| --- |
| title: Chromatography RT Predictor |
| emoji: 🧪 |
| colorFrom: blue |
| colorTo: indigo |
| sdk: gradio |
| sdk_version: "5.25.0" |
| app_file: app.py |
| pinned: false |
| --- |
| |
| # Laboratory-Conditioned Chromatographic Retention-Time Predictor |
|
|
| Predicts retention time with archived fingerprint neural-network fold models for one of 23 laboratory labels represented during training. The interface does not run the GAT/GCN/ExtraTrees stack. |
|
|
| ## Dataset |
| - HighResNPS-derived forensic toxicology data |
| - 3,776 structure--laboratory observations |
| - 1,357 InChIKey connectivity groups and 23 represented laboratory labels |
| - Laboratory-specific chromatographic method metadata are not encoded by this interface |
|
|
| ## Model Architecture |
| The deployed component is an **FPNN** using radius-2, 2048-bit Morgan fingerprints and a learned embedding for the 23 represented laboratory labels. Predictions from the available fold models are averaged. |
|
|
| ## Evaluation Scope |
|
|
| The study evaluation uses molecular-identity-grouped and scaffold-aware holdouts over three prespecified split repetitions. Component and full-stack results are reported separately in the accompanying manuscript. This interface does not claim zero-shot prediction for a new laboratory or chromatographic method. |
|
|
| ## Usage |
| Enter a SMILES string and select the target laboratory. The app returns: |
| - Predicted retention time (minutes) |
| - Uncalibrated fold-model disagreement (standard deviation) |
| - Maximum development-set Morgan-fingerprint Tanimoto similarity as structural context |
| - Key molecular descriptors |
|
|
| Neither the fold-model disagreement nor the similarity score is a predictive interval or a guarantee of prediction reliability. |
|
|
| The `models/` checkpoint directory and `data/training_fps.npz` are required for local deployment. |
|
|
| ## Citation |
| See the accompanying manuscript and versioned analysis record for model and data provenance. |
|
|
