MolCRAFT

A generative model for structure-based drug design is usually judged on the binding affinity of what it produces, but a molecule can score well and still be unusable: generated ligands routinely carry strained ring systems, or adopt poses that rearrange the moment docking software touches them, in which case the reported affinity belongs to the redocked pose rather than to the molecule the model drew. MolCRAFT's authors catalogue three such failure modes across existing SBDD generators — distorted geometry, inferior binding mode, and outright generation failure — and trace them to two causes. Autoregressive samplers collapse onto a narrow set of substructures under the artificial atom ordering generation imposes. Diffusion samplers avoid that but denoise in a hybrid space, discrete atom types alongside continuous coordinates, where high-variance intermediate latents drift off the molecular manifold.

MolCRAFT moves generation entirely into continuous parameter space with a Bayesian flow network. Rather than noising and denoising the molecule itself, the model carries a distribution over the clean molecule — Gaussian means and precisions for atom coordinates, categorical probabilities for atom types — and updates those parameters by Bayesian inference as an SE(3)-equivariant network sharpens its reconstruction. Both modalities then live in the same continuous, differentiable space, with no discrete-continuous seam to bridge and far lower input variance than a hybrid diffusion path. It was built at the Institute for AI Industry Research at Tsinghua University with the University of Illinois Urbana-Champaign and the Shanghai Institute of Materia Medica, and presented at ICML 2024.

It is the first project in the MolCRAFT series and the backbone its siblings build on: MolJO steers this generator with property gradients, and MolPilot reworks its sampling schedule across the two modalities.

#Key Features

• One parameter space for two modalities: Continuous coordinates and discrete element types are both represented as distribution parameters, so a single noise process and a single closed-form Bayesian update cover the whole molecule.
• Noise-reduced sampling: The sampler updates parameters directly from the network's estimate of the clean molecule instead of first drawing a noisy sample from it, removing a variance source that leaves diffusion models completing molecules only at the end.
• Generation for an arbitrary pocket: sample_for_pocket.py takes a PDB file and a reference ligand, clips the pocket to a 10 Å region around it, and samples from the released checkpoint with no retraining.
• Poses that hold up after redocking: Generated conformations are evaluated for strain energy, protein-ligand steric clashes, and RMSD to the redocked pose, not affinity alone.
• Size-controlled comparison: Because Vina Dock rewards larger molecules, results are reported for small and large variants matched to baseline molecule sizes.

#Technical Details

The network is an SE(3)-equivariant architecture conditioned on the protein pocket, trained on CrossDocked2020 under the standard preparation for this benchmark: RMSD-based filtering and a 30% sequence identity split leaving 100,000 training complexes and 100 held-out test proteins, with 100 molecules sampled per test pocket at evaluation. On that test set MolCRAFT reports an average Vina Score of −6.59 kcal/mol against −6.36 kcal/mol for the reference ligands, with Vina Min −7.27 and Vina Dock −7.92 kcal/mol, and a margin of −0.84 kcal/mol over the strongest baseline the paper evaluates at comparable molecular size. Conformational quality follows: the lowest average Jensen-Shannon divergence to reference bond-length (0.319) and torsion (0.300) distributions among the compared methods, and 41.8% of generated poses within 2 Å RMSD of their redocked counterpart. Sampling runs in 100 steps rather than the 1,000 training steps, producing 100 molecules for a pocket in 141 seconds — roughly a 30× speedup over the diffusion baselines timed in the same study — at 96.7% generation success. Ablating the sampling strategy alone moves the Vina Score from −5.42 to −6.51 kcal/mol.

#Applications

The practical use is pocket-conditioned hit generation: point the released checkpoint at a crystal or predicted structure with a known ligand site and it proposes 3D molecules placed in there, with poses stable enough to inspect directly rather than as a sketch for docking to fix. Affinity, drug-likeness, synthetic accessibility, strain energy, and clash counts are all computed over the same samples, so the output arrives as a triaged candidate list. Running it needs the repository's Docker environment and a GPU, placing it with computational chemists rather than at the bench.

#Impact

MolCRAFT carried Bayesian flow networks from unconditional 3D molecule generation into pocket-conditioned design, and pressed the argument that conformational feasibility has to be measured alongside affinity or the reported affinities are false positives. It is now a standard comparison in later SBDD work, including TransDiffSBDD, MolPIF, CByG, READ and SE3-BiLingoMol. The limits are real: it inherits CrossDocked2020's docked-pose training distribution, its results are in-silico Vina and PoseCheck metrics with no experimental validation, and the released weights carry a non-commercial license.