Chapter 22: Multiple Test In Silico Weight-of-Evidence for Toxicological Endpoints Check Access

New European legislation in chemical safety assessment (REACH) requires combining chemical and toxicological information from a variety of sources: chemical structure information, similarity with other chemicals (read-across), in vitro and in vivo laboratory test data, and computerised information and knowledge, i.e. in silico methods. Integrating such diverse sources of compound-related knowledge requires quantitative methods to merge multiple lines of evidence. Prediction of a toxicological property (endpoint) of a chemical from test information is analogous to medical diagnosis. As test predictions may fail, a probabilistic approach is imperative. Through Bayesian statistics, the probability of an endpoint can be estimated from test data on the basis of Diagnostic Likelihood Ratios (DLRs) involving test performance measures to produce correct predictions. An informative measure of Weight-of-Evidence (WoE) can be based on the logarithm of the DLR, which depends on the actual test result(s). A convenient unit of WoE, the deciban, was developed by Turing in the 1940s. Its usefulness for quantifying single test and multiple test (battery) information is demonstrated on an example dataset. Multiple joint test results for given endpoints are scarce, so it is important to assess the influence of small sample size. This can be done with binomial logistic regression, which permits candidate models to be compared. Model selection criteria point to models that optimize in approximating the data, while avoiding overfitting. This is important to reduce predictive uncertainty. The WoE measure developed here can be shown to be a simple linear function of the model parameters describing the joint test data.