Characterization and prediction of chemical functions and weight fractions in consumer products

Abstract

Assessing exposures from the thousands of chemicals in commerce requires quantitative information on the chemical constituents of consumer products. Unfortunately, gaps in available composition data prevent assessment of exposure to chemicals in many products. Here we propose filling these gaps via consideration of chemical functional role. We obtained function information for thousands of chemicals from public sources and used a clustering algorithm to assign chemicals into 35 harmonized function categories (e.g., plasticizers, antimicrobials, solvents). We combined these functions with weight fraction data for 4115 personal care products (PCPs) to characterize the composition of 66 different product categories (e.g., shampoos). We analyzed the combined weight fraction/function dataset using machine learning techniques to develop quantitative structure property relationship (QSPR) classifier models for 22 functions and for weight fraction, based on chemical-specific descriptors (including chemical properties). We applied these classifier models to a library of 10196 data-poor chemicals. Our predictions of chemical function and composition will inform exposure-based screening of chemicals in PCPs for combination with hazard data in risk-based evaluation frameworks. As new information becomes available, this approach can be applied to other classes of products and the chemicals they contain in order to provide essential consumer product data for use in exposure-based chemical prioritization.

Keywords: Chemical function; Chemical prioritization; Consumer products; Cosmetics; ExpoCast; Exposure modeling.

Fig. 1

Workflow for using existing chemical function and weight fraction datasets to build empirical compositions and QSPR classification models for chemical function and weight fraction for use in estimation of chemical exposure.

Fig. 2

Median product weight fractions (WF) associated with 35 harmonized chemical functions in 66 personal care product categories. Heatmap is clustered on both axes to group products/functions that are similar in terms of WF.

Fig. 3

Function and weight fraction bin predictions for chemicals having identified personal care product use (A; N = 538) and function predictions for chemicals with unknown use (B; N = 9658). Chemical observations are on the vertical axis; the two datasets are scaled differently on this axis for size. Chemicals having a high probability are shown in dark blue. Weight fraction predictions (indicated by the colorbar to the right of the main figure) are reported when a function could be predicted with >90% probability. Only functions are predicted for the unknown use chemicals since they may or may not be present in consumer products.