Integrative Chemical-Biological Grouping of Complex High Production Volume Substances from Lower Olefin Manufacturing Streams

Abstract

Human cell-based test methods can be used to evaluate potential hazards of mixtures and products of petroleum refining ("unknown or variable composition, complex reaction products, or biological materials" substances, UVCBs). Analyses of bioactivity and detailed chemical characterization of petroleum UVCBs were used separately for grouping these substances; a combination of the approaches has not been undertaken. Therefore, we used a case example of representative high production volume categories of petroleum UVCBs, 25 lower olefin substances from low benzene naphtha and resin oils categories, to determine whether existing manufacturing-based category grouping can be supported. We collected two types of data: nontarget ion mobility spectrometry-mass spectrometry of both neat substances and their organic extracts and in vitro bioactivity of the organic extracts in five human cell types: umbilical vein endothelial cells and induced pluripotent stem cell-derived hepatocytes, endothelial cells, neurons, and cardiomyocytes. We found that while similarity in composition and bioactivity can be observed for some substances, existing categories are largely heterogeneous. Strong relationships between composition and bioactivity were observed, and individual constituents that determine these associations were identified. Overall, this study showed a promising approach that combines chemical composition and bioactivity data to better characterize the variability within manufacturing categories of petroleum UVCBs.

Keywords: UVCB; ion mobility spectrometry; petroleum; read-across; regulatory risk assessment.

Figure 1

Schematic of the overall experimental design depicting the chemical analytical and bioactivity profiling of neat substances and their respective DMSO extracts (n = 25).

Figure 2

(A) Feature abundances for Low Benzene Naphthas category. The top row depicts the raw abundances of features detected for neat products, product extracts, and the abundance of features characterized by the same molecular formula in neat and corresponding extract substances. The bottom row depicts the same features normalized to the total abundance of features per substance. Dark blue bars denote features present at >1%, light blue bars denote features present at 0.1–1%, and white bars denote features present at <0.1% abundance. Dotted red lines refer to ECHA’s 80% minimum threshold [14] for UVCB characterization. The third plot in each row shows features present in both the neat samples and DMSO extracts (black bars) and features unique to each (grey bars). (B) Hierarchical clustering portraying the chemical similarity of LBN neat substances based on IMS-MS profiles. Substances closer together have the most similar chemical profiles. Colors indicate pre-assigned health hazard groups.

Figure 3

(A) Feature abundances for the Resin Oils category. The top row depicts the raw abundances of features detected for neat products, product extracts, and the abundance of features characterized by the same molecular formula in neat and corresponding extract substances. The bottom row depicts the same features normalized to the total abundance of features per substance. Dark blue bars denote features present at >1%, light blue bars denote features present at 0.1–1%, and white bars denote features present at <0.1% abundance. Dotted red lines refer to ECHA’s 80% minimum threshold [14] for UVCB characterization. The third plot in each row shows features present in both the neat samples and DMSO extracts (black bars) and features unique to each (grey bars). (B) Hierarchical clustering portraying the chemical similarity of RO neat substances based on IMS-MS profiles. Substances closer together have the most similar chemical profiles. Colors indicate pre-assigned health hazard groups.

Figure 4

Total feature raw (left) and percent (right) abundance of unique molecular formulas representing typical constituents for LBN (A) and RO (B) categories, see color legend. Constituents were selected based on the substance profiles [48,49]. More detailed analysis can be found in Tables S11 and S12.

Figure 5

(A) Hierarchical clustering based on bioactivity profiles for LBN DMSO extracts. The name of each substance is colored by the prescribed health hazard group. Corresponding ToxPi diagrams depict overall substance toxicity; each slice represents one cell type, including all assessed phenotypes. Cell types tested include iPSC-derived hepatocytes (purple), endothelial cells (green), cardiomyocytes (pink), neurons (yellow), and HUVEC (blue). Larger pie slices indicate greater toxicity for that substance and cell type. (B) Hierarchical clustering based on bioactivity profiles for RO DMSO extracts. Substance names are again colored by prescribed health hazard groups, and respective ToxPi charts show an overall greater toxicity of RO substances as compared to LBN substances.

Figure 6

(A) Hierarchical clustering based on chemical profiles of neat LBN substances. Substance names are colored by prescribed health hazard groups. (B) IMS-MS chemical profiles depicted as carbon number versus double bond equivalence. Larger bubble sizes and a darker grey color depict more abundant features. Adjacent ToxPi charts show overall bioactivity across iPSC-derived hepatocytes (purple), endothelial cells (green), cardiomyocytes (pink), and neurons (yellow), as well as HUVEC (blue).

Figure 7

(A) Hierarchical clustering based on chemical profiles of neat RO substances. Substance names are colored by prescribed health hazard groups. (B) IMS-MS chemical profiles depicted as carbon number versus double bond equivalence. Larger bubble sizes and a darker grey color depict more abundant features. Adjacent ToxPi charts show overall bioactivity across iPSC-derived hepatocytes (purple), endothelial cells (green), cardiomyocytes (pink), and neurons (yellow), as well as HUVEC (blue).

Figure 8

(A,B) Each scatterplot shows bioactivity (y-axis) for the overall ToxPi score (top), iPSC-derived endothelial cells (middle), and iPSC-derived neurons (bottom) as predicted from the chemical profiles of neat (A) and corresponding extracts (B). Observed bioactivity is shown on the x-axis. Bioactivity prediction was conducted using the penalized regression approach described in Methods. The predicted values were obtained by leave-one-out cross validation, where the prediction model was developed with each sample left out of analysis, and the model applied to the features of the held-out sample. The most informative validations were chosen with the highest prediction r (Pearson coefficient) and the lowest q (false discovery rate value). (C) Correlation plot depicting the hydrocarbon compounds from neat samples that were most significantly predictive of the overall ToxPi score based on cross-validation analyses. Bubble size represents the Pearson correlation between feature abundance and ToxPi score overall as well as for individual cell types. Positive correlations are shown in blue, whereas negative correlations are shown in red. (D) Heatmap depicting the relative abundance of each feature in each sample tested. A darker color indicates higher abundance.