A chemical space model for the exploration of eco-toxicological data

With over 350,000 chemicals and mixtures currently registered for production and use worldwide, around 83% of authorized chemicals lack adequate toxicity data, leaving the majority of chemicals poorly characterized. International agencies urge scientists to develop screening methods to explore, identify, and predict chemical hazards, supporting the prioritization of chemical risk assessment. Here, Tree Manifold Approximation and Projection (TMAP) were applied, with the aim of reducing the dimensionality of large toxicological dataset, providing the foundations to data imputation methods allowing to get an understanding of chemical modes of action. Specifically, TMAP was implemented using MHFP6 fingerprints and the NORMAN SusDat database, which contains over 100,000 compounds. To ensure that the TMAP layout preserves chemical structural similarity, a quantitative parameter optimization procedure was developed. The defined optimal parameter set allowed us to define the embeddings that preserves the most structural similarity among nearest connected neighbors, with similarity progressively decreasing as the distance between nodes increases. Leveraging this approach, a graph-based spatial imputation function was generated to obtain insights into the potential ecotoxicity mechanisms of data poor chemicals using physicochemical properties and CTD toxicogenomic data. The relevance and meaningfulness of TMAP chemical space was explored for Daphnia magna, Pimephales promelas and Algae. Chemical classes known to be structurally similar were found to be grouped together in the TMAP chemical space, while heterogeneous classes were found to be sparse. Data imputation allowed for the identification of known and potential chemical mechanisms of action. Indeed, acetylcholinesterase and transthyretin were confirmed as major mechanisms of action of organothiophosphate and brominated flame retardant toxicity in Daphnia magna and Pimephales promelas, respectively. Overall, transdisciplinary toxicological databases combined with TMAP, stand out as a computationally efficient and suitable method to explore and analyze large datasets, allowing for the inference of associations between chemical structures and chemical hazard identification and other potential applications of this hypotheses-generating tool.