Umeå University's logo

Early warning systems (EWSs) are currently being developed by various authorities aiming at identifying potentially hazardous chemicals before they become a threat to the environment and human health. In this context, patents provide an excellent data source for exploring novel chemistry or the use of chemicals in materials and products. However, analysis of patents is challenging, including unraveling molecular structures presented as graphics depicting various elements, functional groups, and molecular bonds. Our study aims to improve EWS using automated artificial intelligence-based molecular structure recognition methods for encoding these for further hazard analysis. Current structure extraction tools are primarily trained on chemical structures collected from publicly available data sets, and the application of these tools to patent-specific chemical data has received little attention. This paper presents a field study utilizing the three tools Decimer, Molscribe, and Mathpix and assesses their performance in recognizing chemical structures in patents. Two data sets were compiled and curated including (1) diverse organic chemicals and (2) per- and polyfluoroalkyl substances (PFAS). It was revealed that these tools perform well on simpler molecular structures, whereas they struggle with more complex structural features, including repetitive units, cross-bonding, and Markush structures. Furthermore, it was discovered that these tools are extremely sensitive to image artifacts such as noise from lines and dots or distortions. Overcoming these challenges will be critical before implementation in automated EWS and thereby enable screening of patents for rapid and effective identification of potentially hazardous emerging chemicals.

Innovative tools suitable for chemical risk assessment are being developed in numerous domains, such as non-target chemical analysis, omics, and computational approaches. These methods will also be critical components in an efficient early warning system (EWS) for the identification of potentially hazardous chemicals. Much knowledge is missing for current use chemicals and thus computational methodologies complemented with fast screening techniques will be critical. This paper reviews current computational tools, emphasizing those that are accessible and suitable for the screening of new and emerging risk chemicals (NERCs). The initial step in a computational EWS is an automatic and systematic search for NERCs in literature and database sources including grey literature, patents, experimental data, and various inventories. This step aims at reaching curated molecular structure data along with existing exposure and hazard data. Next, a parallel assessment of exposure and effects will be performed, which will input information into the weighting of an overall hazard score and, finally, the identification of a potential NERC. Several challenges are identified and discussed, such as the integration and scoring of several types of hazard data, ranging from chemical fate and distribution to subtle impacts in specific species and tissues. To conclude, there are many computational systems, and these can be used as a basis for an integrated computational EWS workflow that identifies NERCs automatically.