PharmCast

What more training data has bought

Every model in the family measured on the same held-out molecules, as the training set grew from 1,000,000 to 2,114,521.

Why these numbers are comparable. Both training scripts reserve chunks 0 to 19 of the screening collection, 9,975 molecules, and train on everything after. Every model in the table was therefore scored on the identical held-out set. The two scripts name the metric differently, test_median_mcc and val_median_mcc, but it is one measurement on one set.
2.1x
Training set growth across the family
0.893
Agreement at 2,114,521 molecules
58,039
Peptide ensembles in the newest model
6
Models in the family

Agreement against training set size

0.8860.8880.8910.8930.8961.0M1.2M1.5M1.8M2.0M 1.00M1.19M1.36M1.37M1.72M2.11M TRAINING MOLECULES MEDIAN AGREEMENT
Brown points are collection-only models, green points include peptides. Note the vertical scale.
ModelTraining moleculesCollection PeptidesAgreement
PharmCast-S 1.0M1,000,0001,000,000none0.888
PharmCast-S 1.2M1,192,6571,192,657none0.889
PharmCast-S 1.36M1,357,5471,357,547none0.891
PharmCast-SP v11,368,0861,357,54710,5390.890
PharmCast-SP v21,724,8331,688,79436,0390.891
PharmCast-SP v32,114,5212,056,48258,0390.893
What the held-out set does and does not cover. Chunks 0 to 19 are screening collection molecules only, so this metric measures catalogue chemistry. Peptide performance is not in it and is measured separately in the PharmCast-SP report. The screening collection is still being fingerprinted and the peptide corpus is still being extracted, so the training set continues to grow.

Learning curves

0.8560.8690.8820.895 EPOCH
PharmCast-S 1.0M PharmCast-S 1.2M PharmCast-S 1.36M PharmCast-SP v1 PharmCast-SP v2 PharmCast-SP v3
Median agreement by epoch, one line per model, in build order.