Broad chemistry, optimized for drug-like characteristics

Most screening libraries are assembled. Ours is generated, filtered, and scored, which means we can say precisely what is in it and show the measurements behind every claim. This page is that description: what the catalog covers, how it is distributed, and where it is deliberately different from a curated bioactive library.

Narrow window, evenly filled

The defining property of the catalog is a combination that rarely occurs together. It occupies a deliberately narrow region of chemical space, and it fills that region evenly rather than clustering into repeated families.

Every comparison below was measured against 771,668 ChEMBL compounds using the same fingerprint and the same measurement code for both collections. Where the two columns were computed over different-sized populations, the row says so, because sample size affects some of these quantities.

Read those two similarity rows together, because they point in opposite directions and that is the finding. The higher mean pairwise similarity says the catalog sits inside a tighter, intentionally lead-like envelope. The lower nearest-neighbour similarity says that inside that envelope, molecules are further from their closest relative. Narrow region, evenly filled.

For a primary screening campaign that is the shape you want. Every molecule you screen buys new structural information rather than re-confirming a core you have already tested.

How close is the nearest relative?

This is the share of molecules having at least one neighbour at each similarity threshold. Every row below is measured the same way: 2,000 randomly chosen molecules, each compared exhaustively against 2.25 million others. We use direct comparison rather than reading neighbours off the clustering, because clustering systematically overstates how isolated a molecule is.

Catalog molecules are not isolated: essentially all of them have a related neighbour, and three quarters have a close one. The gap opens at the high thresholds, where the reference library’s published analog series show up. That difference is the tradeoff described further down, and it is deliberate.

A deliberately lead-like envelope

The catalog spans a narrower range than the reference on every property axis we measured, most visibly in molecular weight and ring count. Medians sit close to ChEMBL’s; the tails are cut.

That is a filtering decision, not an accident. Candidates outside drug-like property ranges are discarded, as are molecules that look impractical to synthesize and molecules flagged by medicinal-chemistry structural alerts for reactive groups or implausible ring systems. The result is a library where the property screen has already been applied.

Novel, and still growing into new chemistry

Our generative models are trained to be novel. Every candidate is checked against ChEMBL and other sources and discarded if found elsewhere.

That second row is a statement about the generator, not about the catalog. The rate at which they re-emit a compound that already exists in the training data measures how much they memorized. At 0.50% it is roughly four times lower than the size of the two collections alone would predict, and every one of those matches is discarded before import.

The models learned the style of bioactive chemistry rather than memorizing its contents, and the collection is still reaching structures it has not covered before rather than repeating itself.

What this catalog is not

The catalog averages roughly 2.4 molecules per scaffold, and very close analogs are about half as common as in the reference library. It is not built to hand you a ready-made structure-activity series around a hit. A diverse library is optimized for finding starting points; a focused library is optimized for developing them, and those are different collections.

In practice that means follow-up around a confirmed hit is best served by us generating analogs on demand around that specific scaffold.

How we measured this

Every molecule in the catalog was processed, not a sample of it. Each was converted to a Morgan fingerprint (radius 2, 2048 bits, bit vector) using the same code path that powers similarity search in the product, so the numbers on this page describe the same representation the application uses.

Clustering

Molecules were grouped with BitBIRCH, a tree-based clustering method built for large molecular libraries: it scales linearly with library size and takes a single interpretable parameter, a Tanimoto similarity threshold, rather than a target number of clusters. We used the diameter criterion, so a molecule joins a group only if the group’s average pairwise similarity stays above the threshold. Groups were built at two thresholds, 0.45 and 0.65, because coverage and tightness trade off against each other and neither setting alone describes the collection.

A refinement pass follows the initial build, and it is not optional. Tree-based clustering routes each molecule by comparing against group representatives, and in high-dimensional fingerprint space that misroutes: molecules get stranded alone even when a close relative is present. Skipping refinement overstated isolation by roughly threefold in our own first attempt, which is why the figures above lean on direct nearest-neighbour search instead.

The full catalog does not fit in a single clustering tree on ordinary hardware, so it was processed as five contiguous blocks covering the entire id range end to end. The whole pipeline takes about three hours.

The reference arm

Every ChEMBL comparison on this page was computed locally over 771,668 compounds, with the identical fingerprint and the identical code that processed our own molecules. We deliberately avoided quoting published figures for reference libraries: different fingerprints and different implementations produce different numbers, and a comparison assembled from two sources is not a comparison.

What is measured on what

Some figures come from the complete catalog and some from samples, and the distinction matters when reading them.

Samples are drawn across the full id range rather than from the beginning of the catalog. Molecules were added over time and the generator’s settings changed as it went, so the earliest molecules are not representative of the whole. Nearest-neighbour figures come from exhaustive comparison against every other molecule, not from cluster membership, because clustering systematically over-reports isolation even after refinement.

Known limits

Scaffold counts use Bemis-Murcko perception, which treats a single ring-heteroatom change as a different scaffold. That makes our analog-depth numbers conservative, and it applies equally to both collections. Morgan fingerprints ignore stereochemistry, so stereoisomer pairs register as identical. All similarity figures use one fingerprint and one metric; other representations would shift the absolute values, though the catalog-versus-reference comparisons are computed the same way on both sides.