MagicDock

MagicDock is a framework for docking-oriented de novo ligand design: instead of scoring or ranking pre-enumerated molecules, it generates new ligands—both protein binders and small molecules—that are optimized to dock into a target protein. It targets three limitations the authors identify in prior generative approaches: reliance on "pseudo" de novo pipelines that recombine existing fragments, insufficient explicit modeling of the docking interaction, and inflexible handling of different ligand types.

The framework couples differentiable surface modeling with gradient inversion. Proteins and ligands are represented as solvent-accessible surface point clouds, and generation is performed by treating a pretrained, differentiable docking model as an objective and descending gradients through it to sculpt a candidate ligand's coordinates and features. Because the docking knowledge is baked into the model's fixed weights, the same checkpoint supports design across multiple binding scenarios without per-target retraining.

MagicDock was developed by researchers at the Beijing Institute of Technology and released as a preprint in October 2025, with an associated OpenReview submission.

#Key Features

• Gradient-inversion generation: Ligands are produced by optimizing point-cloud coordinates and features through a differentiable, pretrained docking model, so generation is directly steered by learned docking signal rather than a separate scorer.
• Differentiable surface modeling: Both receptors and ligands are encoded as learnable 3D point clouds derived from solvent-accessible surfaces, giving a shared geometric representation for proteins and small molecules.
• VQ-MAE self-supervised pretraining: A vector-quantized masked autoencoder is pretrained on protein and molecule surface data, learning general representations before any docking supervision is applied.
• Unified across ligand types: Customized design procedures let the same framework generate protein ligands and small-molecule ligands within one gradient-inversion approach.
• Fixed-checkpoint design: After pretraining and fine-tuning, generation is an inference-time optimization, so new binders can be designed without retraining the model.

#Technical Details

MagicDock is built as a four-stage pipeline. Stage one performs docking-oriented ligand modeling using solvent-accessible surface point clouds. Stage two pretrains a VQ-MAE that combines masked autoencoding with vector quantization on protein and molecule datasets. Stage three fine-tunes on three supervised docking-related tasks—pocket prediction, interaction prediction, and binding-affinity regression—using datasets including SKEMPI v2, SAbDab, PDBbind 2020, and CrossDocked2020. Stage four generates ligands by inversion, optimizing the ligand point cloud against the differentiable docking objective. Across nine design scenarios, MagicDock reports average improvements of 27.1% over protein-ligand-design baselines and 11.7% over small-molecule-design baselines. Code and supporting materials are provided in the paper's supplementary materials.

#Applications

MagicDock is intended for structure-based drug discovery and protein-binder design, where the goal is to invent new molecules tailored to a specific pocket rather than screen an existing library. Because it handles both antibody/protein binders and small molecules through one surface-based formulation, it can serve teams working on biologics and on small-molecule leads. The gradient-inversion setup lets designers repeatedly probe a target, generating candidates whose predicted docking behavior is optimized before any wet-lab synthesis.

#Impact

MagicDock contributes a unified, docking-centered take on de novo ligand design, bringing gradient-inversion generation—optimizing inputs through a frozen predictive model—into a setting usually dominated by autoregressive or diffusion generators. Reported gains over specialized protein- and molecule-design baselines across nine scenarios suggest the surface-based, docking-in-the-loop formulation is competitive. As a preprint whose evaluation is computational, its real-world value awaits experimental confirmation that generated binders bind and function as predicted.