Assessment of chemical-toxicological relationships in atmospheric aerosols
Daniel O'Sullivan, DAVID HEALY, John Sodeau, Stig Hellebust
University College Cork
Abstract Number: 834
Working Group: Health Related Aerosols
Last modified: July 25, 2011
Human health effects of atmospheric particles may depend on many factors including size, physical state and chemical composition.
The main aim of this study was to link chemical composition of airborne particulate matter (PM2.5) present at different locations in a large natural Harbour and city region with measured toxicological endpoints. Towards this end a model was developed to relate the toxicity profile of ambient PM to its chemical composition. This approach provides an attractive possibility of quickly assessing potential health effects of ambient PM. The model was applied to a readily available bank of “real” ambient samples, collected in Cork Harbour, Ireland, over a twelve month period between April 2007 and April 2008. The samples were analysed for chemical composition and in-vitro toxicity. Five main tests for toxicity were explored: (i) oxidative stress/ROS; (ii) DNA damage/genotoxicity; (iii) cytotoxicity; (iv) inflammatory potential; (v) endotoxin content.
Statistical analysis of the toxicological data was conducted using 2-factor ANOVA (with replicates), in order to quantify the main effects responsible for variations within the datasets (e.g. whether responses are mainly stimulated by changing doses or exposure times.
Chemico-toxicological relationships were explored with principal component analysis and partial least squares regression modelling. Using these dimension reducing techniques, >80% of the variance within the chemical dataset was apportioned to a few principle components, each of which could be characterised by differing source profiles. The possibility that information expressed in the chemical speciation of unknown samples can be used to forecast toxicological responses was explored.
The model, which demonstrated 100% selectivity and sensitivity in assigning the chemical training set samples into one of two classes based on the ROS responses (“High” or “Low”) was estimated by cross-validation to have a predicted error rate of 16% if generalised to an independent dataset.