CoolFace
Apppublic

Te-nok/QT-M2M4-GNNens-demo

sourceHugging Faceupdated 5mo agoView on Hugging Face
0likes
App README

QT-M2M4-GNNens-demo

This Space provides a research-use demo for the QT-M2M4-GNNens workflow associated with the manuscript:

Integrated Molecular Graph Modeling of hERG Assay and Pharmacovigilance Signals for Early Cardiac Safety Prioritization

The demo supports exploratory, pharmacovigilance-informed cardiac safety prioritization using an ensemble of molecular graph models. The output should be interpreted as a ranking or triage score for further safety follow-up, not as a definitive prediction of clinical QT prolongation, torsade de pointes, arrhythmia risk, or regulatory safety status.

Overview

The demo accepts a compound as either a PubChem CID or a SMILES string and returns ensemble-based QT-liability prioritization outputs derived from two model families, M2 and M4.

The QT-M2M4-GNNens workflow integrates public hERG assay-derived information and pharmacovigilance-derived signal information to support early prioritization of compounds for cardiac safety follow-up.

The pharmacovigilance-derived component should be interpreted as signal-informed weak labeling. It should not be interpreted as evidence of causal drug-event relationships.

Intended Use

This Space is intended for:

  • —exploratory cardiac safety prioritization;
  • —research-use comparison of candidate compounds by predicted QT-liability score;
  • —methodological demonstration of the QT-M2M4-GNNens workflow;
  • —educational inspection of ensemble-based molecular graph model outputs;
  • —reproducibility support for the associated manuscript.

Not Intended Use

This Space is not intended to replace:

  • —experimental hERG testing;
  • —clinical ECG evaluation;
  • —exposure-response assessment;
  • —nonclinical cardiac safety evaluation;
  • —expert pharmacovigilance review;
  • —regulatory safety assessment;
  • —clinical decision-making.

The model should not be used as a standalone tool to determine whether a drug or compound is clinically safe or unsafe.

Inputs

  • —PubChem CID
  • —SMILES

Outputs

  • —Used SMILES
  • —Canonical SMILES
  • —InChIKey
  • —p_M2
  • —p_M4
  • —alpha_M2
  • —alpha_M4
  • —p_ens
  • —Tier

Interpretation of the Output

  • —p_M2: predicted QT-liability prioritization score from the M2 model family
  • —p_M4: predicted QT-liability prioritization score from the M4 model family
  • —alpha_M2: ensemble weight assigned to the M2 family
  • —alpha_M4: ensemble weight assigned to the M4 family
  • —p_ens: final ensemble-based QT-liability prioritization score
  • —Tier: rule-based prioritization category derived from the ensemble score

The values should be interpreted as model-derived prioritization outputs. They are not calibrated clinical risk probabilities.

Tier Definition

  • —GREEN: lower-priority QT-liability signal
  • —YELLOW: intermediate-priority QT-liability signal
  • —RED: higher-priority QT-liability signal

The tier labels are intended only for prioritization and should not be interpreted as clinical diagnostic categories or regulatory classifications.

Usage

  1. 1.Enter either a PubChem CID or a SMILES string.
  2. 2.Click Run prediction.
  3. 3.Review the returned ensemble scores and prioritization tier.
  4. 4.Interpret the result as a research-use prioritization output requiring further experimental, clinical, or pharmacovigilance review.

Example SMILES:

  • —CCO
  • —c1ccccc1

Important Limitations

Important limitations include:

  • —pharmacovigilance-derived labels do not establish causal drug-event relationships;
  • —spontaneous reporting systems are subject to reporting bias, notoriety bias, confounding by indication, co-medication effects, and missing denominator information;
  • —hERG assay-derived labels do not fully capture clinical QT prolongation or torsade de pointes risk;
  • —model performance may decrease for compounds outside the chemical domain represented in the training data;
  • —predicted scores are not calibrated clinical risk probabilities;
  • —the model does not account for dose, exposure, metabolism, drug-drug interactions, patient-level risk factors, electrolyte abnormalities, renal or hepatic impairment, or ECG context.

Notes

  • —Research use only.
  • —Not for clinical decision-making.
  • —Not for regulatory decision-making.
  • —Not a replacement for experimental hERG testing or clinical ECG assessment.
  • —This demo is intended for methodological, educational, and reproducibility-support purposes.
  • —Model assets are loaded at runtime from a private Hugging Face model repository via Space secrets.

Code and Reproducibility

The associated code, preprocessing scripts, and reproducibility materials are available in the companion GitHub repository:

https://github.com/tenokiya/ChemoinfoQTgnnensmanuscript_submission

Technical Note

This Space uses a Docker-based deployment and loads the trained model assets at runtime from a private Hugging Face model repository configured through Space secrets.

Because the model assets are loaded through Space secrets, users do not need direct access to the private model repository to run the public demo, provided that the Space is correctly configured and running.

Citation

If you use this demo or the associated model outputs, please cite the associated manuscript once published:

Enokiya T, Yamaguchi T. Integrated Molecular Graph Modeling of hERG Assay and Pharmacovigilance Signals for Early Cardiac Safety Prioritization. Submitted to Drug Safety.