Regression Transformer
GT4SD — Property-Conditioned Molecular Generation
Property-conditioned generation — specify a target logP and the model edits a seed molecule toward it, jointly predicting and generating.
Performance Metrics
| Metric | Value | Description |
|---|---|---|
| QED Prediction RMSE | < 0.06 | Drug-likeness prediction error (all RT variants, Table 1) |
| Generation Novelty | > 99.3% | Novel molecules across all property conditioning settings |
| Property Prediction | Matches/surpasses | Matches or surpasses conventional regression models for QED, ESOL, logP |
Test Set
ZINC + custom property labels, QM9
Notes
A single model both predicts properties and generates molecules conditioned on property targets. The metrics above are from the Nature Machine Intelligence paper, not from Spectra; detailed per-property tables are in that paper (paywalled). Spectra status: experimental, and verified locally rather than in production. The integration loads the IBM crippen_logp checkpoint and returns every molecule with its logP measured by RDKit plus the signed distance from your target, so you can see what a run did rather than taking the target on trust; that behaviour was checked against the answer key the checkpoint ships and a local acceptance sweep, not against a production run. Two limits are worth knowing before planning around it: the model edits a seed molecule rather than building one to order, and a local acceptance sweep measured it saturating near 3.4 logP, so higher targets are accepted but rarely reached; and the IBM model card for this checkpoint lists production-level inference as out of scope. Read the per-molecule measurements rather than assuming the target was met.