Validated Machine-learning approaches: How clustering permitted rapid expert validation of big data to create the iSafeRat® High Accuracy biodegradation model

Persistency is one of the 3 (soon to be 4) key environmental properties of concern and may be considered the most critical one upstream, because rapidly biodegradable substances typically shouldn’t remain sufficiently long in the environment to be able to build up to levels where they would cause extensive environmental harm. 

A set of standard methods were created decades ago to assess ready biodegradability (OECD 301 series and later 310) however these tests have very stringent criteria are often conservative in their predictivity of environmental persistency. Many of these studies have resulted in negative or contradictory outcomes which are only partly due to poor decisions which method to use or to spatiotemporal variability of the inoculum in terms of structural and functional diversity. In silico methods used for regulatory purposes to date have mostly been based on structural fragment contribution and ultimately are not completely applicable when multiple fragments are found collectively in the same molecule, or to KNN approach where the prediction depends to the relevance of the nearest neighbour. KREATiS has recently built its own model based on a machine-learning approach using experimental data of more than 2000 compounds. Also, the dataset has been manually and carefully curated using a two steps approach. First, only experimental studies compliant to OECD301/310 Guidelines were kept. Second a clustering approach allowed assessment of whether similar compounds achieved similar results. During that last step, over 600 clusters of compounds were expert validated in-depth, highlighting, and allowing us to eliminate potential experimental fails that would finally only lead to a wrong conclusion, inducing noise and error in the model. Also, two types of splitting and modelling methods were explored in order to identify the configuration of the model leading to the best performance. The model is currently in beta testing and will shortly be available for ecodesign and regulatory cases alike.

Speakers: Floriane Lauras and Paul Thomas (KREATiS) 

Floriane LARRAS has a PhD in aquatic ecotoxicology and investigated over 10 years the impacts of micropollutants to freshwater organisms using taxonomical, life-history traits and omics-based endpoints.  In 2016, she obtained a Marie Curie grant to explore the molecular mechanisms involved in microbial community responses to pesticides leading to the development of a tool managing omics dose-response curves for ecological risk assessment. She joined KREATiS in 2022 to develop ecotoxicological QSAR models,  prepare prediction documents for client regulatory submission and manage project research to improve in silico approaches for substances like cosmetics, pharmaceuticals or natural complex substances.

Paul Thomas has a Ph.D. in aquatic ecotoxicology and more than 20 years of experience with industrial chemicals, agrochemicals, and biocides gained initially at CIT (now Charles River) where he set up the ecotoxicology service and was study director for 4 years, ATOFINA (now ARKEMA) where he practiced ecotoxicology and Risk Assessment for existing substances and then AkzoNobel (now Nouryon) where he was head of the department, head of the ecotoxicology GLP and research laboratories and organiser of REACH services for the company for over 6 years. He joined CEHTRA in 2008 as director of CEHTRA Lyon specialising in REACH-related services and is the manager of the CEHTRA ecotoxicology team and contributed strongly to numerous successful registrations out of the >600 substances that CEHTRA registered since 2010 and has been working in the area of predictive ecotoxicology since this time. In 2014, while retaining all his roles in CEHTRA, Paul founded KREATiS and was nominated CEO KREATiS is a start-up company specialising in the creation of high accuracy QSARs to replace experimentation for REACH and other regulations. In 2020 he bought the company together with a human health expert, Dr. Carole Charmeau, and they have worked hard with the KREATiS team to bring the company to new heights, adding multiple models to their iSafeRat® model suite, and embracing new technologies such as 3D docking models for studying protein-ligand interactions.