It has been shown that various mathematical transformations of mass spectra may increase classification efficiency. Better separation of classes can be achieved in some cases if transformed, instead of original spectra, are submitted to classification. In addition, some transformation procedures reduce the number of variables and lower the dimensionality of the spectral space, which shortens the computing time.

Because the most common neutral loss is 14 (loss of CH2), the logical spectra transformation is into modulo-14 spectra, which thereafter can be used as input data for PCA. Modulo-14 spectra are calculated as the sum of peak heights at mass-to-charge ratio values shifted by 14. Each modulo-14 spectrum has 14 dimensions (transformed mass-to-charge ratio values) that are significantly lower than regular spectra. Mass Frontier offers modulo-14 transformation with (A1) or without (A2) normalization of such spectra

Classification of mass spectra assists the interpretation of structurally related compounds. Because the characteristic peaks in spectra of structurally related compounds can be shifted due to various substituents, it can be difficult for classification methods to recognize structural similarity. This difficulty can be overcome by transforming spectral data into auto-correlation spectra. The auto-correlation function:

is suitable for detecting periodicity in a series of spectra. In Mass Frontier you can choose auto-correlation transformation with (B1) or without (B2) normalization of mass spectra. Since auto-correlation does not reduce the space dimensionality and requires computing time to be calculated, a classification that uses this transformation is the most time-consuming procedure among the transformation methods.

Mass Frontier allows the user to submit original (not transformed) spectra (C1) to classification as well.

Mass Frontier offers the following spectra transformations:

• Modulo-14 spectra normalized  A1
• Modulo-14 spectra not normalized  A2
• Auto-correlation spectra normalized  B1
• Auto-correlation spectra not normalized B2
• Original spectra (not transformed)   C1  (default)

No general rule exists concerning the selection of an appropriate mass spectrum transformation. Classification methods can be employed for a broad range of problems, and each of them may need a different spectrum transformation. Your objective should be to find the transformation that provides the best separation of classes. The only reliable way to find the best transformation for particular groups of spectra, is to experiment with all of them. Subsequently, the transformation which provides you with the most information will be your first choice when dealing with comparable data.