Confidently Uncertain: Probabilistic Machine Learning to Predict Soil Biotransformation Half-Lives

Predicting environmental persistence of chemicals from
molecular
structure is an open challenge, yet indispensable in regulatory screenings
for potentially harmful substances and to advance the development
of safe-and-sustainable-by-design chemicals. Limited availability
of biotransformation half-life data makes persistence prediction difficult,
and models typically struggle to generalize beyond their training
data. Therefore, reliable estimates of prediction confidence are key.
Here, we propose a probabilistic model for the prediction of soil
biotransformation half-lives. A Gaussian Process Regressor was trained
on 867 mean pesticide half-lives with data uncertainty estimates.
Instead of single half-life values, our model predicts well-calibrated
probability distributions that can be used to calculate a compound’s
probability of being persistent. Although the overall model performance
remains moderate, the predictions are reliable when the confidence
in the prediction is high. We applied our model to pesticide transformation
products with unknown half-lives, and to a database of globally marketed
chemicals. We show that our model is able to identify chemicals that
are known, or suspected to be, persistent in the environment. The
model is available as an online app (https://pepper-app.streamlit.app/) and as a Python library (pepper-lab) to meet diverse user needs.