Compound libraries for virtual screening: ZINC, Enamine REAL and more

What each library lets you buy, dock and trust

17 min

Three nested circles holding fewer and fewer small hollow rings toward the centre, where one green ring is joined by two dashed lines to a large green circle containing the structure of haloperidol.

Which compound library should you screen?

Start with an in-stock collection of a few million compounds when you want assay results within days, move to a make-on-demand space such as Enamine REAL when you need chemistry nobody has on a shelf, and use public sets such as ChEMBL when you need known actives. ZINC is not a vendor. It is a free database that gathers vendor catalogs and prepares them for docking.

The size of a library says less than it seems to. Lyu and colleagues docked 138 million make-on-demand molecules against the dopamine D4 receptor, picked 549 of them to be synthesized, and found that 122 (22%) displaced a radioligand at 10 μM.[4] The hit-rate curve in the same paper put 453,000 D4 ligands in that library, so the screen found about one in 3,700 of them. The shortlist was 0.0004% of what was docked.

Three questions decide which library fits. How much of it can you dock, how many of its molecules can actually be bought, and does your scoring setup already pick known binders out of decoys? The first has a number (Table 3), the second a delivery time and a success rate (Table 1), and the third decides whether a bigger library helps at all.

What libraries exist for virtual screening?

Three kinds of library feed a virtual screen. In-stock catalogs list compounds sitting in a vendor's freezer, ready to ship in days. Make-on-demand spaces list molecules that have never been made but can be built from a vendor's building blocks and reactions. Public bioactivity sets hold compounds with measured activities, which you screen for what is known rather than what can be bought.

LibraryTypeSizeWhat it is good for
Enamine Screening CollectionIn stock4,824,879 compoundsFirst pass with results in days; pre-plated subsets of 100,160 to 460,160 compounds
eMoleculesIn stock20.3 million screening compoundsMany suppliers in one order; average in-stock ship time of 2 days
ZINC20 in-stock setIn stock, aggregatedAbout 4 million compoundsAn "informer set" to screen before a large campaign
Enamine REAL DatabaseMake on demand, enumeratedOver 16.1 billion moleculesDrug-like subset small enough to list and dock
Enamine REAL SpaceMake on demand, combinatorial113.7 billion (BioSolveIT) or 116.3 billion (Enamine)Analog search around a hit; docking only after subsetting
WuXi GalaXi and Mcule UltimateMake on demand2.5 billion and 128 million (2022 snapshot)Additional chemistry not in REAL
ZINC-22Aggregator of make-on-demand37 billion in 2D, 4.5 billion docking-ready in 3D (2022)Free download of tranches already built for docking
eMolecules eXploreMake on demand, combinatorialClaimed over 60 trillion; samples of up to 3 billion can be downloadedReach beyond REAL; treat the count as a space, not a list
ChEMBL 37Public bioactivityRelease of May 2026Known actives, approved drugs, benchmarks and decoy sets
Table 1. Compound libraries used for virtual screening. Sizes are the figures each source states on the date shown; vendor figures are company claims. REAL Space is reported by Enamine and by BioSolveIT with slightly different values. Sources: Enamine (2026), BioSolveIT (2026), eMolecules (2026), Tingle et al. (2023), ChEMBL (2026).[5][7][8][9][10][11][3][12]

Make-on-demand does not mean made on demand for everyone. Enamine states that its compounds are synthesized in 3 to 4 weeks with a success rate above 80%, and splits its REAL compounds into tiers with success rates of 85% or more (s-REAL), over 75% (m-REAL, priced 50% higher) and 65% or more (u-REAL).[5][6] Lyu and colleagues' own orders did better than the headline: 44 of 51 compounds (86%) for AmpC and 549 of 589 (93%) for D4 arrived.[4]

What is ZINC, and which version should you use?

ZINC is a free database from the Irwin and Shoichet laboratories at the University of California, San Francisco that lists purchasable compounds and builds them into 3D, docking-ready files. It has never sold anything. Each version grew with what vendors could offer, and the version you cite matters because the sizes differ by a factor of 300.

ZINC15, published in 2015, held over 120 million purchasable drug-like compounds, a quarter of them available for immediate delivery.[1]

ZINC20, published in 2020, added make-on-demand catalogs and took the database into the billions. Its authors found that over 97% of the core Bemis-Murcko scaffolds in make-on-demand libraries are unavailable from in-stock collections, and that an 88-fold increase in molecules between the two was built on a 16-fold increase in scaffolds.[2]

ZINC-22, published in 2023, moved to billions as the default. It held over 37 billion enumerated compounds in 2D, of which over 4.5 billion had been built in ready-to-dock 3D formats. The 2D set came from Enamine REAL Database (5 billion), Enamine REAL Space (29 billion), WuXi (2.5 billion), Mcule (128 million) and the ZINC20 in-stock set (4 million).[3] The authors expected to add about 300 million molecules a month at full speed.[3]

For new work, take make-on-demand tranches from ZINC-22 and the in-stock informer set from ZINC20, which the ZINC-22 authors say continues to be maintained. Older papers that say "docked ZINC15" mean the 2015 release, and their numbers are not comparable with a ZINC-22 subset. The tranche browser in ZINC-22 splits the database by properties and offers common subsets such as lead-like and fragment-like, so you download only the slice you will dock.[3]

Size is where ZINC becomes a logistics problem. One billion molecules take about 1 TB in each of mol2, SDF and PDBQT formats, and about 200 TB in the db2 flexibase format used by DOCK. The authors estimate that 4.5 billion molecules in db2 form would need nearly a petabyte and almost three months to download at 1 Gbps.[3] Most users never download all of it, and should not try.

How big is Enamine REAL, and what does it contain?

Enamine REAL is a set of molecules that can be built from about 213,000 building blocks using 157 reaction protocols, according to Enamine; BioSolveIT, which hosts the browsable version, counts 211,806 building blocks and 161 reactions.[5][8] Because the molecules are combinations, nobody keeps them in a freezer. The vendor synthesizes the one you order.

Two sizes are quoted. REAL Database is the enumerated, drug-like subset, over 16.1 billion molecules that obey Lipinski and Veber limits.[7] REAL Space is the full combinatorial space, which Enamine puts at 116.3 billion molecules and BioSolveIT at 113.7 billion after its October 2026 update.[5][8] The two pages disagree by about 2%, and neither is dated; cite the one you used.

The growth has been steep. BioSolveIT reports that REAL Space grew from about 650 million molecules at its launch in February 2018 to 113.7 billion, roughly 175 times larger.[8] ZINC-22 captured the Enamine portion at about 34 billion in 2022, so a screen built from it already misses about 70% of today's REAL Space.

Enamine also sells a stock collection for the screening that needs no synthesis. Its Screening Collection of 4,824,879 compounds splits into five parts (HTS 1,949,943; Legacy 1,708,725; Advanced 843,801; Premium 71,880; Functional 250,530), at least 90% pure by LCMS or NMR, and about 225,000 new compounds are added each year.[9]

What other libraries and public sets matter?

eMolecules lists over 50 million unique structures in its catalog, including 20.3 million screening compounds and 4.5 million building blocks, and adds a virtual space called eXplore that it describes as over 60 trillion synthesizable compounds.[10][11] The trillion-scale figure is a count of what the reactions could produce, not of molecules anyone has enumerated, and the page offers downloadable samples of up to 3 billion.

WuXi's GalaXi and Mcule's Ultimate collection are the two other make-on-demand catalogs that ZINC-22 ingested directly, at 2.5 billion and 128 million molecules in 2022.[3] The ZINC-22 authors add that ZINC20 remains their record of over 200 smaller catalogs, which is the long tail for niche chemotypes.[3] Vendors such as Molport, Chemspace and Mcule publish their own counts, which we did not verify and do not quote here.

ChEMBL belongs in a different category: it is not for sale. Release 37 appeared in May 2026,[12] and it holds the measured activities that tell you which compounds are known binders of your target.

It is where you get actives to validate a setup, approved drugs for a repurposing screen (we counted 2,104 approved small molecules in release 37 for our PAINS filter analysis), and the property distributions to build decoys from. PubChem is a similar source of deposited structures and assay results, under different curation.

A worked example: the 138 million molecule D4 screen

The best-documented ultra-large screen is the one from Lyu and colleagues, because the paper reports every stage from library to bench. Table 2 follows the D4 receptor campaign through those stages and shows how the library, the prediction, the shortlist and the result line up. They disagree by orders of magnitude.

StageMoleculesNote
Docked to the D4 orthosteric site138 million43,563 core hours, about 1.2 calendar days on 1,500 cores
D4 ligands predicted in the library453,000 (0.33%)From the observed hit-rate curve, assuming equal hit rates within each score range
Selected for synthesis589 (0.0004%)Chosen from the top 1,000 clusters, by eye or by score alone
Delivered by Enamine549 (93% of those ordered)Make-on-demand synthesis
Hits at 10 μM122 (22% of those tested)Hit rate was 22 to 26% in the top-scoring bins and fell to zero in the worst
Measured with a dose response81Ki from 18.4 nM to 8.3 μM
Table 2. Funnel of the 2019 dopamine D4 screen. Percentages are fractions of the library docked, except where noted. Hits are compounds that displaced more than 50% of specific radioligand binding at 10 μM. Source: Lyu et al. (2019).[4]

Three things follow. The 22% hit rate describes the top of the ranking, not the library, which by the paper's own curve is 0.33% ligands. Selecting compounds by eye or by score alone gave similar hit rates, about 24% each, so the ranking carried the result.[4]

Clustering before docking would have saved time but cost the result. Docking only one representative per cluster cut docking time 22-fold, but the 47 confirmed actives among the top 3,000 ranked molecules fell by an average of 1,121,443 ranks once replaced by their representatives, and only two of the original active scaffolds stayed on the list.[4] Cluster after docking, not before.

What does docking a whole library cost?

The D4 screen gives a unit cost.

Dividing 43,563 core hours by 138 million molecules gives 1.14 core-seconds per molecule, with DOCK 3.7 sampling about 3,300 orientations and 479 conformations of each.[4] Table 3 scales that figure to each library we have discussed. It is our calculation, and it is only as portable as that docking setup: another engine will take more or less time per molecule, and the table should be rescaled by your own per-molecule time.

LibraryMoleculesCore hoursDays on 1,500 coresPDBQT storage
Enamine Screening Collection4.8 million1,523Under 1About 5 GB
Lyu D4 screen138 million43,5631.2About 138 GB
ZINC-22, docking-ready4.5 billion1.42 million39.5About 4.5 TB
ZINC-22, all in 2D37 billion11.7 million324About 37 TB
Enamine REAL Space113.7 billion35.9 million997About 114 TB
Table 3. What docking each library would cost at the D4 screen's speed. We calculated core hours as library size times 1.14 core-seconds per molecule (Lyu et al., DOCK 3.7.2), days as core hours divided by 1,500 cores and 24, and storage from the ZINC-22 figure of about 1 TB per billion molecules in PDBQT format.[4][3]

The point of the table is the last row. Docking all of REAL Space at the speed of the D4 screen would occupy a 1,500-core cluster for nearly three years, while the library keeps growing. Nobody docks REAL Space whole. They subset it by property or search it around a hit, and dock only the molecules that survive.

Does a bigger library find better compounds?

It finds better scores whether or not it finds better compounds.

The best score among N unrelated random molecules improves with N even when none of them binds, which means a bigger library will produce a more impressive top hit by chance alone. We calculated how much: if decoy scores followed a normal distribution with a standard deviation of 1 kcal/mol, the expected best score would sit 3.2 kcal/mol below the mean at 1,000 molecules, 4.9 at a million and 6.1 at a billion.

Library sizeBest decoy, in standard deviations below the meanGain over a 1,000 molecule screen, if one standard deviation is 1 kcal/mol
1,0003.230
1 million4.851.6 kcal/mol
1 billion6.072.8 kcal/mol
100 billion6.773.5 kcal/mol
Table 4. Best score expected from N decoys that do not bind. Idealized simulation, not a docking result: decoy scores are drawn from a normal distribution, and the expected maximum comes from Blom's approximation, checked against a Monte Carlo simulation up to one million molecules (3.25 against 3.23 at 1,000 and 4.89 against 4.85 at one million). Real score distributions are skewed, so the sizes of the effect would differ. We calculated this; it is not from the cited papers.

So a cutoff such as "keep anything below −9 kcal/mol" means different things in a 10,000 molecule screen and in a billion molecule one, and a docking score from one library cannot be compared with a score from another. Cut by rank within the library you docked, as Lyu and colleagues did, and judge by hit rate at the top.

Lyu and colleagues tested whether size helps by mixing known ligands with property-matched decoys and growing the decoy set from tens of thousands to hundreds of millions. When docking enriched ligands in the small library, performance typically improved as the library grew. When docking performed poorly in the small benchmark, it often got worse.[4] A larger library is a multiplier on whatever the scoring function already does.

That is the reason to check a setup on a small set of known binders and decoys before paying to dock a billion molecules. A scoring function that cannot rank 50 known ligands among 5,000 decoys will bury the same 50 under a billion.

How should you prepare a library before docking?

Vendor and database files are not docking inputs yet. A compound has to carry the protonation state, tautomer and stereochemistry that it would have in the assay, and ZINC-22 does this with ChemAxon's protonation and tautomer calculation at or near physiological pH before it builds 3D structures.[3] A list you assemble yourself, from a paper's supplement or a ChEMBL query, needs the same treatment, which Dimorphite-DL enumerates for a pH range.

Filtering comes next, and it should be property filtering first.

The lead-likeness filter keeps molecular weight at or below 350 Da and logP at or below 3, which leaves room for optimization that Lipinski's rule of five does not. Then remove assay-interference chemotypes: the PAINS filters flag 5.8% of approved drugs, so treat an alert as a prompt to check the structure rather than a verdict, and the Brenk filter adds 105 structural alerts for reactive or toxic groups.

After docking, Butina clustering turns a ranked list into scaffolds so a shortlist is not twenty close analogs, and molecular similarity search finds neighbors of a promising hit inside a library you already hold. Check the top poses with PoseBusters before you spend money on synthesis.

Which library for which question?

Your questionStart withAddCheck
Find first hits quickly for a new targetAn in-stock collection of 1 to 5 million compoundsLead-like and PAINS filtersWhether a known active lands in the top 1% of the ranking
Find chemistry outside current catalogsLead-like tranches of ZINC-22 or REALClustering after docking, not beforeVendor success rate and tier before you order
Expand a hit you already haveAnalog search in REAL SpaceDocking of the analogs onlyRedock the original hit to confirm the pose
Test whether your docking setup worksKnown actives from ChEMBL with property-matched decoysTwo scoring functionsEnrichment, not pose RMSD alone
Look for a new use for existing drugsApproved small molecules from ChEMBLThe target's known ligands as controlsWhether known binders rank in the top 1%
Rank a vendor's list you were sentThe list itselfProtonation and 3D preparationWhether any compound in the list is a known interferer
Table 5. Choosing a library by what you need to learn. The check column is the cheap test that catches the usual failure for that choice.

Run the in-stock pass first even when you plan to reach for billions. It costs hours rather than weeks, it tests the whole chain from receptor preparation to assay, and it tells you what hit rate to expect before you commit to synthesis.

Virtual screening library tools on ProteinIQ

The order is prepare, filter, dock, then check. A hosted AutoDock Vina job takes up to 10 ligands, and larger lists run through the API or batches, so use the filters to cut a library to the size you intend to pay for. The structure-based virtual screening and ligand-based virtual screening pages describe the surrounding workflows, and small-molecule discovery carries the hits on to ADMET triage and simulation.

TaskTool
Fetch a known compound from ChEMBL by IDChEMBL Download
Fetch a compound from PubChem by name, CID or InChIKeyPubChem Download
Enumerate protonation states for a pH rangeDimorphite-DL
Keep lead-like compoundsLead-likeness filter
Flag assay-interference substructuresPAINS filter
Flag reactive and toxic groupsBrenk filter
Calculate properties for a whole listMolecular descriptors
Find analogs of a hit in your own listMolecular similarity search
Prepare SMILES as PDBQT ligandsSMILES to PDBQT
Dock and score a ligand setAutoDock Vina, GNINA, smina
Group a ranked list into scaffoldsButina clustering
Check that top poses are physically plausiblePoseBusters
Predict ADMET properties of the shortlistADMET-AI
Table 6. Library preparation, docking and checking tools on ProteinIQ.

Frequently asked questions

What is the ZINC database?

ZINC is a free database of commercially available compounds from the Irwin and Shoichet laboratories, built so that the molecules can be docked directly. ZINC-22, the 2023 release, holds over 37 billion compounds in 2D and over 4.5 billion in ready-to-dock 3D formats.[3] You download subsets by property tranche rather than the whole thing.

Is ZINC15 still worth using?

ZINC15 held over 120 million purchasable compounds when it was published in 2015.[1] It is the version cited in older docking papers, so it is useful for reproducing them. For a new screen, ZINC-22 and the in-stock set in ZINC20 cover everything it did and billions more, and their subsets are the ones vendors currently list.

How big is Enamine REAL?

Enamine reports 116.3 billion make-on-demand molecules in REAL Space, and BioSolveIT 113.7 billion after its October 2026 update; the enumerated REAL Database is over 16.1 billion.[5][8][7] Enamine states that synthesis takes 3 to 4 weeks with a success rate above 80%.

How many compounds should I dock in a virtual screen?

Dock as many as your validated setup can rank and your budget can synthesize, and no more. A million-compound screen of a lead-like subset takes about 316 core hours at the D4 screen's speed, which is about a day on a 16-core workstation, while all of REAL Space would take nearly three years on 1,500 cores (our calculation, Table 3).

Are make-on-demand compounds as reliable as in-stock ones?

They are slower and less certain. Enamine's REAL tiers carry success rates from 65% to 85% or more, and Lyu and colleagues received 86% and 93% of the compounds they ordered.[6][4] Plan for some to fail and order a few extra.

Is ZINC free to use?

ZINC-22 is described by its authors as freely available for access and download, and the article is published open access.[3] The compounds themselves are not free: ZINC links them to vendors, and you buy from the vendor.