MODA

Structure-based molecular design is usually carried out one task at a time. A medicinal chemist growing a fragment, closing a linker between two anchors, swapping a scaffold, or decorating side chains typically reaches for four different generative models, each trained on its own curated dataset with its own loss and its own heuristics. MODA — "Mask Once, Design All" — collapses those four editing problems into a single objective: mask a contiguous spatial fragment of a ligand sitting in its protein pocket, then regenerate the missing atoms by diffusion denoising.

Because each editing task corresponds to a different rule for choosing which atoms to mask, one masked-reconstruction objective covers all of them, and a Bayesian mask scheduler samples a task-specific masking rule for every molecule at every training iteration. The result is a single multi-task backbone trained in one stage, with no pretrain-then-finetune step and no task labels, control tokens, or auxiliary classifiers at inference.

MODA was posted as an arXiv preprint in July 2025 by a group led from Shenzhen University, with collaborators at the University of Nottingham Ningbo China and UC San Diego. It sits alongside pocket-conditioned 3D generators such as Apo2Mol and FLOWR.root, but its central claim is about breadth rather than any single task: that multi-task training buys transfer to unseen targets and design settings.

#Key Features

• Four editing tasks in one backbone: Fragment growing, linker design, scaffold hopping, and side-chain decoration are all expressed as masked substructure denoising, so a single set of weights serves what would normally require four specialist models.
• Bayesian mask scheduler: A sampling policy picks among chemically motivated masking rules — the exit atom plus its two-bond neighborhood for fragment growing, the shortest path between anchors for linkers, the Bemis-Murcko core for scaffolds, and non-ring terminal groups for side chains.
• Native 3D, SE(3)-equivariant representation: Atom types and Cartesian coordinates are fed directly to the backbone rather than SMILES strings, preserving stereochemistry and non-covalent contacts with the pocket.
• Single-stage training: The unified masking-denoising objective removes the pretrain-finetune gap; a pretrained variant (Model-C-PT) scores only marginally better across the reported metrics.
• Zero-shot transfer: The multi-task model handles de novo design and lead optimization on held-out protein-ligand pairs without per-target refitting or force-field refinement.

#Technical Details

MODA is a joint continuous-discrete diffusion model: Gaussian noise is applied to the coordinates of masked ligand atoms and categorical noise to their atom types, with the protein pocket supplied as conditioning context. Pocket atoms are themselves perturbed with small Gaussian noise as a regularizer that discourages overfitting to rigid binding-site conformations. At inference, lead-optimization runs initialize the ligand at the center of mass of the reference substructure and de novo runs at the pocket center, preserving SE(3) invariance. Training data is built from the CrossDocked protein-ligand set, with each complex annotated for one or more tasks; per-task splits run from roughly 88,000 to 114,000 masked examples, and the fragment, linker, and scaffold sets overlap heavily, which is what lets shared geometric priors form.

Three paradigms are compared: single-task training and evaluation (Model-A), multi-task training with single-task evaluation (Model-B), and fully multi-task (Model-C). On the four-subtask de novo test set, Model-C reaches Vina Dock scores of −8.83 kcal/mol on linker design and −8.45 on fragment growing with unique-correctness above 93 percent, against six diffusion baselines (TargetDiff, DiffSBDD, DiffBP, FLAG, D3FG, DecompDiff). On the zero-shot de novo set it holds a negative raw Vina Score of −2.05 kcal/mol where several baselines go positive, with a 30.1 percent improvement rate, QED 0.55, SA 0.62, and mean Lipinski compliance of 4.72.

#Applications

The intended user is a medicinal chemistry team working from a known pocket. Rather than maintaining separate models for hit elaboration, PROTAC and bivalent linker design, scaffold hopping around an intellectual-property or liability constraint, and side-chain tuning, a group can run all four edits through one checkpoint and get 3D poses generated in the context of the target. Because the model transfers to unseen targets without retraining, it fits early hit-to-lead triage, where a new pocket appears faster than a task-specific dataset can be assembled.

#Impact

MODA's contribution is a demonstration that a single-stage multi-task diffusion routine can substitute for the pretrain-finetune pipelines that dominate structure-based generation, and that the shared backbone gains rather than loses accuracy relative to specialists on most tasks. The evidence has clear boundaries: the work is a preprint that has not been peer-reviewed, no code or model weights have been released, and every result is in silico, resting on docking scores and cheminformatics filters that correlate imperfectly with measured binding. Single-task training also retains an edge on scaffold hopping, so the multi-task backbone is not uniformly better. The masking formulation nonetheless states compactly something the field has circled for a while: the canonical ligand-editing tasks are one operation applied at different scales.