Application of fuzzy logic for automated interpretation of mass spectra

The advent of laboratory automation has given researchers the capability to acquire a plethora of data so that data interpretation is now the rate limiting step, which is particularly true in mass spectrometry used for high-throughput screening. Fuzzy logic is a natural way to encapsulate the processes used by human experts to interpret mass spectra because mass spectra can be described linguistically. Unfortunately, the vast majority of approaches to automated mass spectral analysis have relied on neural network or chemometric strategies. Artificial neural network interpretation schemes tend to become a “black box” that lacks transparency in the decision process and require greater skill in choosing an architecture and training method than the mass spectrometrist may realize and, therefore, are prone to pitfalls that may not be visible without expensive validation efforts. Chemometric techniques tend to mishandle unanticipated spectral data, often obscuring valuable information. In contrast, fuzzy logic rules are straightforward mathematical expressions of plain language; thus, the decisionmaking process is transparent. In addition, fuzzy methods can be used to highlight the presence of unexpected, unknown spectra. In the case of data sets in which there is little a priori knowledge about the solution, other data discovery methods can be used to assist in forming the rule base. A prudent approach uses these methods to abstract key features from data sets to form fuzzy rules, but then relies on fuzzy logic to capture the human inference process to form the decision basis implemented in the classifier. Examples illustrate the fuzzy methods needed to describe mass spectra, to develop fuzzy rule bases, to employ fuzzy logic approaches to classify spectra, and to construct methods to determine hard classifications (i.e., defuzzification) for final decision making as well as methods to resolve uncertainty in classification with information provided by spatially related data points in mass spectra based chemical imaging. The illustrations presented provide a bridge to span the gap between scientists using mass spectrometry and practitioners of fuzzy logic.