Comment on: “Utilizing Artificial Neural Networks in MATLAB to Achieve Parts-Per-Billion Mass Measurement Accuracy with a Fourier Transform Ion Cyclotron Resonance Mass Spectrometer” by D. Keith Williams Jr., Alexander L. Kovach, David C. Muddiman, and Kenneth W. Hanck. J. Am. Soc. Mass Spectrom. 20, 1303–1310 (2009)

This is an excerpt from the content

When calibrating a FT-ICR MS system [1], the authors report that “results …demonstrated that a fit using artificial neural networks (ANN) provided a better fit of data than the multiple linear regression (MLR) method previously published” [2]. However, the article’s Figure 3a (that “illustrates the residuals generated by the fitting of the calibration data by MLR”) shows the largest two oligomers to be outliers. The residuals shown in Figure 3b from the ANN show the outlier extreme displacements are gone from the largest two oligomers, but now displacements along the line of smaller oligomers reflect an underlying curvature in the ANN fit.

Five statistical tests were run that showed non-normality when ML was fit but not when ANN was run, but the tests were likely detecting kurtosis (long tailed distributions) due to the outliers. The kurtosis was diminished after fitting the compliant ANN, which masked the outliers. However, the above-mentioned displacements testify that passing normali