PDBPocketDiffusionMolecules
A 3D-aware pipeline
from structure to drug candidates.
Step 1
Input
PDB fetch / upload
RCSBPDB parser
Step 2
Detection
Find binding cavities
P2Rank 2.5
Step 3
Selection
Pocket / ligand / atom
3D viewermanual
Step 4
Generation
100-step diffusion
PocketXMol
Step 5
Filter & rank
Drug-likeness rerank
RDKitconstraints
Step 1
Input
PDB fetch / upload
RCSBPDB parser
Step 2
Detection
Find binding cavities
P2Rank 2.5
Step 3
Selection
Pocket / ligand / atom
3D viewermanual
Step 4
Generation
100-step diffusion
PocketXMol
Step 5
Filter & rank
Drug-likeness rerank
RDKitconstraints
P2Rank
ML pocket detectorScores every solvent-accessible surface point on the protein for how ligand-friendly it looks, then clusters high-scoring points into pockets.
PocketXMol
Cell 2026 · foundation model3D diffusion model. Starts from noise inside the pocket and iteratively refines atom positions and types over 100 steps until valid molecules emerge — pocket geometry stays fixed.
Inside a generation run
100 stepst = 100
pure noise
Random atom cloud in pocket
t = 66
rough shape
Atoms drift toward geometry
t = 33
chemistry emerges
Elements + bonds crystallize
t = 0
valid molecule
Sanitized → SMILES + SDF