When considering the above points, in balance with other reasonable explanations, it would appear that data analysis and subsequent interpretation represents the key area in which the two groups adopted fundamentally different approaches that may help to explain the discrepancies observed. Snyder and Saunders employed an intensity-based method that analysed the two channels of a two-colour microarray independently whereas Stabler et al. used a ratio-based method that analysed the two channels of a two-colour microarray in combination.
Snyder and Saunders used a metric (pON) that reflects the probability of there being a hybridisation signal for each spot in each channel independently. Whilst this seems a reasonable approach it takes no account of any relative intensity levels; a gene may be called present in two strains if the pON threshold is exceeded in both yet the relative intensity level may be significantly higher in one strain than the other. For example, for a pathogen gene that was highly divergent in N. lactamica there may still be a low but sufficient signal intensity above background to call the gene as present in N. lactamica using the pON metric, yet this level of hybridisation would be insignificant when compared to the signal intensity achieved with a pathogen; therefore a more accurate call for this gene should be absent or highly divergent rather than present. This also raises the issue that the PMT gain settings during scanning may have an impact on the number of genes passing the pON threshold and thus being assigned as present. If this were true then a normalisation strategy would need to be employed to account for this to ensure the definition of presence is consistent across different arrays and strains and is not dependent on scanning. Furthermore, the comparison of absolute intensities for different genes can be affected by factors such as the concentration, length, and Tm of reporter elements. The authors have essentially adopted a one-colour approach to the analysis of two-colour data, although generally some form of normalisation and an attempt to address the issue of relative intensity would be included. However, in their study relative intensities would not have been informative as no pathogens were included in the comparison.
Stabler et al. approached data analysis in a more common approach for two-colour data by first removing unreliable low intensity data based on image analysis QC flags, normalisation of ratio data to remove any systematic effects and applying ratio cutoffs to determine genes present or absent in comparison to the control strain hybridised to each array as a common reference. The GACK software used to set dynamic ratio cutoffs and make the 'present', 'divergent' or 'absent/highly divergent' calls was set at its most conservative, classified as 'present' if the estimated probability of presence (EPP) = 100% and 'absent/highly divergent' if EPP = 0% with 'divergent' genes between these two extremes. Snyder and Saunders wrongly indicate that the calls of absence by Stabler et al. were based purely on a lack of hybridisation; if this were the case then we would agree with the other possible technical reasons for this that they describe in detail. In actual fact, the calls of absence by Stabler et al. are based on the ratio of intensities for the test strain channel and the common reference strain control channel. Therefore, the call of absence depends on direct comparison to the positive signal intensity for the same spot which serves as an internal control. This presents one of the great strengths of two-colour analysis that helps to circumvent the highlighted problems of missing or poorly performing reporter elements as the same reporter is being used in each channel of the comparison.
These fundamental differences in data analysis present the most likely source of discrepancies between the two studies as all subsequent interpretations are based on the issue of gene presence in different strains. Snyder and Saunders were likely to overestimate gene presence and unable to discriminate conserved genes from highly divergent genes as relative intensities are not taken into account nor any direct comparison to pathogens. In contrast, Stabler et al. are likely to underestimate gene presence due to the perhaps overly strict criteria applied to gene selection, that is, the need for a 100% certainty of gene presence in every strain within a group, that would exclude any gene just missing these thresholds in a single strain. Given these two likely extremes it is not surprising that so many discrepancies were reported.