SMI-TED 289M
IBM FM4M (SMI-TED) — Molecular Generation & Property Prediction
Large encoder-decoder foundation model (289M params) pre-trained on 91M SMILES. Generation + property prediction.
Sources
Performance Metrics
| Metric | Value | Description |
|---|---|---|
| Validity (fragment) | 99.99% | Generated SMILES that are chemically valid (MOSES benchmark) |
| Scaffold Novelty | 99.99% | Novel Bemis-Murcko scaffolds |
| Internal Diversity | 85.65% | Pairwise Tanimoto distance within generated set |
| FCD Score | 1.153 | Fréchet ChemNet Distance (lower = more realistic) |
| BBBP (ROC-AUC) | 92.26% | Blood-brain barrier permeability classification |
| ClinTox (ROC-AUC) | 94.27% | Clinical toxicity classification |
| BACE (ROC-AUC) | 88.24% | BACE-1 inhibitor classification |
| ESOL (RMSE) | 0.611 | Aqueous solubility prediction (log mol/L) |
| Lipophilicity (RMSE) | 0.552 | Lipophilicity prediction (logD) |
| FreeSolv (RMSE) | 1.223 | Solvation free energy prediction (kcal/mol) |
Test Set
MOSES benchmark (generation), MoleculeNet (property prediction)
Notes
Pre-trained on 91 million SMILES from PubChem. Evaluated on both MOSES generation benchmarks and MoleculeNet property prediction tasks. Classification results are ROC-AUC (higher is better), regression results are RMSE (lower is better).