Spectrum interpretation pattern recognition

Because < this well-developed ability to perceive shapes, chemists often use picturesaa present their data. For example, in spectroscopy, a spectrum is plotted as arontinuous curve rather than represented in tabular form. Tlie human eye can perceive the presence or absence of peaks, and interpretations are made accardingly. Using the computer to perceive" these sliapes or to enhance recogrftion abilities is the goal of pattern recognition. 

The development and widespread use of computers and microprocessors in control laboratory instruments has made it possible to fully automate a laboratory, including interfacing instruments directly to a LIMS. In the fully automated laboratory, a sample is logged into a LIMS, then transferred to a laboratory where it is prepared for analysis by a robot, which then transfers it to an autosampler or analyzer. Once analyzed, the data is transferred through a communications link to a device which could convert the raw data into information that a customer needs. For example, in a simple case, a nmr spectrum could be compared to spectra on file to yield an identification of an unknown. In more complex instances, a data set could be compared to standards and by using pattern recognition techniques the LIMS would be able to determine the source of a particular raw material. Once the data is reduced and interpreted, the LIMS becomes the repository of the information. A schematic for such a fully automated laboratory is shown in Figure 2 (6). 

Two main approaches have been used to interpret bacterial mass spectra. First, pattern recognition methods have been used, in which reference spectra are obtained under standardized conditions to generate a fingerprint collection. An unknown bacterial spectrum obtained under the same conditions is then compared to the collection and matched if possible. 

When there is no spectrum in the library that can be used to elucidate a chemical structure, interpretative methods are needed. The methods of pattern recognition and of artificial intelligence must then be used. As a result, different chemical structures will be obtained as candidates for the unknown molecule. To verify an assumed structure, simulation of spectra becomes important. In a final step, the simulated spectrum could be compared with that measured. 

Wonders must not be expected from conceptually rather simple methods of pattern recognition for the solution of very complex problems (like the interpretation of a spectrum or the description of structure-activity-relationships). Pattern recognition can be seen only as one part of a computer-assisted interpretation system for chemical data. Accentuation should be given to "assisted because a complete processing by a machine of sophisticated data interpretations in chemistry is unrealistic and uneconomical at least for the next two decades. 

Three different approaches have been used for computer-assisted interpretations of chemical data. 1. Heuristic methods try to formulate computer programs working in a similar way as a chemist would solve the problem. 2. Retrieval methods have been successfully used for library search (an unknown spectrum is compared with a spectral library). 3. Pattern recognition methods are especially useful for the classification of objects (substances, materials) into discrete classes on the basis of measured features. A set of characteristic features (e.g. a spectrum) of an object is considered as an abstract pattern that contains information about a not directly measurable property (e.g. molecular structure or biological activity) of the object. Pure pattern recognition methods try to find relationships between the pattern and the "obscure property" without using chemical knowledge or chemical prejudices. 

Spectrum interpretation is extraction of structural properties flora spectroscopic data. Methods of pattern recognition and of supervised statistical learning aroused (Section 8.5). 

Two types of pattern recognition technique " have been applied to spectrum interpretation. Supervised methods (see Supervised Learning) are limited to one or more predefined structural classes and require representative training sets for each to develop the classifier. Unsupervised methods (see Unsupervised Learning) partition a set of spectra into clusters with common structural features on the basis of spectral features alone. No predefined classes are required. 

Unsupervised methods of pattern recognition have received less attention in spectrum interpretation. In one application to the interpretation of IR spectra, an ordered binary hierarchical tree was generated from a set of spectra of known compounds using the three-distance clustering method. There were no predefined substructures. 

Nuclear magnetic resonance (NMR) spectroscopy is the most informative analytical technique and is widely applied in combinatorial chemistry. However, an automated interpretation of the NMR spectral results is difficult (3,4). Usually the interpretation can be supported by use of spectrum calculation (5-18) and structure generator programs (8,12,18-21). Automated structure validation methods rely on NMR signal comparison using substructure/ subspectra correlated databases or shift prediction methods (8,15,22,23). We have recently introduced a novel NMR method called AutoDROP (Automated Definition and Recognition of Patterns) to rapidly analyze compounds libraries (24-29). The method is based on experimental data obtained from the measured ID or 2D iH,i C correlated (HSQC) spectra. 

---