RAPID AND QUANTITATIVE-ANALYSIS OF METABOLITES IN FERMENTER BROTHS USING PYROLYSIS MASS-SPECTROMETRY WITH SUPERVISED LEARNING - APPLICATIONTO THE SCREENING OF PENICILLIUM-CHRYSOGENUM FERMENTATIONS FOR THE OVERPRODUCTION OF PENICILLINS

The combination of pyrolysis mass spectrometry (PyMS) and artificial n eural networks (ANNs) can be used to quantify levels of penicillins in strains of Penicillium chrysogenum and ampicillin in spiked samples o f Escherichia coli. Four P. chrysogenum strains (NRRL 1951, Wis Q176, P1, and P2) were grown in submerged culture to produce penicillins, an d fermentation samples were taken aseptically and subjected to PyMS. T o deconvolute the pyrolysis mass spectra so as to obtain quantitative information on the titre of penicillins, fully-interconnected feedforw ard artificial neural networks (ANNs) were studied; the weights were m odified using the standard back-propagation algorithm, and the nodes u sed a sigmoidal squashing function. In addition the multivariate linea r regression techniques of partial least squares regression (PLS), pri ncipal components regression (PCR) and multiple linear regression (MLR ) were applied. The ANNs could be trained to give excellent estimates for the penicillin titre, not only from the spectra that had been used to train the ANN but more importantly from previously unseen pyrolysi s mass spectra. All the linear regression methods failed to give accur ate predictions, because of the very variable biological backgrounds ( the four different strains) in which penicillin was produced and also of the inability of models using linear regression accurately to map n on-linearities. Comparisons of squashing functions on the output nodes of identical 150-8-1 neural networks revealed that networks employing linear functions gave more accurate estimates of ampicillin in E. col i near the edges of the concentration range than did those using sigmo idal functions. It was also shown that these neural networks could be successfully used to extrapolate beyond the concentration range on whi ch they had been trained. PyMS with the multivariate clustering techni que of principal components analysis was able to differentiate between four strains of P. chrysogenum studied, and was also able to detect p henotypic differences at five, seven, nine or 11 days growth. A crude sampling procedure consisting of homogenised agar plugs proved applica ble for rapid analysis of a large number of samples.