TransDiffSBDD

Structure-based drug design (SBDD) asks a generative model to invent a molecule that fits a particular protein pocket. The output is awkward for any single architecture, because a ligand is two things at once: a discrete graph of atoms and bonds, and a continuous set of 3D coordinates describing how that graph sits in the site. Diffusion models handle the coordinates well but tend to blur the discrete chemistry; autoregressive language models handle the chemistry well but must quantize coordinates into tokens, discarding the continuity that binding geometry depends on.

TransDiffSBDD, released as a preprint in March 2025 by researchers at Tsinghua University, Microsoft Research AI for Science, McGill University, and Mila, assigns each modality to the architecture suited to it. A GPT-style causal transformer reads a hybrid-modal sequence and emits the ligand's SMILES string; its output vector then conditions a diffusion MLP that samples the ligand's atomic coordinates.

The second idea is ordering. The hybrid-modal sequence lists the pocket's atoms as token-coordinate pairs, then the ligand's SMILES tokens, and only then the ligand's coordinates. That layout encodes a causal claim: once a ligand's graph is fixed and the pocket is known, the binding pose is largely determined. Generation proceeds graph-first, pose-second, unlike pocket-conditioned generators such as PocketXMol that sample atom identities and positions together.

#Key Features

• Hybrid-modal sequence: Protein-ligand complexes are serialized into one stream alternating discrete tokens with continuous 3D vectors, so a single causal model sees the pocket, the ligand's chemistry, and its geometry without a separate encoder for each.
• Causal modality ordering: Placing all ligand coordinates after the SMILES tokens makes pose generation explicitly conditional on the molecular graph rather than simultaneous with it.
• Integrated transformer and diffusion: The token head produces SMILES autoregressively while a diffusion network on the same backbone produces coordinates, trained under a joint objective that sums cross-entropy loss with a weighted diffusion loss.
• Pocket-conditioned generation: Because the pocket's atoms and coordinates are part of the input sequence, the pretrained model generates candidates for a given pocket directly, with no retraining step required to see a new target.
• Optional per-target RL refinement: After pretraining, the transformer backbone can be treated as an RL agent — the diffusion MLP stays frozen — and updated with a regularized maximum-likelihood loss against a property reward such as Vina score, QED, or synthetic accessibility.

#Technical Details

The backbone follows a GPT-style design: a token embedding block for discrete symbols, a linear layer projecting 3D coordinates into the same space, positional embeddings, causal transformer layers, and two output heads — an MLP token head and a denoising diffusion MLP. Pretraining draws on roughly 209M 3D small molecules from the Uni-Mol dataset alongside 100,000 protein pocket-ligand complexes from the CrossDocked2020 training split, augmented by SMILES randomization and by centering each complex on the ligand's center of mass before random 3D rotation.

Evaluation uses the standard CrossDocked2020 protocol: 100 held-out pockets with low sequence similarity to training, 100 generated molecules each, scored with AutoDock Vina. The reported results, which include per-target RL fine-tuning, give a Vina Dock of −9.37 kcal/mol, SA of 0.75, diversity of 0.81, QED of 0.48, a Vina Score of −6.02 kcal/mol, and a success rate of 83.9% — the fraction of molecules reaching Vina Dock below −8.18 kcal/mol, QED above 0.25, and SA above 0.59. Baselines include Pocket2Mol (−7.15 Vina Dock, 24.4% success), TargetDiff (−7.80, 10.5%), DecompDiff (−8.39, 24.5%), MolCRAFT (−9.25, 36.1%), and Reinvent with Vina docking (−9.18, 76.7%); the reference ligands score −7.45 with a 25.0% success rate.

#Applications

The model targets early hit generation for a protein whose binding site is structurally characterized: given pocket coordinates, it proposes chemically valid molecules together with a docked pose, and the RL stage lets a project team weight affinity against drug-likeness and synthesizability as the campaign requires. Case studies on targets 1R1H and 4PXZ produce candidates with better docking scores than the reference ligands and distinct scaffolds — the behavior medicinal chemists want from a generator used for scaffold exploration rather than analogue enumeration.

#Impact

TransDiffSBDD's contribution is architectural: it shows that separating a molecule's discrete and continuous descriptions across two generative mechanisms, and ordering them so geometry follows chemistry, outperforms giving both to one mechanism. The success-rate margin is the headline result, and it lands in the metric closest to a practical multi-objective filter. The caveats are those of the SBDD literature generally — Vina, QED, and SA are computational proxies, and the evaluation is entirely in silico — plus two specific to this work: the reported numbers depend on a per-target RL stage that adds compute for every new pocket, and no code or weights have been released, so the results cannot yet be independently reproduced. The paper is a preprint awaiting peer review.