Machine Learning Demonstrates Dominance of Physical Characteristics over Particle Composition in Coal Dust Toxicity

Abstract

Mine dust has been linked to the development of pneumoconiotic diseases such as silicosis and coal workers' pneumoconiosis. Currently, it is understood that the physicochemical and mineralogical characteristics drive the toxic nature of dust particles; however, it remains unclear which parameter(s) account for the differential toxicity of coal dust. This study aims to address this issue by demonstrating the use of the partial least squares regression (PLSR) machine learning approach to compare the influence of D₅₀ sub 10 μm coal particle characteristics against markers of cellular damage. The resulting analysis of 72 particle characteristics against cytotoxicity and lipid peroxidation reflects the power of PLSR as a tool to elucidate complex particle-cell relationships. By comparing the relative influence of each characteristic within the model, the results reflect that physical characteristics such as shape and particle roughness may have a greater impact on cytotoxicity and lipid peroxidation than composition-based parameters. These results present the first multivariate assessment of a broad-spectrum data set of coal dust characteristics using latent structures to assess the relative influence of particle characteristics on cellular damage.

Keywords: coal mine dust; cytotoxicity; oxidative stress; partial least squares regression; particle-cell relationships.

Figure 1

Summary of the major and minor compositional characteristics of the coal dust particles as determined by QEMSCAN analysis. Major mineral constituents were identified where CM = carbonaceous matter, Clys = clays, Qz = quartz (SiO2), Py = pyrite (FeS2), Cal = calcite (CaCO₃). In the context of the clays and sulfates, these categories comprise kaolinite [Al₂Si₂O₅(OH)₄] and Illite [K_0.6(H₃O)_0.4Al_1.3Mg_0.3Fe_0.1²⁺Si_3.5O₁₀(OH)₂·(H₂O)], gypsum [Ca(SO₄)·2(H₂O)], szomolnokite [Fe²⁺(SO₄)·(H₂O)], jarosite [KFe³⁺(SO₄)₂(OH)₆], and rhomboclase [HFe³⁺(SO₄)₂·4(H₂O)] respectively. (b) Minor mineral distribution among samples, where FeOx, Rt, Dol, and Sd refer to iron oxide minerals, rutile (TiO₂), dolomite [CaMg(CO₃)₂] and siderite [Fe²⁺(CO₃)] respectively. For reference, the minerals classed as iron oxides were hematite (Fe₂O₃) and goethite [Fe³⁺O(OH)]. (c) Major element concentrations measured using XRF spectrometry and calculated as the wt % element analyzed. Carbon and oxygen represent the remainder as loss on ignition. (d) Distribution of the elements Si, Fe, and Ca, among the mineral hosts identified. The data are represented as a percentage relative to the total amount of Si, Fe or Ca shown in Figure 1c.

Figure 2

Differential expression of cytotoxicity and lipid peroxidation among coal dust exposed THP-1 cells (72 h time point). (a–c) represent the samples that have been subdivided into high, medium, and low cytotoxic responses, respectively, based on clusters observed at the 350 μg/mL concentration and the gradient of samples in these clusters. For reference, the relative error reported at each concentration was less than 3.5% across the samples analyzed. (d) displays the distribution of MDA produced by the THP-1 cells from a 72 h exposure with 350 μg/mL of the particle samples. Based on the results three groups of samples could be identified, namely low, medium, and high MDA-release samples. (e) further classes the samples based on their measured cytotoxic response. The amount of MDA released relative to the negative control was normalized by the percentage of viable cells based on the results of the cytotoxicity assay. For reference, the relative error across all measurements was less than 0.2 units.

Figure 3

Loading plots representing the particle parameters that are either positively or negatively correlated, as well as the parameters correlated with MDA or LDH releases. To interpret the loading plot, parameters that cluster together display an association suggesting they behave similarly, while parameters that oppose each other show opposing effects in the model system. The particle parameters are represented by the red, orange, and gray dots, and the two responses are indicated by blue triangles. (a,b, c,d and e,f) represent pairs of loading plots categorized by mineral and element-based data, mineral-specific parameters, and general particle characteristics, respectively. The pairs contrast the two responses where (a,c,d) related to the MDA-based model and (b,e,f) relate to the LDH-based model. For (a,b) the minerals have the XRD and QS, this indicates that the value was derived from XRD and QEMSCAN analysis, respectively. For C and D the liberation state and CRY data for the minerals quartz = qz, pyrite = py, and clays = clys (kaolinite = kln) are abbreviated as follows: Lib = liberation, MLib = moderately liberated, MEcap = mostly encapsulated, FEncap = fully encapsulated, and CRY = crystallite size. VIP scores were used to class the characteristics by importance to the model, where the influence of each variable was categorized as highly (VIP > 1.25), moderately (1 < VIP < 1.25), or less influential (VIP < 1). The total variance explained by components 1 and 2 across all variables in the MDA model were comp1 = 21.86%, comp2 = 15.09%. The total variance explained by components 1 and 2 across all variables in the LDH model comp1 = 19.94%, comp2 = 12.39%.

Figure 4

Coefficient plot comparing the relative impact of the particle characteristics based on their coefficient value against the measure of relevance computed for each parameter in relation to the model (based on VIP scores). A and B represent the coefficients plots for the MDA and LDH-based models, respectively. To interpret the coefficients plot, the x-axis represents the model coefficient in the form Y = BX, where B is the coefficient and is calculated for each explanatory variable “X” included. The magnitude of the coefficient gives an indication of the relative importance of the parameter, and the direction (positive or negative) yields whether the parameter either promotes or depresses the response. The Y-axis represents the VIP score calculated for each parameter. This stratifies which parameters played a significant role in the model. A VIP score >1 was used as the threshold for significance.