Virtual Screening

Virtual screening ranks a library of candidate molecules for a target. In structure-based workflows, ranking may use docking poses, scoring functions, learned affinity models, or filters.

Given a target pocket $P$ and a molecule library $\mathcal{L}$, a simple screening objective is:

$$\mathrm{rank}(L)=S(P,L,{\hat{X}}_L)$$

where ${\hat{X}}_L$ is the generated or selected pose for ligand $L$, and $S$ is a scoring function.

Enrichment at top $k$ can be written as:

$${\mathrm{EF}}_k=\frac{\text{actives in top }k/k}{\text{actives in library}/|\mathcal{L}|}$$

Screening Contract

Virtual screening is a ranking task over a specific candidate pool:

$$\mathcal{C}=\{(L_i,r_i,a_i){\}}_{i=1}^N$$

where $L_i$ is a ligand candidate, $r_i$ is its prepared representation, and $a_i$ is activity or relevance metadata when available.

For a reusable note, state:

Field	Question
target context	protein, pocket, receptor state, cofactors, waters, and assay context
library source	known actives, measured inactives, decoys, purchasable library, generated set, or public benchmark
ligand preparation	standardization, protonation, tautomer, stereochemistry, conformers
pose policy	docked pose, known pose, generated pose, ensemble, or no pose
scoring policy	docking score, learned score, consensus, filter, or reranker
denominator	all attempted compounds, successfully docked compounds, or post-filtered compounds

The denominator is part of the claim. Reporting only post-filtered molecules can overstate practical screening quality.

Ranking Pipeline

A structure-based screening workflow is often:

$$L\to \tilde{L}\to \{{\hat{X}}_{L,j}{\}}_{j=1}^J\to s(L)\to \mathrm{rank}(L)$$

Each arrow can fail. A paper should separate ligand preparation failure, pose generation failure, scoring failure, and final ranking quality.

Metrics

Use ranking metrics that match the screen:

Metric	Use	Main Caveat
EF@K or EF at top fraction	early active enrichment	sensitive to active prevalence and decoys
ROC-AUC	global ranking over positives/negatives	can look good while early enrichment is weak
BEDROC or early-recognition metric	top-heavy ranking	parameter choice affects interpretation
precision@K	fixed wet-lab or compute budget	depends on candidate pool and label completeness
hit rate after filters	practical pipeline yield	denominator must include failed candidates

For docking-based screens, pose plausibility and enrichment should be reported separately.

Checks

• Is the library filtered for chemistry, duplicates, and assay compatibility?
• Are decoys realistic, or do they make the task artificially easy?
• Is ranking driven by pose quality, affinity, ligand bias, or dataset artifacts?
• Are top-ranked molecules inspected for validity and novelty?
• Are negative sets and applicability domain reported?
• Is pose generation failure separated from scoring failure?
• Is early enrichment reported instead of only global classification metrics?
• Are failed, invalid, or filtered molecules included in sample accounting?
• Is the target split new protein, new ligand scaffold, new pair, or within-family interpolation?

Related

• Pocket
• Pose generation
• Scoring function
• Binding affinity
• Ranking metrics
• Negative set
• Applicability domain
• Metric
• Protein-ligand docking