BridgeDock: Taming Generative Co-Folding Prior for Molecular Docking with Diffusion Bridge

AI summary

Overview

Research area: AI for structural biology and computational drug discovery, specifically deep-learning–based molecular docking and biomolecular co-folding models.

Technical level: Advanced. The paper assumes familiarity with diffusion models (score-based/EDM preconditioning), Schrödinger bridges and Denoising Diffusion Bridge Models (DDBM), and the AlphaFold 3 (AF3) coordinate-denoising formulation.

Scope (one sentence): The paper introduces BridgeDock, a framework that turns a pretrained AF3-style co-folding denoiser into a flexible docking model by recasting docking as an apo-to-holo diffusion bridge that starts from the given unbound structures instead of Gaussian noise, with a reparameterization that keeps each bridge state compatible with the pretrained denoiser's noise-level interface.

What This Paper Is About

Conventional docking starts from provided structures — an apo (unbound) receptor in flexible docking, or a holo receptor in rigid docking — and must predict the bound protein–ligand complex. Modern co-folding models such as AlphaFold 3, Boltz-1 and Protenix are far more accurate but generate complexes from sequence-derived inputs (MSAs and optional templates) starting from pure Gaussian noise, so they never use the provided apo structure as a structural prior. That mismatch leaves them with a much larger conformational search space and weaker robustness in difficult or few-step settings. BridgeDock's goal is to keep the strong pretrained co-folding prior while explicitly building the docking input into the generative trajectory, by modeling the apo-to-holo transition as a diffusion bridge.

Key Contributions

1. An apo-to-holo diffusion bridge formulation for flexible docking. Docking is recast as a stochastic process anchored at the apo-conditioned initial structure (apo protein coordinates plus an initial ligand placement) and terminating at the holo target, using the variance-exploding (VE) bridge of Denoising Diffusion Bridge Models rather than AF3's noise-to-data process.
2. An AF3-aligned bridge parameterization. Because raw bridge states change both the coordinate parameterization and the noise-level interface the pretrained denoiser was trained on, the authors algebraically reparameterize each interior bridge state into AF3's noisy-target form x_GT + σ·ε, and map bridge time to an AF3-compatible noise level σ_t^align = c_t / b_t. They show this mapping preserves the bridge's signal-to-noise ratio relative to the target endpoint (SNR ∝ b_t²/c_t²), so pretrained co-folding weights remain usable.
3. A bridge-specific training objective and sampler. A bridge-time reconstruction weight w_bridge(t) = 1 / c_out(t)² is applied to the aligned denoising loss, and a Diffusion Bridge Implicit Models (DBIM) sampler is adapted to the apo-to-holo VE bridge, initialized from a near-apo interior time rather than the singular endpoint where b_σmax = 0.
4. Benchmark evidence plus ablations isolating where the gains come from. Comparisons against the same Protenix-Mini backbone with (i) no apo use, (ii) apo coordinates as auxiliary conditioning only (AF3-SFT-Apo), and (iii) a bridge without AF3 alignment, across 20-, 5- and 1-step budgets.

Main Findings

• Consistent gains on PoseBusters v2 at the default 20 steps: BridgeDock reaches lDDT 90.41% vs 84.79% for the AF3 baseline, %<2 Å 73.21 vs 68.61, %<5 Å 89.47 vs 84.69, and mean ligand RMSD 2.26 Å vs 2.64 Å; median RMSD is unchanged at 1.09 Å for both.
• Few-step generation is a major advantage: with only 1 denoising step BridgeDock obtains lDDT 90.77%, %<2 Å 69.86, %<5 Å 88.04, mean RMSD 2.19 Å and median 1.08 Å, versus the ODE-sampled AF3 baseline's 83.91%, 69.76, 86.51, 2.33 Å and 1.19 Å. A single BridgeDock step exceeds both the 5-step and 20-step AF3 baselines in lDDT and in both RMSD threshold percentages. At 5 steps BridgeDock gives lDDT 91.23%, 71.77, 88.52, 2.33 Å (mean) and 1.04 Å (median), versus 84.26%, 68.23, 84.40, 2.72 Å and 1.14 Å for the ODE baseline.
• Outperforms prior docking methods on the PDBBind v2020 time-based split: BridgeDock reaches lDDT 92.61%, %<2 Å 71.00, %<5 Å 89.40, mean RMSD 2.55 Å, median 1.09 Å, versus the Protenix baseline's 88.08%, 69.00, 86.30, 2.81 Å and 0.99 Å. Literature-reported numbers for other methods are much lower on the threshold metrics (e.g. DiffDock-Pocket 41.70 / 74.90 / 2.60 Å median; Re-Dock 42.90 / 76.40 / 2.40 Å; FlexDock 39.70 / – / 2.50 Å; GNINA 6.60 / 28.00 / 7.20 Å; SMINA 3.60 / 20.50 / 7.30 Å). Note that median RMSD for BridgeDock is slightly higher than the Protenix baseline (1.09 vs 0.99 Å).
• Gains come mainly from harder cases, not easy ones: the authors interpret the flat or slightly worse median RMSD alongside improved mean RMSD and threshold percentages as evidence that the method reduces the tail of difficult failures rather than polishing already-easy poses.
• Apo conditioning alone helps, but the bridge helps more: AF3-SFT-Apo (a zero-initialized fully connected layer encoding apo structure as extra conditioning, keeping the noise-to-structure path) reaches lDDT 87.53%, %<2 Å 70.14, %<5 Å 86.12, mean 2.48 Å, median 1.01 Å — better than the AF3 baseline but well below BridgeDock (90.41, 73.21, 89.47, 2.26 Å, 1.09 Å).
• Alignment is essential: the unaligned bridge variant gives only 85.23% lDDT, 68.42 %<2 Å, 86.60 %<5 Å, 2.77 Å mean and 1.22 Å median — small and unstable gains. The large gap to the aligned version indicates the benefit comes from combining the bridge with preservation of the pretrained co-folding prior, not from the bridge objective in isolation.
• The model is not given pocket information at test time: apo protein conformations are predicted from amino-acid sequence, ligand conformations come from RDKit, and the ligand is initialized at the geometric center of the apo protein rather than an oracle pocket location. The three literature flexible-docking baselines, by contrast, use ligand-derived pocket information under their respective known-pocket protocols.

Methodology in Plain English

1. Start from a pretrained co-folding backbone. The authors use Protenix-Mini, the lightweight open-source AF3-style implementation, and continue training it on 40,000 protein–ligand complexes curated from Protenix's original training data for 80,000 steps on 8× A100 GPUs, which takes approximately five days.
2. Replace "noise to structure" with "unbound to bound." Instead of corrupting only the ground-truth holo coordinates with Gaussian noise, the forward process mixes the apo structure, the holo target and noise with time-dependent coefficients, so at the start of the bridge the state sits at the apo structure and near the end it approaches the holo complex.
3. Translate bridge states back into the language the pretrained network speaks. Subtracting the apo contribution and dividing by the holo coefficient turns each bridge state into an ordinary "clean target plus scaled noise" state, and the corresponding noise level is read off as the noise-to-target ratio. Because the signal-to-noise ratio is preserved, the pretrained denoiser sees inputs that look like what it was trained on.
4. Keep the apo structure as an explicit condition. In addition to defining the bridge endpoint, the apo coordinates are passed through a separate apo encoder as conditioning features at every training and inference step, so removing the apo term from the coordinate input does not remove apo information.
5. Train with a bridge-weighted reconstruction loss (the AF3 aligned reconstruction objective reweighted by the bridge-time factor), retaining AF3's auxiliary terms such as smooth-LDDT and optional bond-geometry constraints.
6. Sample efficiently. Inference uses a DBIM-style implicit sampler over a decreasing uniform time grid starting slightly inside the singular endpoint (σ_max − 10⁻³), with a bootstrap denoising pass from the centered apo structure to initialize the bridge state.
7. Evaluate on PoseBusters v2 and a PDBBind v2020 time-based split using five metrics: lDDT, percentage of predictions under 2 Å and 5 Å ligand RMSD, and mean and median ligand RMSD.

Why This Matters

Impact on research. The paper shows a general recipe for transferring large noise-conditioned diffusion models to data-conditioned scientific prediction problems — here, turning a sequence-conditioned co-folding generator into a structure-conditioned docking generator without discarding pretrained weights. It is described as the first attempt to combine diffusion bridge theory with co-folding priors for docking, and the alignment trick (reparameterizing bridge states and matching SNR to the pretrained schedule) is architecture-agnostic in principle.

Real-world applications:

• Structure-based drug discovery: fast flexible docking of candidate ligands into unbound (apo) receptor structures, which are far more common experimentally than holo structures.
• Virtual screening at scale: few-step and even one-step generation makes large library screening cheaper at inference time.
• Refinement of predicted structures: using experimentally or predictively obtained unbound conformations as a starting point rather than re-predicting complexes from sequence.
• Beyond protein–ligand systems: the authors position bridge-based generation as applicable to general structure prediction and other data-to-data generative tasks.

Industry relevance. Pharmaceutical and biotech computational chemistry groups already run docking pipelines (e.g. SMINA, GNINA) and are adopting co-folding models; a method that preserves co-folding accuracy while consuming the apo structures companies actually have, and that works at low step counts, directly addresses cost and throughput constraints in industrial screening.

Future Directions

• Extending beyond Protenix-Mini. The backbone is a lightweight AF3-style variant trained on 40,000 complexes for 80,000 steps; whether the gains hold for full-scale co-folding backbones and larger curated datasets is not reported.
• Understanding the median-RMSD trade-off. BridgeDock improves mean RMSD and threshold percentages but has slightly higher median RMSD than the Protenix baseline on PDBBind (1.09 vs 0.99 Å) and than AF3-SFT-Apo on PoseBusters (1.09 vs 1.01 Å); how to capture the tail improvements without any regression on typical cases is an open question.
• Broader conformational-change regimes. The paper focuses on apo-to-holo rearrangement in flexible docking; extending the bridge to larger backbone-level flexibility and to other complex types (the monomer, protein–protein and protein–nucleic acid cases that co-folding models handle) is a natural next step.
• Generalizing the alignment mechanism. Applying the same bridge-state reparameterization and SNR-matching to other pretrained denoising backbones and to non-structural data-to-data scientific generation tasks remains to be tested.

Target Audience

Researchers and practitioners in machine learning for structural biology and computational drug discovery — especially those working on diffusion/flow models, co-folding architectures, or docking — who want to reuse large pretrained denoisers for structure-conditioned prediction. It also suits computational chemists evaluating whether co-folding accuracy can be brought into conventional apo-structure docking workflows, and ML methodologists interested in diffusion bridges and pretrained-prior alignment. Readers need a working knowledge of diffusion model preconditioning and protein–ligand structural representations; beginners will find the bridge algebra and sampler details demanding.

Authors’ abstract

Molecular docking aims to predict the three-dimensional structure of a protein-ligand complex; in the conventional structure-based setting, it is conditioned on an input protein structure and ligand, often starting from an apo receptor in flexible docking or a holo receptor in rigid docking. Recent co-folding models exemplified by AlphaFold 3 achieve strong accuracy in protein-ligand complex prediction, but they typically generate complexes from sequence-derived inputs and therefore do not directly use apo structures as structural priors in docking, leading to a larger conformational search space and weaker robustness in difficult or few-step settings. We present BridgeDock, a framework that adapts pretrained co-folding backbones to flexible docking by modeling the apo-to-holo transition as a diffusion bridge, allowing generation from the initial protein-ligand state rather than from Gaussian noise. To make this bridge formulation compatible with pretrained denoisers, we introduce an alignment mechanism that maps bridge states to the original denoising schedule of the pretrained model. Experiments on standard flexible docking benchmarks show that BridgeDock consistently outperforms strong co-folding baselines. Moreover, BridgeDock is computationally efficient at inference time and remains effective even with very few denoising steps.