Predicting PPARγ Potency: Consensus Model & Deep Learning

How NovaMechanics combined neural network binding affinity prediction with ensemble machine learning classification to build OECD-validated models for PPARγ antagonist activity — applied to screen PFAS compounds for endocrine disruption.

Molecular Diversity • 2025
View Publication

The Challenge

Identifying PPARγ Modulators for Drug Discovery & Chemical Safety

PPARγ is a nuclear receptor central to glucose and lipid metabolism, making it a prime target for antidiabetic therapies. Identifying potent antagonists computationally is essential for both drug discovery and assessing environmental chemical risks, particularly for per- and polyfluoroalkyl substances (PFAS) that may disrupt endocrine function.

Existing computational approaches often treat binding affinity and biological activity separately, and many lack the regulatory validation needed for real-world decision-making in chemical safety assessment.

6,587
Compounds from Tox21 bioassay dataset
34
PFAS compounds screened
OECD
Regulatory-compliant validation

Our Approach

A synergistic dual-model strategy combining deep learning and consensus classification

Curate bioassay data and perform molecular docking

Gathered 6,587 compounds from PubChem Tox21 bioassay data and performed molecular docking against the PPARγ ligand-binding domain using the Enalos Asclepios KNIME pipeline to generate binding affinity scores.

Build deep learning binding affinity model

Trained a fully connected neural network on Extended Connectivity Fingerprints (ECFPs) to classify compounds as strong or weak binders based on molecular docking scores, providing a first-pass screen for PPARγ interaction.

Develop consensus classification for antagonism

Built three base classifiers — Random Forest, Support Vector Machine, and k-Nearest Neighbours — to predict antagonistic activity, then integrated them into an ensemble bagging strategy for improved accuracy.

Validate models to OECD guidelines

Rigorously validated both models according to OECD principles, ensuring generalisability and sufficient efficiency in detecting the minority class of active antagonists through cross-validation and external test sets.

Screen PFAS and deploy on Enalos Cloud

Applied the validated workflows to screen 34 commercially available PFAS compounds, forecasting their binding strength and antagonistic activity against PPARγ. Models deployed as interactive web applications on the Enalos Cloud Platform.

Results at a Glance

DL + ML
Synergistic Strategy
Neural network binding + ensemble classification in a unified workflow
RF/SVM/kNN
Consensus Ensemble
Three diverse classifiers combined via bagging for robust antagonism prediction
PFAS
Safety Screening
34 commercially available PFAS compounds assessed for PPARγ disruption
OECD
Regulatory Ready
Both models validated per OECD QSAR principles with defined applicability domains
FAIR
Open Access
Data and models shared through interactive Enalos Cloud web applications
ECFP
Molecular Fingerprints
Extended Connectivity Fingerprints capture structural features for binding prediction

Related Publication

Peer-Reviewed Paper

Predicting peroxisome proliferator-activated receptor gamma potency of small molecules: a synergistic consensus model and deep learning binding affinity approach powered by Enalos Cloud Platform

Antoniou M., Papavasileiou K.D., Tsoumanis A., Melagraki G., Afantitis A. — Molecular Diversity, 2025 — DOI: 10.1007/s11030-025-11230-6