High Throughput Screening (HTS) is an effective means to determine the chemical compounds which are efficient on a given biological target. The experiments are carried out using a standard format, the 96-well plate, where six wells are controls whose expected values are known. However the measurement techniques are subject to variation which renders the assessment of an experiment difficult. In the context of quality control of an industrial task, a novelty detection method can be employed to determine abnormal or unusual outputs where the novel points can be defined as the observations which have extreme values compared to other measures observed under the same experimental conditions. The new method proposes to screen an additional set of three plates featuring only control wells which constitute the reference data to compare the plates. This set of plates is used to estimate the distribution of the control values. In the foirst place, a Gaussian Mixture Model is trained within the EM algorithm. A point is declared "novel" if its probablility is below a novelty threshold. The technique is compared to a traditional approach of outlier detection. The choice of this threshold is investigated together with alternative approaches to the problem.
|Date of Award||1997|
|Supervisor||Ian T. Nabney (Supervisor)|