The current model is on the model weights page. Checksums are in SHA256SUMS. Each release below is selectable in the model menu on the compare and fingerprint pages.
PharmCast SCP v9 superseded
Screening collection + ChEMBL + peptides, trained 30 August 2026. Superseded by SCP v10 on 1 September 2026.
- Download
- pharmcast_scp_v9.pt 32,818,533 bytes (31.3 MB)
- SHA-256
- 1c0be82e38e1eff1f13dd0d633263e151c8dde5fad9db676a4a3a027209699c6
- Training set
- 5,126,378 molecules for gradient updates, drawn from a sealed snapshot of 5,178,160: 4,099,846 screening collection, 947,414 ChEMBL, 130,900 loop peptides. 1% (51,782) was reserved as a molecular-weight-stratified stopping set.
- Fingerprint width
- 10,549, one output per pharmacophore
- Parameters
- 8,045,877. Inputs 2,059, binary2048 features, hidden 1024 and 512, batch 256, learning rate 0.001, identical in shape to v8.
- Training
- 100 epochs on the Apple GPU in 353 minutes, the epoch-73 weights restored as the best stopping loss (0.0235018).
- Accuracy
- Measured on molecules the model cannot have seen. The
catalog and ChEMBL rows are every molecule fingerprinted after the
v9 corpus was sealed, 93,994 and 49,200, so n differs because the corpora
are built at different rates. The peptide row is the sequence-novel
peptide test set, 500 loops from PDB entries at 1.8 Å or better
and R-free 0.23 or better, excluding every entry used to build the training
corpus and every loop whose amino acid sequence already appears in
it, so those peptides are novel by sequence rather than merely by SMILES.
Per fingerprint rather than per pair, median MCC is 0.911 on all 93,994 catalog molecules fingerprinted after the snapshot was sealed. ChEMBL is the weakest population, with error about 50% higher than catalog chemistry and roughly one pair in five outside 0.05; error rises to 0.041 above 600 Da, where corpus coverage is thinnest.Population n Median error Within 0.05 Pearson r Ranking Catalog chemistry 93,994 0.015 92% 0.976 0.931 Loop peptides, test set 500 0.019 86% 0.979 0.943 ChEMBL, activity backed 49,200 0.026 76% 0.936 0.889 Reference against itself 0.006 0.995 ceiling - License
- Apache-2.0, same as the code. Keep the NOTICE file.
PharmCast SCP v8 superseded
Screening collection + ChEMBL + peptides, trained 28 August 2026. Superseded by SCP v9 on 30 August 2026.
- Download
- pharmcast_scp_v8.pt 32,737,957 bytes (31.2 MB)
- SHA-256
- d138903777958ff98e7e531086e5cbb65cd62291b7864c1971925e5ac8f42ab5
- Training set
- 4,239,599 molecules for gradient updates: 3,543,460 screening collection, 629,090 ChEMBL, 67,049 loop peptides. Drawn from a sealed snapshot of 4,282,423, of which 1% (42,824) was reserved as a molecular-weight-stratified stopping set.
- Fingerprint width
- 10,549, one output per pharmacophore
- Parameters
- 8,045,877. Inputs 2,059, hidden 1024 and 512, identical to v7.
- Training
- 100 epochs, the epoch-66 weights restored as the best stopping loss (0.0230779). This is the first PharmCast model trained on the GPU (Apple MPS) rather than the CPU. No wall-clock figure is quoted: training times here are machine dependent and were withdrawn from every published surface.
- Accuracy
- On every molecule fingerprinted after the v8
corpus was sealed, so none was seen in training: 96,916 catalog
compounds, 63,500 ChEMBL and 11,500 loop peptides.
Per fingerprint rather than per pair, median MCC is 0.909 on catalog chemistry, with a 10th percentile of 0.840 and a 90th of 0.947.Population Median error Within 0.05 Pearson r Ranking Catalog chemistry 0.016 91% 0.975 0.930 Loop peptides 0.019 84% 0.979 0.944 ChEMBL, activity backed 0.024 78% 0.937 0.890 All three combined 0.019 85% 0.972 Reference against itself 0.006 0.995 ceiling - Against the release it replaced
- Measured head to head on one common held-out set that neither model was trained on, v8 is ahead on ranking accuracy, median error and Pearson r in all three populations. The one metric where it is not is per-molecule MCC on loop peptides, 0.895 against 0.896. The clearest gain is ChEMBL, where the corpus grew 46%.
- License
- Apache-2.0, same as the code. Keep the NOTICE file.
PharmCast SCP v7 superseded
Screening collection + ChEMBL + peptides, trained 27 August 2026. Superseded by SCP v8 on 28 August 2026.
- Download
- pharmcast_scp_v7.pt 32,690,725 bytes (31.2 MB)
- SHA-256
- ad5aee5c1fdee405b12fed6d3deb5da01c7679d368de18a6e703f9d0b72d4aa8
- Training set
- 3,724,667 molecules: 3,263,245 screening collection, 430,915 ChEMBL, 30,507 loop peptides
- Fingerprint width
- 10,549, one output per pharmacophore
- Parameters
- 8,045,877. Inputs 2,059, hidden 1024 and 512.
- Training
- Early stopped at epoch 100 of a 300 epoch budget; the weights are epoch 57, the best stopping loss (0.0230824655), restored. Held out 37,623 molecules.
- Measured against the reference
- On 25 drug pairs with full reference ensembles (100 ETKDGv3 conformers, UFF relaxed), mean absolute error in PharmSim is 0.043, median 0.031. The model reads slightly high: mean signed error +0.030.
- License
- Apache-2.0, same as the code. Keep the NOTICE file.