Search

Large volumes of data are routinely collected during bioprocess operations and, more recently, in basic biological research using genomics-based technologies. While these data often lack sufficient detail to be used for mechanism identification, it is possible that the underlying mechanisms affecting cell phenotype or process outcome are reflected as specific patterns in the overall or temporal sensor logs. This raises the possibility of identifying outcome-specific fingerprints that can be used for process or phenotype classification and the identification of discriminating characteristics, such as specific genes or process variables. The aim of this work is to provide a systematic approach to identifying and modeling patterns in historical records and using this information for process classification. This approach differs from others in that emphasis is placed on analyzing the data structure first and thereby extracting potentially relevant features prior to model creation. The initial step in this overall approach is to first identify the discriminating features of the relevant measurements and time windows, which can then be subsequently used to discriminate among different classes of process behavior. This is achieved via a mean hypothesis testing algorithm. Next, the homogeneity of the multivariate data in each class is explored via a novel cluster analysis technique called PC1 Time Series Clustering to ensure that the data subsets used accurately reflect the variability displayed in the historical records. This will be the topic of the second paper in this series. We present here the method for identifying discriminating features in data via mean hypothesis testing along with results from the analysis of case studies from industrial fermentations, (Copyright 2000 Academic Press.)

An important step in data analysis is class assignment which is usually done on the basis of a macroscopic phenotypic or bioprocess characteristic, such as high vs low growth, healthy vs diseased state, or high vs. low productivity. Unfortunately, such an assignment may lump together samples, which when derived from a more detailed phenotypic or bioprocess description are dissimilar, giving rise to models of lower quality and predictive power. In this paper we present a clustering algorithm for data preprocessing which involves the identification of fundamentally similar lots on the basis of the extent of similarity among the system variables. The algorithm combines aspects of cluster analysis and principal component analysis by applying agglomerative clustering methods to the first principal component of the system data matrix. As part of a rational strategy for developing empirical models, this technique selects lots (samples) which are most appropriate for inclusion in a training set by analyzing multivariate data homogeneity. Samples with similar data structures are identified and grouped together into distinct clusters. This knowledge is used in the formation of potential training sets. Additionally, this technique can identify atypical lots, i.e., samples that are not simply outliers but exhibit the general properties of one class but have been given the assignment of the other. The method is presented along with examples from its application to fermentation data sets., (Copyright 2000 Academic Press.)