Predicting Five Rat Acute Toxicity Endpoints with ANNE Models using ADMET Predictor™

Authors: Daga PR, Lawless M, Waldman M, Fraczkiewicz R, Clark RD, DiBella J, Bolger MB
Conference: SOT
Keywords: ADMET, ANNE, artificial neural network ensemble, QSTR, Toxicity
Software: ADMET Predictor®
Division: PBPK

Introduction

  • Alternative methods are being explored to predict the toxicity of chemicals to reduce use of animals.
    • Laboratory/Animal tests are costly in time and money
  • • Cheminformatics (QSTR) presents a good alternative to animal testing
    • Once the model is ready, predictions can be made quickly

Why Artificial Neural Network Ensemble (ANNE)?

  • Toxicity prediction is a tough problem
    • Multiple underlying mechanisms of action
    • Datasets studied (e.g., rat LD50) are large and chemically diverse
    • Multiple and wide variety of data sources
    • Simple regression methods like MLR may prove insufficient
  • Ensemble methods, such as ANNE and Random Forest, have proven to be robust enough to tackle this intensive task
  • Five endpoints were provided to model
    • Rat LD50 and “Very Toxic”, “Non Toxic”, “EPA Cat” & “GHS Cat”
    • The labels in the four end-points are dependent upon rat LD50

Society of Toxicology 54th Annual Meeting and ToxExpo, March 22-26, 2015, San Diego, California

By Pankaj R. Daga, Michael Lawless, Marvin Waldman, Robert Fraczkiewicz, Robert D. Clark, John DiBella, and Michael B. Bolger