Predicting low dose effects for chemicals in high through-put studies.

High through-put studies commonly use automated systems with 96-well plates in which multiple chemicals are tested at multiple doses using log-2 dose increments after a suitable incubation period. There are typically multiple (ranging from five to eleven) doses on each chemical, and occasionally plate replications of the dose-response studies. The target endpoint for such studies is typically the LC50, but for some chemicals, there may be multiple doses below a benchmark dose where there is no apparent adverse response relative to control response. We show how an estimation approach can lead to clearly interpretable results about response in the low dose region using data from a high throughput study of 2189 chemicals on yeast. Accurate estimates can be obtained of response for study chemicals by using best linear unbiased predictors (BLUPs) in a mixed model, and summarized via plots with expected response (assuming no low-dose effect) with confidence intervals for response below the benchmark dose for each chemical, providing an informative summary of response at low doses. We conclude that this approach can provide valuable insights that would be missed if the observational data were only considered through the lens of statistical methods appropriate for experimental studies.