AMDiff

Structure-based drug design asks a generative model to invent a molecule that fits a specific protein pocket in three dimensions. Two families of methods dominate, and each fails in a characteristic way. Atom-based generators place one atom at a time and can reach any corner of chemical space, but with no explicit geometric constraint they emit bond lengths and angles no chemist would draw — strained fused cyclopropanes, seven-membered aromatic rings, rings that never close. Motif-based generators assemble molecules from a curated fragment vocabulary and so stay chemically sane, but they cannot express any substructure the vocabulary omits, and they routinely botch the linkers between fragments.

AMDiff — the Atom-Motif Consistency Diffusion Model — declines to choose. It runs one hierarchical diffusion process over both views of the same molecule simultaneously. The atom view predicts atom types and coordinates; the motif view constructs a motif tree of functional groups and larger fragments; an interaction network passes information between them at every denoising step. Motifs hand the atom view clustering information about which atoms belong together, and atoms hand the motif view precise positioning information. Ligand-protein contacts are modelled by an equivariant graph neural network so that generated conformations respect the geometry of the pocket, and binding-site features enter as conditional inputs under classifier-free guidance.

The model comes from the Hong Kong University of Science and Technology (Guangzhou), with collaborators at the Hong Kong University of Science and Technology and Shenzhen Bay Laboratory. It belongs to the same pocket-conditioned generative family as MODA and PocketXMol, and is distinguished within it by treating atom-level and motif-level representations as two views to be reconciled rather than as competing design units.

#Key Features

• Atom-motif consistency: A joint-training paradigm reconciles a fine-grained atomic view with a coarse motif view of the same ligand, so generated molecules keep fragment-level coherence without inheriting a fragment library's blind spots.
• Cross-view interaction network: Bidirectional message passing sends clustering signals down from motifs to atoms and positional detail up from atoms to motifs at each step of the reverse diffusion chain.
• Pocket conditioning with classifier-free guidance: Binding-site features act as the conditioning signal, letting the same checkpoint be steered toward different targets at inference time without retraining.
• Topological fingerprints: Persistence diagrams encoded as topological fingerprints capture multi-scale shape features of the growing ligand, complementing the geometric message passing.
• Robustness to pocket perturbation: Generation was probed against site-directed mutants and against binding pockets defined at radii from 4 to 30 angstroms.

#Technical Details

AMDiff was trained on CrossDocked2020, a collection of 22.5 million docked protein-ligand complexes, filtered to remove entries RDKit cannot parse, poses with RMSD above 2 Å, and complexes below 40% sequence identity — yielding 100,000 training complexes with a further 100 proteins held out for testing. Training ran for 600,000 steps at batch size 4 with Adam at an initial learning rate of 5e-4 under exponential decay, taking roughly 16 hours on an NVIDIA RTX 4090.

Evaluation sampled 100 molecules for each of the 100 held-out targets, 10,000 molecules in total, against liGAN, AR, Pocket2Mol, GraphBP, DecompDiff, and the motif-based FLAG. AMDiff reported 98.9% validity (FLAG, the next best, 98.1%), diversity 0.672, novelty 0.663, QED 0.479, and a synthetic accessibility score of 0.684 — the highest among the generative baselines, though still short of the 0.727 measured on the reference test-set ligands. Mean AutoDock Vina affinity was -7.466 kcal/mol, against -7.502 kcal/mol for the reference ligands. Bond-angle and dihedral-angle KL divergences against reference molecules quantify how closely the generated geometries track real chemistry.

#Applications

The intended use is lead identification and early hit-to-lead work on a target whose structure is known. Case studies designed candidate inhibitors for two kinases of clinical importance: anaplastic lymphoma kinase (ALK, PDB 3LCS), a driver in non-small cell lung cancer and anaplastic large cell lymphoma, and cyclin-dependent kinase 4 (CDK4, PDB 7SJ3). Because the pocket enters only as a conditioning input, medicinal chemists can point the trained checkpoint at a new target, a resistance mutant such as ALK G1202R or S1206Y, or a differently sized pocket definition, and sample candidates without collecting target-specific training data.

#Impact

AMDiff addresses a real dichotomy in the structure-based generative literature, where atom-level and motif-level methods have largely developed as separate lineages with complementary failure modes. Code is released under the MIT license with trained weights distributed through Google Drive, and the work was peer-reviewed and published in Chemical Science after preprint release. The write-up itself is the least reusable part of the release: the version of record is CC BY-NC 3.0 and the arXiv posting carries only arXiv's perpetual non-exclusive distribution license, so neither permits the redistribution the code license does. The evaluation is entirely in silico: validity, drug-likeness, and docking scores are computational proxies, and no generated compound was synthesized or assayed, so the ALK and CDK4 results demonstrate that the model produces plausible, pocket-complementary chemotypes rather than confirmed binders.