1. Introduction

As stated by the authors the chromosomes contain the same bases and associated proteins. They expect that even if the measured chromosome spectra are more or less alike it should be possible, however, to detect smaller structural differences between the chromosomes by spectroscopy [1], (page 5391). Therefore, the sensitivity and stability of the experimental setup are crucial parameters in order to be able to differentiate between signals representing the three chromosomes. At that point it could have been informative if the authors of the cited work have demonstrated the reproducibility of their experimental set-up. The authors present difference spectra with intensities which appear to be close to the noise limit. No patterns in these difference spectra seem to justify that they are able to identify and differentiate between the three human chromosomes. The demonstration of classification is based on the output from an advanced data analytical procedure: Non-linear discriminant analysis (GDA). The GDA procedure has been used separately on two dataset with two different results as outcome. Unfortunately, test results demonstrating the validity of the classification models are not presented.

2. The reproducibility and significance of the measured Raman spectra

The authors of Ref. [1] points to several subtle differences in the chromosome spectra shown in Fig. 3(a) of their work. Inspection of the spectral peaks presented in Fig. 3(a) reveals that the spectral features, which are left back in the spectra after subtraction are mutually translated and apparently show the same positive-negative loop pattern. The accuracy of the wavenumber axis is given by the authors to be within 1 cm^-1. This is a relative large shift. How can one be certain that this shift is caused by morphological and/or chemical differences between the chromosomes and not by the uncertainty in wavennumber axis?

Even if the spectral differences shown in Fig. 3(a) are real there seem not to be any statistical support for discrimination between the chromosomes. Figure 3(b) depicts the intensity ratios at various wavelength with standard deviation for each mean value. In the text the authors claim that the ratios of the three groups are different. But taken into account the standard deviation shown in Fig. 3(b) in combination with table 1(b) it seems that the three spectra overlap substantially. Further in the text they claim that the standard deviations (SD) representing #1 and #2 overlap too much. It is further stated that #3 is quite unlike the other two groups. Reading this plot [Fig. 3(b)] one can actually not say that there is any significant difference between the three types of chromosomes. If a comparison, for instance between the two first mean values in table 1 (b) (#1: 1.97 and #2: 1.95) were done using a double sided t-test (N=22, SD_#1=0.25 and SD_#2=0.21) a p-value equal to 0.7775 comes up. We cannot reject the null hypothesis that states that the two mean values are equal.

The statistical analysis presented in Fig. 3(b) and table 1 and the difference spectra shown at the bottom of Fig. 3(a) seems to point in opposite direction. Take for instance plot ‘f’ (f=(b-c)*3) that seems practically to be a residual representing noise. Difference spectra ‘d’ and ‘e’ gives the derivative like residual. Based on these three difference spectra can we then conclude that ‘b=c’ although ‘c’ is claimed to be different from both ‘a’ and ‘b’?

Furthermore, it is claimed (page 5392) that the variability observed in Fig. 3(b) (and table 1) is not caused by instrument differences (or perhaps other experimental conditions that could change during the measurements). The authors could have documented this statement by demonstrating the reproducibility of the experimental set-up.

3. Experimental design and data analysis

The results of applying discriminate analysis is presented in Figs. 4(a) and 4(b). The input to the data analysis program is the spectra exemplified in Fig. 3(a). It could be tempting to accept the results depicted in Fig. 4. But a closer inspection of Figs. 3(a) and 3(b) seems to indicate that the plots presented in Fig. 4 are a result caused by other reasons. Grouping of the observations in their classes #1, #2 and #3 can equally well be a result of overfitting or incomplete experimental design. Sources of variances that can contribute to measurements are day to day variations such as changes in probe volume temperature, recalibration of the experimental set-up, different operators of the experimental equipment and grouping of samples representing different chromosomes before the measurements are taken place. If these sources of variance contribute with an amount equal to or larger than the potential inherent difference in the spectra some kind of experimental design has to be applied in order to rule out these sources (recall from Fig. 3(b) and Table 1 that the differences between group means were not significantly different). The importance of choosing an appropriate experimental design is discussed in Ref. [3].

Inspection of Figs. 4(a) and 4(b) reveals that the cluster of points belonging to #1, #2 and #3 in Fig. 4(a) falls in different parts of the GDA1 – GDA2 plots. If the underlying difference spectra contain specific spectral features from each chromosome the model based on 6 donors over 12 days should be able to predict which classes the chromosomes from one individual belongs to. The purpose of doing the data analysis is to obtain a model which can be used to predict the type of chromosomes trapped in the laser beam in future experiments. A comparison of Figs. 4(a) and 4(b) seems to indicate that the major part of the variability in the data comes from other sources than real spectral differences between chromosomes. The GDA procedure is a supervised classification model that the authors train on a set of data. But although the model can learn to distinguish the different groups of the training examples it remains to be shown that the model can generalize to other cases. Accordingly, the authors should have kept part of the examples for validation, used the single donor spectra or alternatively use some kind of cross-validation procedure. Observing that the generated clusters in Figs. 4(a) and 4(b) are organized in different ways with respect to the GDA1 and 2 axes could indicate a problem with respect to generalization. Presentation of the results based on validation data set is common practice in papers where chemometric methods are used for data analysis [2, 3, 4].