Ultra-large library docking: billions of compounds, better hits

For decades the intuition was that bigger screening libraries would mostly add noise: more compounds means more false positives scoring well by chance, so a docking campaign across a hundred million molecules should bury its real hits. The opposite turned out to be true. When libraries grew from a few million to hundreds of millions of make-on-demand compounds, hit rates and potencies went up. Understanding why is the single most useful thing to know about modern structure-based virtual screening.

What “make-on-demand” means

The libraries that drove this shift are not shelves of physical vials. They are combinatorial enumerations of compounds thatcould be synthesized on request from a catalog of building blocks and a set of reliable, validated reactions. The best-known is Enamine’s REAL space. You search it computationally, dock the virtual molecules, then order only the few hundred top scorers for physical synthesis and assay. The synthesis success rate for these reactions is high enough (typically 70-80%) that the virtual library behaves like a real one for screening purposes.

The size matters because chemical space is astronomically underexplored. A few million purchasable compounds is a rounding error against the ~10^60 drug-like molecules that could in principle exist. Make-on-demand libraries pushed the accessible slice from millions to billions, and that extra coverage is where the new chemotypes were hiding.

The result that changed the field

Lyu et al. (2019) docked around 138 million make-on-demand compounds against two targets: the enzyme AmpC beta-lactamase and the dopamine D4 receptor. The numbers were unusually good. For D4, of 549 compounds tested, 122 were active, a 22% hit rate, and they spanned 81 new chemotypes including a 180-picomolar, subtype-selective agonist. For AmpC they found one of the most potent non-covalent inhibitors of that enzyme reported at the time. The headline lesson: with a large enough, chemically diverse library, the very top of the docking rank-ordered list is enriched for genuine binders, not artifacts.

Why does more compounds help rather than hurt? Two reasons. First, a bigger library samples the pocket’s ideal complementary shape more finely, so the best-fitting molecules fit better in absolute terms. Second, with millions of candidates you can afford to be ruthless at the top: you only ever test the extreme tail of the score distribution, where the enrichment is strongest. You are not testing average compounds, you are testing the best-in-100-million.

How you dock a billion compounds without a supercomputer-year

Brute force does not scale forever. Two strategies make the very largest spaces tractable:

• Massive parallel brute force. Platforms like VirtualFlow (Gorgulla et al., 2020) distribute the dock across thousands of cloud or cluster cores with near-perfect scaling. That team prepared more than 1.4 billion molecules in ready-to-dock format and screened over a billion against KEAP1, recovering a nanomolar inhibitor.
• Combinatorial shortcuts. V-SYNTHES (Sadybekov et al., 2022) avoids enumerating the full space at all. It docks the building-block fragments (synthons) first, keeps the best scaffold-synthon seeds, then grows only the promising ones, guided by the reaction rules of the library. It searched an 11-billion-compound space while explicitly docking less than 0.1% of it, and validated hits against ROCK1. This is the direction the field is heading as libraries pass tens of billions.

The caveats that still apply

Scale does not repeal the limits of a docking score. Scoring functions remain the weak link: they approximate binding free energy and systematically mis-rank some chemotypes. Bigger libraries amplify any pose or protonation-state error because you are operating at the extreme tail where small artifacts get promoted. Practical screens still depend on a good receptor structure, careful pocket definition, and re-scoring or visual triage of the top poses before anything gets ordered. The size of the library raises the ceiling; it does not fix a bad setup.

Try the docking yourself

You do not need a billion-compound cluster run to use the same logic. The endpoint of any ultra-large screen is a short list of individual candidates you want to inspect carefully against your target and its mutants. Open Studio and dock your shortlisted compounds against the target structure, then read the pose and the ADMET panel before you commit synthesis budget. Comparing the wild-type and mutant ΔΔ on a handful of top scorers is exactly the triage step a large screen leaves you with.

Liganx brings molecular docking online and free in the browser, so the per-compound validation that follows a large virtual screen does not need a local install. Running molecular docking on your final shortlist is the cheapest insurance against ordering the wrong molecules.

Primary sources

• Lyu J, et al. Ultra-large library docking for discovering new chemotypes. Nature 566, 224-229 (2019). doi:10.1038/s41586-019-0917-9
• Gorgulla C, et al. An open-source drug discovery platform enables ultra-large virtual screens. Nature 580, 663-668 (2020). doi:10.1038/s41586-020-2117-z
• Sadybekov AA, et al. Synthon-based ligand discovery in virtual libraries of over 11 billion compounds. Nature 601, 452-459 (2022). doi:10.1038/s41586-021-04220-9