Quantitative predictions from chemical read-across and their confidence measures

Abstract

In silico modeling new approach methodologies (NAMs) are viewed as a promising starting point for filling the existing gaps in safety and ecosafety data. Read-across is one of the most widely used alternative tools for hazard assessment, aimed at filling data gaps. However, there are no systematic studies or recommendations on the measures to identify the quality of read-across predictions for the data points without any experimental response data. Recently, we have reported a new similarity-based read-across algorithm for the prediction of toxicity (biological activity in general) of untested compounds from structural analogues (the tool available from https://sites.google.com/jadavpuruniversity.in/dtc-lab-software/home). Three similarity estimation techniques such as, Euclidean distance-based similarity, Gaussian kernel function similarity, and Laplacian kernel function similarity are used in this algorithm. As the confidence of predictions for untested compounds is important information, we have addressed this issue here by consideration of several similarity and error – based criteria. The role of these measures in discriminating high and low residual query compounds is studied in three different approaches: (a) comparison of means of a measure for high and low residual groups; (b) development of classification models for absolute residuals to identify the contributing measures; (c) application of the sum of ranking differences (SRD) approach to identify the measures closer to the reference rank defined by the absolute residuals. Finally, the frequency of occurrences of different measures in the three approaches is compared. The results from three data sets with 10 divisions of source and target compounds in each case indicate that weighted standard deviation of the predicted response values appear to be the most deterministic feature for the reliability of predictions followed by different similarity-based features. The derived reliability measures will provide a greater confidence to the quality of quantitative predictions from the chemical read-across tool for new query compounds.