Scoping when bit length of molecular fingerprints determines the reliability of computational tools in environmental chemistry and toxicology

Many modern computational tools to describe or predict the behaviour of organic chemicals rely on molecular fingerprints. Fingerprints are used as descriptors for structure-activity relationship modelling and structure-based clustering of chemicals across the fields of environmental chemistry, technology and (eco)toxicology. Here, we highlight how Extended Connectivity Fingerprint (ECFP), RDKit fingerprint, Atom Pair fingerprint, and Torsion fingerprint, among others, are susceptible to "bit collisions", which result in the same fingerprint for different molecular structures in a given dataset. Bit collisions can be introduced during hashing, where two distinct fragments result in the same hash, or more commonly, where the folding operation assigns two different structures to the same bit index. These collisions can occur at the molecular level, where two distinct substructures in a molecule collide, or at the dataset level, where collisions occur due to molecular aggregation. Using two curated datasets (pesticide degradation and global chemical inventory data), we show that bit collision is inversely related to the bit size of fingerprints and is more pronounced in Atom Pair and RDKit fingerprints compared to ECFP and Torsion fingerprints. In addition, molecular size and atom count are positively related to bit collisions. At the widely used ECFP4 with a bit length of 1024, bit collisions occurred for approximately 50% of the molecules in both tested datasets. Finally, we showcase how bit collisions can be detrimental for predictive modelling of targets with high certainty, such as hydrophobicity, while their effect is less visible for more uncertain and "noisy" targets such as persistence. Overall, this study provides guidance on fingerprint and bit size selection for predictive modelling in environmental chemistry, ecotoxicology and environmental technology.