Herein, we challenge one of the main conclusions in the Review on colocalization recently published in Journal of Cell Science (Aaron et al., 2018), that the Manders' overlap coefficient (MOC) is a valuable coefficient for assessing colocalization by co-occurrence.
The underlying theme of the Review is that colocalization comprises two distinct phenomena, co-occurrence and correlation. We are pleased that our proposal that colocalization should be treated as two distinct but complimentary measures is gaining acceptance (Adler et al., 2008; Adler and Parmryd, 2007, 2013). The division is powerful, in that it allows the growing number of colocalization coefficients to be characterized and compared. Accordingly coefficients can be categorized as measuring either co-occurrence, the extent of a common distribution, or correlation, the strength of the relationship between intensities. This scheme also exposes a third group of coefficients that report a mix of co-occurrence and correlation, which we termed ‘hybrids’. Our detailed studies conclude that both the MOC (Manders et al., 1993) and the more recently introduced Hcoeff (Herce et al., 2013) are hybrid coefficients (Adler and Parmryd, 2010; Adler and Parmryd, 2014). The problem with hybrid coefficients is interpretation, since they fail to differentiate between widely differing combinations of co-occurrence and correlation (Fig. 1). In the datasets shown in Fig. 1, a MOC value of 0.6 covers a co-occurrence of between 0.37 and 0.82 (a scale of 0–1), depending on the correlation.
The MOC reports co-occurrence poorly since it is affected by the degree of correlation. Three paired datasets (200×200 pixels) whose correlations extend across the full range (1, 0 and −1) had their co-occurrence (M1 or fraction of area common to both – in this simulation they are the same) progressively altered while their co-occurrence M2 remained maximal. The initial distributions were Gaussian, clipped to a range of mean±twice the s.d., with the lowest value above zero. The co-occurrence was progressively reduced by setting an increasing number of pixels in one dataset to zero.
The MOC reports co-occurrence poorly since it is affected by the degree of correlation. Three paired datasets (200×200 pixels) whose correlations extend across the full range (1, 0 and −1) had their co-occurrence (M1 or fraction of area common to both – in this simulation they are the same) progressively altered while their co-occurrence M2 remained maximal. The initial distributions were Gaussian, clipped to a range of mean±twice the s.d., with the lowest value above zero. The co-occurrence was progressively reduced by setting an increasing number of pixels in one dataset to zero.
Aaron et al. discuss the MOC at some length and advocate its use as a measure of co-occurrence. We find this surprising since our earlier investigation concluded that the MOC had little value (Adler and Parmryd, 2010) and Aaron et al. present no observations to challenge or alter our conclusion. In fact, their data supports our view; in their Fig. 6, two cells with differing correlations (Pearson: 0.76 and 0.36) and co-occurrences (M1 and M2: 0.99 and 0.44, and 0.68 and 0.71) are nonetheless reported by the MOC to be almost identical (0.68 and 0.72, on a scale of 0–1). Hybrid coefficients confuse rather than inform, since different combinations of correlation and co-occurrence can produce the same numerical values.
We have for more than a decade advocated that the most efficient way of describing patterns of colocalization is to report both correlation and co-occurrence. Acceptance of this scheme leaves no role for hybrids like the MOC, and the additional evidence we have now provided strengthens our earlier conclusion that “The MOC is not suitable for making measurements of colocalization either by correlation or co-occurrence” (Adler and Parmryd, 2010).
Footnotes
Funding
This work is supported by a grant to I.P. from the Swedish Research Council (2015-04764).
