You are here:
Transparency in QSXR Modeling: Modernizing to a Machine Learning Paradigm (ACS 2023)
Citation:
Charest, N., C. Lowe, D. Chang, C. Ramsland, T. Martin, AND A. Williams. Transparency in QSXR Modeling: Modernizing to a Machine Learning Paradigm (ACS 2023). ACS, San Francisco, CA, August 13 - 17, 2023. https://doi.org/10.23645/epacomptox.23823828
Impact/Purpose:
N/A
Description:
The modernization of Quantitative Structure-Activity/Property Relationship modeling has been inextricably entwined with the adoption of machine learning algorithms ranging from ‘shallow learners’ such as k-Nearest Neighbors and Random Forests to ‘deep architectures’ such as feed-forward and convolutional neural networks. The Organisation of Economic Co-operation and Development prescribes guidelines for bolstering the transparency and fidelity of QSA/PR models, which can occasionally come to odds with the pervasive notion that modern algorithms are ‘black boxes’ that elude thorough interpretation. This talk examines this argument and offers discussion around how the 5 principles of the OECD guidance remain amenable to machine learning, with examples taken from literature or other research to indicate the limitations of the ‘black box’ paradigm, and how it need not stand as a barrier to interpretable models that are suitable for regulatory adoption. Discussion will include methods for interpreting common machine learning algorithms, ideas for promoting transparency in the model construction process, examples of mechanistically framing QSA/PR machine learning models, and related topics.
URLs/Downloads:
DOI: Transparency in QSXR Modeling: Modernizing to a Machine Learning Paradigm (ACS 2023)
PRESENTATION.PDF (PDF, NA pp, 1300.1 KB, about PDF)