Designing open, multi-class computational strategies to classify infrared spectroscopy data derived from the Syrian hamster embryo (SHE) assay

The Syrian hamster embryo (SHE) assay (pH6.7) is being touted as a ??3R?s?? alternative in animal laboratory studies. In the SHE assay, traditionally, colonies are counted and scored by eye to determine the transforming potential of test chemicals. Application of infrared (IR) spectroscopy opens up the possibility of comparing test chemicals with negative and positive controls in a high-throughput fashion (1) through objective pattern recognition methods. Such methods are under development under the 1) ??openness??, and 2) multiple-class requirements: 1) computer systems need to be ??open?? to new data to refine existing classifiers; 2) furthermore, the existence of multiple classes (i.e., chemical treatment conditions) calls for composite architectures containing many simple classifiers instead a single complicated one. In this study we present two classification strategies contemplating these two principles. The proposed strategies are compared to well established ??closed??, single-model classifiers. The dataset used in the study was derived from a SHE assay where eight treatment conditions were present vehicle control (DMSO), D-M, Ba]P, 3-MCA, Anthracene, o-T, 2,4-dT, and MNNG] (2). From the assay, IR spectra (n?14,000) were obtained using attenuated total reflection Fourier-transform IR spectroscopy. Gradual feeding of the proposed models with training data is shown to gradually improve the classification of test data. Segregation of data along chemical mode of action was observed. Overall, the results strengthen arguments towards using the SHE assay in toxicological assessments and point to IR spectroscopy as a possible alternative to visual scoring.