Structure-based virtual screening, step by step
Maintained by the Talindrew team · Last updated
Structure-based virtual screening ranks a library of small molecules by how well they are predicted to bind a target protein's pocket. A typical workflow runs in a fixed order: obtain or predict the target structure, locate druggable pockets, prepare a compound library, dock each compound into the chosen pocket, filter by predicted ADMET and drug-likeness, discard physically implausible poses, check selectivity against anti-targets, and combine the evidence into one ranking. Docking scores alone rank poorly across chemotypes, which is why the later filters matter.
Watch · 90 sec
Drug Discovery in 90 Seconds: Build and Run a Screen in Talindrew's AI Workspace
Chapters
The workflow, step by step
Each step consumes the previous one's output; changing one step only invalidates the steps after it.
- Target structure: an experimental PDB entry when one exists, otherwise a predicted structure (ESMFold, Boltz-2 or AlphaFold 3-class models), checked for confidence around the pocket.
- Pocket prediction: find druggable cavities with a tool such as P2Rank or fpocket and pick the one that matters biologically.
- Compound library: approved drugs for repurposing, fragments, natural products, or your own SMILES.
- Docking: place each compound in the pocket and score poses with AutoDock Vina or a GPU docking engine such as Uni-Dock.
- ADMET and drug-likeness: predicted absorption, distribution, metabolism, excretion and toxicity, with ADMET-AI or rule filters.
- Pose checks: reject poses with clashes or impossible geometry using PoseBusters, and rescore the shortlist, for example with Boltz-2 affinity.
- Selectivity: dock the shortlist against anti-targets and penalise compounds that bind them as strongly.
- Ranking: weigh the evidence into one ordered list, then design analogs of the best hits for the next round.
How reliable are docking scores?
AutoDock Vina's score is an empirical estimate in kcal/mol that is useful for ordering compounds against one pocket but correlates only loosely with measured affinity. Treat it as one line of evidence, and confirm top hits experimentally.
Tools by step
| Step | Common open tools | In Talindrew's FoldRx |
|---|---|---|
| Structure | ESMFold, Boltz-2, AlphaFold 3-class models | ESMFold, Boltz-2, a structure you bring, your own model |
| Pockets | P2Rank, fpocket | P2Rank (default), fpocket |
| Docking | AutoDock Vina, Uni-Dock | Vina (default), Uni-Dock |
| ADMET | ADMET-AI, RDKit rules | ADMET-AI, RDKit rules |
| Pose checks | PoseBusters, Boltz-2 affinity | Both |
| Analogs | CReM | CReM |
Frequently asked questions
How many compounds should a first screen dock?
A few hundred is enough to see whether a pocket discriminates between chemotypes and runs in minutes; widen the library once the pipeline's ranking makes sense for known binders.
Do I need a predicted structure if the protein is in the PDB?
No. An experimental structure of the right conformation is usually the better starting point; use prediction when there is none, or to model a mutant or a different construct.
References
- Krivák R, Hoksza D. P2Rank: machine learning based tool for rapid and accurate prediction of ligand binding sites from protein structure. J Cheminform 10: 39 (2018). https://doi.org/10.1186/s13321-018-0285-8
- Trott O, Olson AJ. AutoDock Vina: improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. J Comput Chem 31(2): 455–461 (2010). https://doi.org/10.1002/jcc.21334
- Eberhardt J, Santos-Martins D, Tillack AF, Forli S. AutoDock Vina 1.2.0: new docking methods, expanded force field, and Python bindings. J Chem Inf Model 61(8): 3891–3898 (2021). https://doi.org/10.1021/acs.jcim.1c00203
- Swanson K, Walther P, Leitz J, et al. ADMET-AI: a machine learning ADMET platform for evaluation of large-scale chemical libraries. Bioinformatics 40(7): btae416 (2024). https://doi.org/10.1093/bioinformatics/btae416
- Buttenschoen M, Morris GM, Deane CM. PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences. Chem Sci 15: 3130–3139 (2024). https://doi.org/10.1039/D3SC04185A
- Polishchuk P. CReM: chemically reasonable mutations framework for structure generation. J Cheminform 12: 28 (2020). https://doi.org/10.1186/s13321-020-00431-w
Related
Design your Structure-based virtual screening, step by step construct in Talindrew.
Free to use. No download required. Works in any modern browser.