PharmCast-SP v3

The composite model, and what the peptides bought

PharmCast-SP v2 predicts a full 3D pharmacophore fingerprint from a SMILES, trained on the screening collection plus ensemble-enhanced protein loop peptides.

The last run finished 2026-08-20 15:09:16: 2,056,482 screening collection molecules plus 58,039 loop peptides, 2,114,521 in all, 40 epochs in 258 minutes. Validation agreement finished at 0.893, and it is converged rather than truncated: the last ten epochs moved it by +0.0011.

0.893

Validation median agreement with the real fingerprint

2,114,521

Training molecules

58,039

Loop peptides, the composite ingredient

8,051,520

Parameters

Training

Agreement climbed from 0.869 at epoch 1 to 0.893 at epoch 40, best 0.893. The curve is flat over the last quarter, so this model stopped because it had finished rather than because the budget ran out.

How it performs, on three chemistries at once

Held out pairs from everything the model is asked to handle: collection molecules fingerprinted after the training snapshot, loop peptides whose names it never saw, and real ChEMBL compounds above 600 molecular weight, outside the size range the collection covers.

Real pfpall similarity against PharmCast across three chemistries — Real pfpall similarity across, PharmCast up. Dashed line is perfect agreement.

Test chemistry	Pairs	Median error	Within 0.05	Correlation
screening collection	1,200	0.02	85%	0.97
loop peptides	1,200	0.02	84%	0.98
large ChEMBL compounds	1,200	0.07	38%	0.61
All three together	3,600	0.03	69%	0.89

Read the combined row as coverage, not as a single accuracy: it blends two chemistries the model handles well with one it handles coarsely.

What the peptides bought

The peptide corpus is the whole reason this model differs from PharmCast-S, so the honest question is what it changed. On peptides it is decisive: median error 0.02 and correlation 0.98, on molecules whose sequences the model never saw. That is the composite doing exactly what it was built to do.

On everything else, S and SP are indistinguishable. Ranking accuracy on catalogue chemistry is 91.3% for SP v2 against 90.9% for S. On real ChEMBL compounds above 600 molecular weight it is 71.0% against 72.2%. Those gaps are far inside the noise. If peptide performance is not a deliverable, the peptides are not currently paying for themselves on ordinary chemistry, and that should be stated rather than glossed.

Where it degrades, and by how much

Chemistry	Ranking accuracy	Median error	Correlation
Catalogue chemistry, median MW 347	91.3%	0.02	0.96
Large molecules, median MW 638	71.0%	0.07	0.62

Both sets have essentially the same similarity spread, so this is a real size effect rather than a harder task. The ground truth is not the problem: rebuilding 149 of those large molecules with a different embedding seed reproduces pair similarity to 0.006 at correlation 0.995, so pfpall agrees with itself an order of magnitude more tightly than PharmCast agrees with it.

Under optimisation pressure the surrogate inflates. Measured on a CHIP design campaign that used this model as its scoring function: seventeen times the search budget raised the median surrogate score by 0.20 and the median real score by 0.02, and the surrogate's median error over survivors went from +0.06 to +0.24. A genetic algorithm pointed at PharmCast will eventually optimise the error rather than the molecule. Use it to rank and to filter, never to decide, and rescore with the real calculation before quoting anything.