New approach to regression uncertainty analysis and applications to drug design

Authors: Waldman M, Clark RD
Conference: ACS
Software: ADMET Predictor®
Division: PBPK

Motivation

  • When can you trust a decision/prediction from a machine learning model?
    • Many examples of machine learning/AI failures (just Google “recent AI failures”)
  • What is the “expected” accuracy of a quantitative prediction?
    • Drug candidate with predicted low solubility
      • “Distrust” the model – expected accuracy is poor – large prediction uncertainty
      • Synthesize anyway and measure the solubility
  • “Trust” the model – expected accuracy is good – small prediction uncertainty
    • Move on to another compound – don’t bother to synthesize

By Marvin Waldman and Robert Clark

ACS National Meeting and Expo in San Diego, CA August 25-29, 2019

Download Slides