Protein-Ligand Docking

Protein-ligand docking estimates how a small molecule may bind to a protein binding site. A useful docking workflow separates pose generation, pose filtering, and affinity or ranking models.

Docking can be framed as searching over ligand pose variables:

$$X^{\text{*}}=\arg \underset{X}{\min }S(P,L,X)$$

Here $P$ is the protein or pocket, $L$ is the ligand, $X$ is the ligand pose/conformation in the binding site, and $S$ is a scoring function. In learned methods, $S$ may be replaced or complemented by a generative model.

More explicitly, the ligand pose can be written as:

$$q=(t,R,\tau ),X=X(q)$$

where $t$ is translation, $R$ is rotation, and $\tau$ is a vector of torsion angles. Classical docking searches over $q$; learned pose generators may sample $q$ or directly predict $X$.

What Docking Produces

Docking output is usually a set of candidate ligand poses in a binding site. A pose is not automatically a binding explanation. It has to be checked for geometry, chemistry, protein-ligand contacts, and whether the assumed binding site is meaningful.

Modeling Views

• Search problem: explore ligand translations, rotations, conformers, and sometimes side-chain flexibility.
• Pose generation problem: sample or optimize candidate geometries conditioned on protein and ligand context.
• Ranking problem: use a scoring function to order candidate poses.
• Generative problem: directly sample plausible complex structures with learned geometry.
• Evaluation problem: distinguish native-like poses from invalid or overfit predictions.
• Screening problem: rank a ligand library for virtual screening.

Working Questions

• How should pose plausibility be checked before using a scoring function?
• Which failures are geometric, chemical, or data-driven?
• How should docking outputs be evaluated against physical constraints and benchmarks such as PoseBusters?
• How much protein flexibility is needed for the target being modeled?
• Is the method using a known pocket, a predicted pocket, or blind docking?

Pipeline Sketch

1. Prepare receptor and ligand structures.
2. Generate candidate poses with an explicit pose generation policy.
3. Filter implausible poses using geometry and chemistry checks.
4. Rank candidates with a scoring function.
5. Record uncertainty and failure modes.

Component Equations

A useful decomposition is:

$$\{{\hat{X}}_k{\}}_{k=1}^K\sim G_{\theta }(P,L,c)$$ $$s_k=S_{\varphi }(P,L,{\hat{X}}_k)$$ $${\hat{X}}_{\mathrm{top}}=\underset{k}{\mathrm{select}}({\hat{X}}_k,s_k)$$

This separates generator $G_{\theta }$, scoring function $S_{\varphi }$, and selection rule. If a paper changes all three, the improvement cannot be assigned to only one component.

Failure Modes

• Plausible-looking pose with invalid ligand geometry.
• Low score caused by scoring shortcuts rather than real binding signal.
• Correct pocket but wrong ligand orientation.
• Good RMSD on an easy split but poor generalization to new proteins or scaffolds.
• Protein treated as rigid when the relevant binding mode requires flexibility.
• Score convention inversion between docking tool, learned scorer, and benchmark script.
• Filtering invalid poses before reporting metrics without counting filtered failures.

Practical Checks

• Validate ligand identity, stereochemistry, bond lengths, clashes, and pocket overlap.
• Report pose quality separately from binding affinity or enrichment metrics.
• Compare against classical docking baselines when possible.
• Track whether evaluation uses crystal structures, predicted structures, apo structures, or sequence-only input.
• Use PoseBusters-style plausibility checks before claiming state-of-the-art structure generation.

Related

• Pocket
• Structure-based drug discovery
• Pose generation
• Pose quality
• Binding affinity
• Virtual screening
• Equivariant GNN
• Flow matching
• Diffusion model
• Structure-based modeling
• Leakage
• LLM Wiki