Virtual screening

Virtual screening is the process of using computational methods to evaluate large numbers of existing compounds against a biological target (usually a protein) and rank them by predicted activity. Rather than testing every compound in a large library experimentally — which may have millions of entries — you screen computationally and then test only the top-ranked candidates.

Two main approaches:

Structure-based virtual screening uses the 3D structure of the target protein (from X-ray crystallography, cryo-EM, or AlphaFold prediction) to dock candidate molecules into the binding site and score the predicted interaction. The score estimates binding affinity.

Ligand-based virtual screening doesn’t require a protein structure. Instead, it uses known active compounds as a reference — finding new molecules similar in shape, pharmacophore, or molecular fingerprint to compounds already known to work. Useful when no experimental structure is available.

How AI has changed virtual screening:

Classical docking programs (Glide, AutoDock) use physics-based scoring functions. AI-based scoring functions and property prediction models (Chemprop, DeepDock) can score candidates faster and sometimes more accurately by learning from experimental data rather than physics alone. Large-scale virtual screening campaigns that previously took weeks on a cluster can now run faster on smaller hardware.

Limitations:

• Docking scores correlate imperfectly with actual binding affinity — false positives are common
• Protein flexibility is hard to model; the protein conformation in a crystal structure may not reflect the active conformation
• Predictions on novel scaffolds (structurally different from training data) are less reliable

Related terms: Structure-Based Drug Design, De Novo Design

Related guide: Health & Medicine