README.md

---
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.