Bleed-Through

Technical description

M4 is a deterministic screen for implausible relationships between the channels of a fluorescence image. Two fluorophores imaged on one specimen share the same physical sample and are recorded through overlapping spectral windows, so their channels show a moderate, spatially structured correlation: the emission spectra overlap, and the optics and detectors leak a fraction of each signal into the other channel, an effect known as spectral bleed-through or crosstalk. A figure assembled from separately generated or separately sourced channels lacks that shared physical origin and tends toward zero correlation, while a figure in which one channel was duplicated into another shows a correlation near one. M4 splits the image into its red, green, and blue channels, divides each into a grid of 16 by 16 blocks, computes the Pearson correlation of every channel pair within each block, and aggregates the mean absolute correlation per pair. A channel that carries no signal is excluded, because its correlation is undefined rather than zero. The image must be at least 64 by 64 pixels, or the indicator returns a zero score and records that it was skipped.

How it works

Each block contributes one Pearson correlation per channel pair. For two flattened blocks a and b,

r = sum((a - a_bar)(b - b_bar)) / sqrt(sum((a - a_bar)^2) sum((b - b_bar)^2)),

which is zero when a channel block has no variance. The mean absolute correlation of a pair is the average of |r| over all blocks, giving three pair statistics for the red-green, red-blue, and green-blue pairs.

Before scoring, each channel is tested for signal: a channel whose standard deviation across the image is below 1.0 is treated as inactive, and any pair that involves it is dropped, since bleed-through cannot be assessed against an empty channel. For each surviving pair the mean absolute correlation c is mapped to a pair score. A correlation in the natural band, 0.1 <= c <= 0.95, scores zero. A correlation below 0.1 scores 5(1 - c / 0.1), rising toward five as the channels approach independence. A correlation above 0.95 scores 5(c - 0.95) / (1 - 0.95), rising toward five as the channels approach a copy. The final score is the mean of the scored pairs, capped at five, and is zero when fewer than two channels carry signal. Individual blocks whose absolute correlation falls below 0.1 or above 0.95 are reported as findings, labelled independent or copied, ordered by how far they sit outside the natural band, and each carries its block bounding box. The metadata records the overall mean inter-channel correlation, the per-pair means, the block count, and the list of channels that were active.

Score thresholds

Score	Meaning
0 to 1	Channel correlations sit in the natural bleed-through band, consistent with a genuine multi-channel acquisition.
2 to 3	One channel pair is near-independent or near-identical, a possible assembly or duplication or a genuinely separated label.
4 to 5	A channel pair is essentially independent or essentially copied. Consistent with channels assembled separately or one channel duplicated into another.

Why this matters

Multi-channel fluorescence figures are a common vehicle for fabrication, because a reader cannot see at a glance whether the channels belong together. The physics gives one handle: real channels share a specimen and a spectral overlap, so colocalization analysis, the standard quantitative treatment of multi-channel fluorescence, measures their relationship with exactly the statistics M4 uses. The Pearson correlation coefficient is the most widely used colocalization measure and is the basis of the practical guides that the microscopy community relies on [2], complemented by the Manders overlap coefficients that quantify the fraction of one channel coinciding with the other [1]. The same literature stresses that colocalization must be assessed only on channels that carry signal and against the bleed-through that genuine acquisitions always show, which is why automated significance frameworks such as the Costes randomization test were developed to separate real overlap from chance [3], and why guided treatments of the subject devote attention to crosstalk as a confound [4]. M4 turns this body of method around: instead of measuring colocalization to study biology, it reads the channel correlation to ask whether the channels could have come from one acquisition at all, flagging the two unphysical extremes of perfect independence and perfect identity.

Limitations

The correlation of fluorophores is biology-dependent, so the low-correlation cue is the weaker of the two. Two genuinely separated labels, such as a nuclear stain and a sparse membrane marker, legitimately show near-zero correlation, and a real image with minimal spectral overlap can sit below the natural band without any manipulation; a low score here is suggestive, not conclusive. The high-correlation cue is firmer, since channels approaching identity rarely arise from independent fluorophores, but a dye pair with heavy spectral overlap can still correlate strongly. The 16-pixel block bounds the spatial resolution, and the absolute-correlation average is upward-biased on small blocks, which raises the apparent floor. The screen reads the red, green, and blue channels of the rendered image, not the original acquisition channels, so a composite that maps fluorophores unusually onto RGB, or that uses more than three channels, is only partially represented. Inactive channels are excluded to avoid a false independence flag. Reuse of a whole field of view across panels is the job of indicator M2, and noise-based composite cues live in sibling indicators, so M4 stays on the inter-channel relationship.

Theoretical background

M4 rests on the optics of multi-channel fluorescence. When two fluorophores label one specimen, each is excited and emits over a band of wavelengths, and the bands overlap; the microscope separates them with filters that cannot be perfectly selective, so a fraction of each fluorophore's emission is recorded in the other channel. The result is that the two channels are coupled by both shared structure, where the labels co-occur, and instrumental crosstalk, where one signal leaks into the other detector. This coupling produces a correlation that is positive and moderate but rarely extreme, and it is spatially organised by the specimen rather than uniform. Channels generated or sourced independently have no shared specimen and no crosstalk, so their correlation tends to the null value of unrelated noise, near zero. A channel copied from another, with or without light retouching, has correlation near one. The Pearson coefficient measures exactly this linear coupling, and the requirement that a channel carry variance is intrinsic to the coefficient, since it is undefined for a constant signal. Reading the correlation against the natural bleed-through band turns the physics of spectral overlap into a test of whether a set of channels could share one origin.

References

1. Manders EMM, Verbeek FJ, Aten JA. Measurement of co-localization of objects in dual-colour confocal images. Journal of Microscopy. 1993;169(3):375-382. DOI: 10.1111/j.1365-2818.1993.tb03313.x
2. Dunn KW, Kamocka MM, McDonald JH. A practical guide to evaluating colocalization in biological microscopy. American Journal of Physiology-Cell Physiology. 2011;300(4):C723-C742. DOI: 10.1152/ajpcell.00462.2010
3. Costes SV, Daelemans D, Cho EH, Dobbin Z, Pavlakis G, Lockett S. Automatic and quantitative measurement of protein-protein colocalization in live cells. Biophysical Journal. 2004;86(6):3993-4003. DOI: 10.1529/biophysj.103.038422
4. Bolte S, Cordelières FP. A guided tour into subcellular colocalization analysis in light microscopy. Journal of Microscopy. 2006;224(3):213-232. DOI: 10.1111/j.1365-2818.2006.01706.x