EESS talk on "Machine learning as tool to predict the toxicity of chemicals"
Event details
| Date | 14.12.2021 |
| Hour | 12:15 › 13:15 |
| Speaker | Dr Marco Baity Jesi, Scientist, Department Systems Analysis, Integrated Assessment and Modelling, EAWAG |
| Location | Online |
| Category | Conferences - Seminars |
| Event Language | English |
Abstract:
We apply machine learning methods to predict chemical hazard focusing on fish acute toxicity across taxa. We analyze the relevance of taxonomy and experimental setup, and show taking them into account can lead to large improvements in the classification performance. We quantify the gain obtained by introducing the taxonomic and experimental information, compared to classifying based on chemical information alone. Among the identified relevant features, the species is the single most important one, surpassing any single chemical descriptor. We use our approach with standard machine learning models. We are able to obtain accuracies of over 93% on datasets where, due to noise in the data, the maximum achievable accuracy is expected to be below 95%. Most of our models “outperform animal test reproducibility” as measured through recently proposed metrics, and the best performances are obtained by random forests and deep neural networks. However, we analyze the metrics that lead to such kinds of statements, and show that the comparison between machine learning performance and animal test reproducibility should be addressed with particular care.
Short biography:
Dr Marco Baity-Jesi develops machine-learning algorithms and uses data-oriented methods to study aquatic systems. After a PhD in theoretical physics in cotutorship between Rome and Madrid, he spent two years as a postdoc in Paris, working at École Normale, and two years in New York, at Columbia University. Since 2019 he is based in Zurich, as a Group Leader in machine learning at Eawag.
We apply machine learning methods to predict chemical hazard focusing on fish acute toxicity across taxa. We analyze the relevance of taxonomy and experimental setup, and show taking them into account can lead to large improvements in the classification performance. We quantify the gain obtained by introducing the taxonomic and experimental information, compared to classifying based on chemical information alone. Among the identified relevant features, the species is the single most important one, surpassing any single chemical descriptor. We use our approach with standard machine learning models. We are able to obtain accuracies of over 93% on datasets where, due to noise in the data, the maximum achievable accuracy is expected to be below 95%. Most of our models “outperform animal test reproducibility” as measured through recently proposed metrics, and the best performances are obtained by random forests and deep neural networks. However, we analyze the metrics that lead to such kinds of statements, and show that the comparison between machine learning performance and animal test reproducibility should be addressed with particular care.
Short biography:
Dr Marco Baity-Jesi develops machine-learning algorithms and uses data-oriented methods to study aquatic systems. After a PhD in theoretical physics in cotutorship between Rome and Madrid, he spent two years as a postdoc in Paris, working at École Normale, and two years in New York, at Columbia University. Since 2019 he is based in Zurich, as a Group Leader in machine learning at Eawag.
Practical information
- General public
- Free
- This event is internal
Organizer
- EESS - IIE
Contact
- Prof. Janet Hering, EAWAG Director