PharmCast · training trajectory

What more training data has bought

How agreement with the reference calculation has moved as the training set has grown, with every model in the family scored on one common held-out population.

PharmCast, how it works What it is trained on Try it live at pharmcast.ai
Why these numbers are comparable. Every model on this page is scored on the same 155,648 screening collection test cases, real purchasable compounds the training-set ingest filter excluded from every model. The points differ only by the model, so the curve measures training-set size and nothing else.
0.881Median agreement, PharmCast version 10
5,887,229Molecules version 10 trains on
155,648Held-out test cases, the same for every model
9Models on one axis

Agreement against training set size

Median agreement with the reference fingerprint against training set size, one point per release, every point measured on the same held-out set of 155,648 screening collection test cases. Agreement rises from 0.833 at 1.72 million training molecules to 0.881 at 5.89 million and flattens at the top end.
Median per molecule agreement against training-set size, one point per release, all on the same 155,648 held-out test cases.

Agreement reads 0.833 at 1.72 million training molecules, 0.837 at 2.11 million, 0.840 at 2.60 million, 0.845 at 2.89 million, 0.855 at 3.28 million, 0.877 at 3.72 million, 0.879 at 4.24 million, 0.880 at 5.13 million and 0.881 at 5.89 million. The returns from growth are real and they are diminishing: the last three releases sit within two thousandths of one another, which is what says where further growth in the training set is likely to land.

What went into each model

Stacked bars of training molecules per model release, split into screening collection, ChEMBL and loop peptides, rising to 5,887,229 molecules at version 10.
Training-set composition and total across the three sources for each model. PharmCast version 10 trains on 5,887,229 molecules.

The three sources are the Enamine screening collection, activity-backed ChEMBL compounds from 142 to 1000 Da, and peptide loops excised from crystal structures. Every molecule is fingerprinted with the reference 100-conformer calculation, which is what makes the set slow to build and what makes the curve above worth reading.

Talk to us about PharmCast

We are happy to walk through the method, the validation, and where the limits sit.

Email us How PharmCast works Where we use it: ChIP

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.