Feature Learning For Improved Multiplex Disease Diagnosis

Abstract This proposal is for CATTSa feature learning technique optimized for use in multiplex mass spectrometryMSfingerprinting assaysMS fingerprints consist of a large number of chemical speciesleading to very high dimensional feature spacesand subsequent high false discovery ratesCATTS aims to reduce the size of this spaceby using knowledge of the underlying biochemistryas well as general purpose clustering algorithmsOur preliminary results demonstrate thatwhen used as a feature learning technique for a variety of classification methodsCATTS significantly improves assay sensitivityThis proposal takes our existing implementation of CATTS and extends it to support additional feature learning algorithms and classification methodsAdditionallyits performance as a multiplex assay strategy will be tested on both protein and lipid MS fingerprint librarieswith an eye towards commercializationRelevance to public healthThe detection of pathogens via mass spectroscopy fingerprinting is rapidly becoming a standard technique for clinical microbiologyHoweverhigh false detection rates and conflicting multiple identifications limit applicabilityand make interpretation of results difficultOur work on CATTS aims to improve the statistical performance of these assaysPreliminary results from studies on one dataset we intend to apply CATTS to suggest that UTIs andin some cases antimicrobial resistancecan be detecteddirectly from patient samplesHoweverthe statistical methods currently employed arenandapos t reliable enoughthe further development of CATTS will accelerate the development of thisand other mass spectroscopy based assays