Introduction. Spectroscopic analysis of urine samples from laboratory animals can be used to predict the efficacy and side effects of drugs. This employs methods combining 1 H NMR spectroscopy with quantification of biomarkers or with multivariate data analysis. The most critical steps in data evaluation are analytical reproducibility of NMR data (collection, storage, and processing) and the health status of the animals, which may influence urine pH and osmolarity. Methods. We treated rats with a solvent, a diuretic, or a nephrotoxicant and collected urine samples. Samples were titrated to pH 3 to 9, or salt concentrations increased up to 20-fold. The effects of storage conditions and freeze-thaw cycles were monitored. Selected metabolites and multivariate data analysis were evaluated after 1 H NMR spectroscopy. Results. We showed that variation of pH from 3 to 9 and increases in osmolarity up to 6-fold had no effect on the quantification of the metabolites or on multivariate data analysis. Storage led to changes after 14 days at 4°C or after 12 months at −20°C, independent of sample composition. Multiple freeze-thaw cycles did not affect data analysis. Conclusion. Reproducibility of NMR measurements is not dependent on sample composition under physiological or pathological conditions.

Recently, Ensembles of local experts have successfully been applied for the automatic detection of drug-induced organ toxicities based on spectroscopic data. For suitable Ensemble composition an expert selection optimization procedure is required that identifies the most relevant classifiers to be integrated. However, it has been observed that Ensemble optimization tends to overfit on the training data. To tackle this problem we propose to integrate a stacked classifier optimized via cross-validation that is based on the outputs of local experts. In order to achieve probabilistic outputs of Support Vector Machines used as local experts we apply a sigmoidal fitting approach. The results of an experimental evaluation on a challenging data set from safety pharmacology demonstrate the improved generalizability of the proposed approach.

A new approach for the automatic detection of drug-induced organ toxicities based on Nuclear Magnetic Resonance Spectroscopy data from biofluids is presented in this paper. Spectral data from biofluids contain information on the concentration of various substances, but the combination of only a small subset of these cues is putatively useful for classification of new samples. We propose to divide the spectra into several short regions and train classifiers on them, using only a limited amount of information for class discrimination. These local experts are combined in an ensemble classification system and the subset of experts for the final classification is optimized automatically. Thus, only local experts for relevant spectral regions are used for the final ensemble classification. The proposed approach has been evaluated on a real data-set from industrial pharmacology, showing an improvement in classification accuracy and indicating relevant spectral regions for classification.

INTRODUCTION:The aim of safety pharmacology is early detection of compound-induced side-effects. NMR-based urine analysis followed by multivariate data analysis (metabonomics) identifies efficiently differences between toxic and non-toxic compounds; but in most cases multiple administrations of the test compound are necessary. We tested the feasibility of detecting proximal tubule kidney toxicity and phospholipidosis with metabonomics techniques after single compound administration as an early safety pharmacology approach.METHODS:Rats were treated orally, intravenously, inhalatively or intraperitoneally with different test compounds. Urine was collected at 0-8 h and 8-24 h after compound administration, and (1)H NMR-patterns were recorded from the samples. Variation of post-processing and feature extraction methods led to different views on the data. Support Vector Machines were trained on these different data sets and then aggregated as experts in an Ensemble. Finally, validity was monitored with a cross-validation study using a training, validation, and test data set.RESULTS:Proximal tubule kidney toxicity could be predicted with reasonable total classification accuracy (85%), specificity (88%) and sensitivity (78%). In comparison to alternative histological studies, results were obtained quicker, compound need was reduced, and very importantly fewer animals were needed. In contrast, the induction of phospholipidosis by the test compounds could not be predicted using NMR-based urine analysis or the previously published biomarker PAG.DISCUSSION:NMR-based urine analysis was shown to effectively predict proximal tubule kidney toxicity after single compound administration in rats. Thus, this experimental design allows early detection of toxicity risks with relatively low amounts of compound in a reasonably short period of time.

The early detection of drug-induced organ toxicities is one of the major goals in safety pharmacology. Automating this process by classification of metabolic changes based on the analysis of 1H nuclear magnetic resonance spectra improves this process. In this paper we propose an ensemble classification system based on support vector machines trained on diverse ldquoviewsrdquo on the data. These views are created by variation of preprocessing techniques and the final classification is achieved by voting on an optimized selection of all experts. Results of an experimental evaluation on a challenging data-set from industrial safety pharmacology show the effectiveness of the proposed approach w.r.t. the detection of drug-induced toxicity.

We present an approach for the automatic classification of Nuclear Magnetic Resonance Spectroscopy data of biofluids with respect to drug induced organ toxicities. Classification is realized by an Ensemble of Support Vector Machines, trained on different subspaces according to a modified version of Random Subspace Sampling. Features most likely leading to an improved classification accuracy are favored by the determination of subspaces, resulting in an improved classification accuracy of base classifiers within the Ensemble. An experimental evaluation based on a challenging, real task from pharmacology proves the increased classification accuracy of the proposed Ensemble creation approach compared to single SVM classification and classical Random Subspace Sampling.

The annotation of song changes in music mixes created by DJs or radio stations for direct access in digital recordings is, usually, a very tedious work. In order to support this process we developed an automatic song change detection method which can be used for arbitrary music mixes. Stochastic models are applied to music data aiming at their segmentation with respect to automatically obtained abstract generic acoustic units. The local analysis of these stochastic music models provides hypotheses for song changes. Results of an experimental evaluation processing music mix data demonstrate the effectiveness of our method for supporting the annotation with respect to song changes