoseBusters v2, using 20 denoising steps, BridgeDock increased the proportion of predictions with ligand root-mean-square deviation below 2 angstroms from 68.61% to 73.21% relative to the AF3-style baseline. This metric measures ligand-position error; lower values indicate closer agreement with the reference pose. Mean ligand error fell from 2.64 to 2.26 angstroms, while the median remained unchanged at 1.09 angstroms.

The authors interpret that pattern as evidence of fewer difficult failures, rather than simply better refinement of easy cases. It is not an improvement in every statistic: on the PDBBind test split, BridgeDock raised the below-2-angstrom proportion from 69.0% to 71.0% and reduced mean error, but its median error increased from 0.99 to 1.09 angstroms.

The PDBBind table also includes published results for earlier docking methods. Those comparisons require care: the earlier methods used known-pocket protocols, and three flexible docking baselines received ligand-derived pocket information. Their results therefore come from different input settings, not an identical experimental protocol.

Fewer steps, with qualifications

BridgeDock also performed competitively with very few denoising steps. On PoseBusters, one step produced 69.86% of ligand poses below 2 angstroms, compared with 69.76% for the one-step baseline using an ODE sampler. Its local structural fidelity score, lDDT, improved more substantially, from 83.91% to 90.77%.

These results support the paper’s claim about sampling efficiency, but step counts are not a measured wall-clock speedup. The paper’s checklist acknowledges that inference wall-clock times are not exhaustively reported.

Ablation experiments support the importance of combining the two main ideas. Supplying apo coordinates as auxiliary conditioning improved the baseline, but the full aligned bridge achieved stronger aggregate results. A bridge without the alignment mechanism did not reliably improve ligand accuracy.

Additional FoldBench evaluation extended the comparison to two backbone sizes. On 509 retained targets, success rates rose from 41.65% to 42.63% for Protenix-Mini and from 56.97% to 59.33% for Protenix-Base. This evaluation excluded 49 longer targets and selected the highest-confidence prediction among five per target, using one random seed.

A research tool, not decisive evidence

The authors acknowledge dependence on starting-structure quality, potentially problematic missing cofactors or modified residues, and poorly represented pocket states. They also report no error bars, confidence intervals or repeated-run significance tests, and the submission lacks a complete code-and-checkpoint release.

BridgeDock’s central contribution is a way to reuse learned structural knowledge while letting an informative starting structure shape generation. The reported benchmarks support that approach; the authors still call for broader prospective evaluation, expert inspection and experimental validation before predictions guide decisive discovery choices.