Wiki · Bio × AI Methods · concept
Molecular docking
Predicting how a small molecule fits into a protein's pocket — the pose — and scoring how well it binds.
Molecular docking predicts how a small molecule settles into a protein’s binding pocket, and how tightly. Given a protein structure and a candidate drug-like molecule, docking answers two questions: what does the bound arrangement look like — the pose — and how good is it? The pose is a 3D placement of every atom; the score is a single number meant to stand in for binding strength. Get both right and you can tell, in silico, whether a compound is worth making.
How it works
Classical docking is a search plus a score. The search explores possible poses — sliding, rotating, and flexing the molecule inside the pocket to find arrangements that fit. The scoring function then rates each pose using a physics-flavored estimate of the interactions holding it in place: shape complementarity, hydrogen bonds, electrostatics, buried hydrophobic surface. Tools like AutoDock Vina and Gold are the classical workhorses (GNINA adds a CNN-based scorer on top of this search). A newer line uses deep learning end-to-end: DiffDock casts docking as a diffusion process, starting from a random placement and iteratively denoising it toward a plausible pose — the same generative idea behind a diffusion model for images, aimed at molecular geometry.
Why it matters (for bio × AI)
Docking is the engine of virtual screening — you can dock millions of compounds against a target and rank them before touching a lab bench, an enormous filter on an enormous chemical space. But the scoring function is the weak link: ranking correlates only loosely with real binding, so top-scoring hits routinely disappoint. Deep-learning docking is fast and sometimes places poses well, yet the PoseBusters benchmark showed a sharp catch — a pose can have low RMSD to the truth and still be physically invalid, with clashing atoms or impossible bond geometry, and generalization to unseen proteins lags. It’s the “read the supplement” caveat of the field: a confident number is not a correct structure.
Related
See AlphaFold for the predicted structures that increasingly serve as docking targets, diffusion model for the generative method behind tools like DiffDock, and protein design for the generative flip side — building a protein to bind a target rather than fitting a small molecule into one.