What PharmCast Is Trained On

Where this collection came from

Enamine is a commercial supplier of screening compounds and building blocks. It markets this as the world's largest screening collection, and the point it presses hardest is physical availability: these are compounds synthesized in its own labs and held in stock, supplied as neat samples of roughly 50 to 150 milligrams or as 10 millimolar DMSO solutions, not a virtual enumeration. Around 3.12 million are held in US stock with two to three day delivery inside the USA, and a further 1.08 million in Ukraine stock. Roughly 225,000 new compounds are added each year, and material is quality controlled by LCMS or proton NMR to at least 90 percent purity.

That distinction matters for how this page should be read. Everything counted here is real, orderable material. Enamine separately publishes much larger virtual spaces, most notably REAL Space, enumerated by applying validated reactions to in-stock building blocks. None of that is included here.

Source: the June 2026 SDF release, Enamine_screening_collection_sdf_202606, from enamine.net/compound-collections/screening-collection, which currently states 4,774,670 compounds.

What the compounds look like

Properties from a random sample of 200,000 compounds. Heavy atom counts are from all 4,612,044. Bond counts run from 8 to 50.

This is a lead-like library rather than a fragment or a biologics-adjacent one. The median compound weighs 344, carries three rings of which two are aromatic, five rotatable bonds, and sits at cLogP 2.7. Only 8 percent carry defined stereochemistry, so the collection is predominantly flat, achiral scaffolding decorated with substituents.

Scaffold diversity

Reducing each molecule to its Bemis-Murcko scaffold, the ring systems and the linkers between them with all side chains stripped, gives a direct read on how many distinct structural frameworks the library is built from.

The framework diversity is high. Roughly seven in ten molecules bring a scaffold no other molecule in the sample shares, and no single framework dominates: the commonest is plain benzene at 2.1 percent, and the top ten together account for only 4.0 percent. Single rings head that list, so the table below steps past them to the commonest frameworks carrying two rings or more, which is what the library is really assembled from.

How much is distinct chemistry

Scaffold counts describe frameworks. They say nothing about how close any two whole molecules are. For that, every compound was compared against every other compound in the collection, with no sampling, keeping one number per compound: the highest Tanimoto similarity it reaches against any other, itself excluded. Compounds are represented by 2048-bit Morgan fingerprints at radius 2.

Reading the two results together

These two measurements pull in opposite directions, and both are true. At the framework level the library is broad: 679 distinct scaffolds per thousand molecules, most of them appearing once. At the whole-molecule level it is dense: the median compound has a neighbor at 0.714 and 55 percent sit within 0.70 of something else.

The resolution is that the collection is wide in frameworks and deep in analogs around them. Many scaffolds are present, and each is decorated many times over with closely related substituent patterns. For model building this matters directly. A random split will place near-identical analogs on both sides and report an optimistic number, so scaffold-aware splitting is the honest choice. For screening, uniform sampling spends much of its budget re-asking the same structural question.

The protein loop corpus

The second corpus is peptides taken from experimentally determined protein structures, described here because it is the same fingerprint space.

Structures come from an RCSB query for entries at 2.5 Angstrom resolution or better with R-free at or below 0.22, which returned 71,018 entries; 70,818 were parsed and 69,450 yielded usable fragments. Loops are runs of 2 to 6 residues lying between annotated secondary structure, read from the helix and strand records in the mmCIF itself rather than re-derived. That gives 1,524,535 loop instances across 135,408 protein chains, a median of 15 per entry, collapsing to 117,741 distinct sequences and 133,136 distinct capped molecules at the time of that extraction.

The peptide corpus has grown since. It now holds 136,494 fingerprinted peptides. PharmCast version 10 trains on 122,994 of them; the remaining 13,500 are reserved as a test set and are never trained on.

Each fragment is capped with its real flanking atoms, an acetyl built from the preceding residue and an N-methylamide from the following one, so it never presents a free amine and acid the protein does not have. Backbone continuity is enforced by requiring the carbon to nitrogen distance between consecutive residues to fall between 1.0 and 2.0 Angstroms rather than trusting residue numbering, which rejected 54,520 candidates that residue numbering alone would have accepted.

Repeat observations are kept rather than deduplicated, and the reason is measurable. Taking 1,500 sequences seen in three or more entries and comparing their observed fingerprints against each other, the median agreement is 0.878. Most instances of a loop are similar, as you would expect, but 18% of sequences fall below 0.70, carrying genuinely different pharmacophores between one crystal structure and the next, and only 19% are near copies. Collapsing them to one representative would discard that.

The activity-backed ChEMBL corpus

The third corpus is 1,214,214 activity-backed ChEMBL compounds from 142 to 1000 Da, each carrying at least one recorded activity measurement. It exists to extend the model past the mass and property range the screening collection covers, on chemistry that has been made and measured rather than enumerated.

What version 10 trains on

A further 13,500 peptides are held back as a test set and are never trained on, so they are not counted here.

The method behind PharmCast is described in Muskal, S. M.; McGregor, M. J. PharmCast: rapid generation of three-dimensional pharmacophore fingerprints from two-dimensional structure without conformer generation. bioRxiv 2026, doi: 10.64898/2026.09.02.748999.