Metalloprotein docking requires a ligand pose to satisfy both global pocket fit and the identity and geometry of metal-coordinating atoms. General-purpose diffusion docking models do not explicitly condition generation on externally specified complete ligand donor sets, and their predicted poses can exhibit incorrect metal-donor relationships. Metal-aware models address coordination during training and generation; here we ask whether external donor information can correct already-trained general-purpose models without retraining. We propose Complete-Identity Contract Guidance (CICG), which applies a supplied complete donor set through translation, rotation, and torsion updates, and Historical Contract Transfer (HCT), which supplies donor identities from earlier complexes at compatible receptor sites. We compare HCT with trajectory-local controls retaining the historical donor count with or without elemental composition. On 86 assessable HCT-covered queries, HCT-guided CICG increases frozen-confidence-selected joint success from 21.5% to 36.6% on SurfDock and from 25.0% to 34.9% on DiffDock-L. The joint criterion requires correct global ligand placement, complete native donor recovery, coordination geometry, preserved stereochemistry, and absence of fatal receptor clashes. Composition-matched guidance reaches 33.14% and 34.30%, leaving the additional benefit of fixed historical membership unresolved. Our results therefore support donor-set guidance as a conditional correction for frozen docking models, while identifying donor-source reliability and physical-quality preservation as the main barriers to broader use.
Figures & tables
Figure 1: Supplied-donor correction of a frozen docking model. (a) CICG applies donor-conditioned corrections CD after late sampling updates of the frozen generator Gθ . (b) The query-ligand graph shows alternative donor sets supplied by HCT (green) and trajectory-local nearest selection (orange). Both enter the same correction operator; the ribbon view shows a historical reference complex. (c) Frozen-confidence-selected SurfDock poses illustrate the metal–donor geometry. Gray surfaces show the receptor pocket; blue and green ligand carbons distinguish no guidance and HCT + CICG. Dashed connectors show Mn–donor distances; values are ordered Mn–O9 / Mn–O11. Structural details are given in Appendix F.1 .
Figure 2: Structural outcomes and coordinate-bank utility. (a) Frozen-confidence-selected SurfDock poses for a recovery (7YBA) and a regression (9FVV). Ligand carbon atoms are blue for no guidance and green for HCT + CICG; oxygen, nitrogen, and the metal are red, blue, and purple. The protein is shown as a gray cutaway surface, and dashed lines connect the transferred donor atoms to the metal. Each pair uses the same camera, and labels report ligand RMSD and joint success. (b) Selected joint-success counts from the no-guidance and HCT-guided banks under frozen-confidence and post-hoc donor-aware selection on the same 86 queries and two seeds. With donor-aware selection held fixed, success increases from 51 to 64 of 172 observations (+7.56 percentage points; paired 95% group-bootstrap CI [1.11, 14.04]), comprising 21 recoveries and eight regressions. Banks containing a successful candidate increase from 91 to 96, while selection misses decrease from 40 to 32.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3: Conditional paired effects and query-weighted intervention benefit in SurfDock. The left plot shows conditional changes in joint success for HCT guidance against no guidance; the right plot shows their query-weighted contributions and the overall contrast Δ=pg+(1−p)ℓ . Both plots use the same percentage-point scale and frozen-confidence selection. Historical donor membership is classified retrospectively using native donor labels. Exact membership comprises 69 queries (138 query–seed pairs), with a net gain of 30 joint successes ( g=+21.74 percentage points); inexact membership comprises 17 queries (34 pairs), with a net loss of 4 ( ℓ=−11.76 percentage points). The weights p=69/86 and 1−p=17/86 give contributions of +17.44 and −2.33 percentage points, summing to Δ=+15.12 percentage points ( 26/172 ). Native labels define evaluation strata only. These descriptive means have different stratum denominators; no confidence intervals are drawn. The fixed donor-aware comparison appears in Figure 2 (b).
Figure 4: Contract membership and observed metal–donor geometry. Each selected output is overlaid on the native ligand (thin gray), used only for evaluation. (a) Q023, 7YBA IO9 A802/Mn A801: no-guidance/HCT samples 36/32; native and historical donors O9/O11. (b) Q002, 9FVV AKG A801/Mn A805: samples 19/10; native donors O3/O5, historical donors O2/O5. All outputs are independently frozen-confidence-selected at seed 43, from distinct trajectories; the original size/ID case-selection rule above is retained. Atom labels are one-based input-graph indices, mapped to zero-based arrays in the manifest. Right-hand plots measure Mn distances to the historical-contract atoms in native, no-guidance and HCT coordinates on one 0–6.5 Å scale. The native O2 point in (b) evaluates the historical atom’s reference position; O2 is not a native donor. The dashed 2.15 Å line is the fixed Mn–O target; structural connectors identify measured distances rather than inferred bonds. The recovery reaches approximately 2.15 Å, while the regression retains 5.99/4.29 Å: assigned membership need not be realized. Labels reproduce stored ligand RMSD and joint-success flags. Recorded fatal-clash flags are zero for all four outputs. Native geometry uses open gray circles, no guidance brown circles, and HCT teal diamonds. Receptor context and camera/crop are fixed within each case; coordinates are unchanged, without alignment or minimization. These examples use frozen-confidence selection; the fixed donor-aware comparison is shown in Figure 2 (b). Exact distances, inputs, sample IDs, atom mapping and cameras are archived with the editable figure.
Diffusion models have become a dominant approach to structure-based drug design. However, because they are trained primarily to fit reference structures, they often place ligands in physically implausible poses that clash with the protein. We introduce MLIPDiff, an inference-time guidance method that can be applied to pretrained diffusion models for ligand generation. The method uses physics-informed guidance from UMA (Universal Models for Atoms), a machine-learned interatomic potential (MLIP). The guidance combines four weighted energy terms (protein-ligand interaction, ligand strain, steric clash, and pocket anchoring) and uses the original model checkpoint without retraining. On CrossDocked2020, MLIPDiff completely eliminates protein-ligand collisions while preserving the ligand's local geometry, improving bond-length distributions and physical plausibility relative to the base model. Compared to models such as NucleusDiff that are specifically designed to reduce atomic clashes, MLIPDiff achieves clash-free structures without compromising local geometry. It maintains the base model's PoseBusters pass rate and attains the best median Vina Score, Vina Min, and Vina Dock, surpassing both the base model and NucleusDiff, with a median of 59.6% of ligands per pocket binding more strongly than the reference. MLIPDiff thus addresses a limitation of prior methods by generating physically plausible structures with favorable predicted binding affinity using a single frozen model and no additional training.
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.
Shikun Feng, Zipeng Yang, Qiuyi Li +4
Zhongguancun Academy, Beijing, China. · Zhongguancun Institute of Artificial Intelligence, Beijing, China. · School of Physical Science and Technology, Beijing University of Posts and Telecommunications, Beijing, China. +2
Flexible docking is a highly practical task, aiming to predict the joint configuration of a protein-ligand complex from a protein apo structure and a ligand SMILES string. Diffusion and flow-based generative models have made this task tractable by modeling the posterior distribution over the ligand pose and the protein conformation conditioned on the apo input. The central challenge, however, remains the long-tailed apo-to-holo structure shift, where a small fraction of protein undergo large conformational changes. In this work, motivated by the practical observation that downweighting large shift examples improves generalization, and by the mixed nature of the long tail consisting of genuinely difficult examples and noisy ones, we propose to use curriculum learning for this task, gradually incorporating large conformational change examples during training. With a linear pacing function, we show that an existing method can be strongly enhanced. On the PDBBind ESMFold apo benchmark, with relaxation and confidence-based pose selection, the resulting model GCLDock achieves a 46.52% top-1 success rate (ligand RMSD less than 2A) and 40.26% when additionally requiring PoseBusters validity, outperforming existing baselines.
Wenyin Zhou, Hossein Azizpour
KTH Royal Institute of Technology · Science for Life Laboratory