MLTox

Enhancing Toxicological Testing through Machine Learning

Abstract

We plan to use machine learning (ML) methods to predict the effects of chemicals on aquatic species.

Our main goal is to ​ use a combination of data from in-vivo (whole organisms) and in-vitro (cell culture) experiments to infer the effects of chemicals on organisms for which no testing data is available (both for the chemical and for the organism).

In the literature, this kind of problem is also known as across-chemical (and across-species) extrapolation. Usually, extrapolation across chemicals is performed using measures of chemical similarity under the assumption that similar chemicals will be similarly toxic to the same species. Extrapolation across species can be performed based on measured chemical effects on some species and the similarity between species, either by phylogenetic distance or sequence/structure similarity of known molecular targets of the chemicals, if at all available, or through similarity in physiological traits.

Given the enormous number of chemicals and potentially affected species, extrapolation chemical-by-chemical or species-by-species is a daunting task.

 

Started

September 2021

ONGOING

PI / Partners

Department Systems Analysis, Integrated Assessment and Modelling (EAWAG)

  • Prof. Dr. Kristin Schirmer
  • Dr. Marco Baity Jesi
  • Dr. Christoph Schür

Description

Problem:

Ecotoxicological testing requires investing large amounts of money, workforce, and time, in addition to the animal suffering for in-vivo tests. There are global efforts to reduce or replace animal testing for both ethical and feasibility concerns for human and environmental risk assessment. Indeed, a ​ paradigm shift is needed to ensure a toxic-free environment as proposed, e.g. in the EU’s Green Deal.

Proposed approach:

  • In WP1, we will train standard ML models on fish data and will compare them to more elaborate models.

  • In WP2, we will analyze how much and under which conditions the usage of in-vitro data can improve the predictions of our ML models.

  • In ​ WP3​ , we will use our models to gain a better understanding of the nonlinear relationships that connect species, chemicals, and related toxicity.

  • In WP4​ , we will explore methods for improving the performance of our models and we will release an open-source package with our models.

Impact:

With the work proposed here, we will provide new means to protect the environment from toxicants, which allow us to significantly ​ reduce or even replace experiments on animals, by combining ML, in-vitro tests, and pre-existing in-vivo data.

Publications

Bibliography

Luechtefeld et al. (2018) Machine Learning of Toxicological Big Data Enables Read-Across Structure Activity Relationships (RASAR) Outperforming Animal Test Reproducibility. Toxicological Sciences, Volume 165, Issue 1, September 2018, Pages 198–212 Machine Learning of Toxicological Big Data Enables Read-Across Structure Activity Relationships (RASAR) Outperforming Animal Test Reproducibility