Chemical hazard and risk assessments often use physical-chemical properties to categorize and identify chemicals of concern. In Europe, recent legislation regarding the classification, labelling and packaging of consumer products includes new chemical hazard categories, including persistent, mobile, and toxic (PMT) and very persistent and very mobile (vPvM). While persistency and toxicity are currently used in chemical regulation, mobility is a new criterion for chemical hazard classification. As a result, the mobility of thousands of chemicals on the global market must be assessed. Mobility is determined based on the organic carbon–water partition ratio (K_OC) or octanol–water partition ratio (K_OW) for neutral compounds and the octanol–water distribution ratio (D_OW) for ionizable compounds. However, measured K_OC values can vary by orders of magnitude depending on the environmental conditions particularly for ionizable compounds.

Ideally such property data is available in machine-readable formats in standardized units with meta data regarding the environmental conditions of the measurement. However, current regulatory databases and experimental datasets have thousands of K_OC data values of diverging quality available, and the environmental conditions associated with this data are not easily accessible. Data must be cross-referenced to ensure the name and CAS number associated with a K_OC value matches the SMILES used in prediction models. The data must be tidied up to remove duplicated entries and incorrectly reported units.

In this work, we use all available K_OC data for chemicals on the global market in a consensus-based approach, which considers the number of K_OC values and their individual variability and uncertainty, and source of K_OC data. Data sources include the OECD's eChemPortal and QSAR Toolbox, experimental log₁₀ K_OC datasets, and multiple K_OC prediction models of varying quality and domains of applicability.

With Bayesian statistical inference approaches we aggregate values and their errors from these different sources to derive a probability distribution for the "true" log₁₀ K_OC value. The use of probability distributions is advantageous in these instances because we are less concerned about the environmental conditions for which the K_OC is measured and can apply mobility classifications based on relative risk tolerance of the uncertainty for a given K_OC value.