Report 098 · Lab Science
Every cell biologist has been told about the edge effect, usually as folklore, usually with the advice to skip the outer ring. A group in Bahrain actually measured it. The outer wells came in 35 percent low, the effect reached three rows into the plate, and the plate brand mattered more than the technique. Then a second study found the effect running the other direction, which is the part that should worry you.
A 96-well plate looks like 96 copies of the same small experiment. It is not. It is a shallow tray of liquid sitting in a warm, humid box that gets opened all day, and the wells near the outside lose water faster than the wells in the middle. Everyone in cell culture knows this. It has a name, the edge effect, and it usually gets handled with a shrug and a habit: do not use the outer ring.
What almost nobody has is a number. So a group at RCSI Bahrain went and got one, and published it in Biochemistry and Biophysics Reports in 2021 under a title that does not hedge: "The edge effect: A global problem. The trouble with culturing cells in 96-well plates."
Their design is deliberately mundane, which is what makes it useful. SW480 colorectal cancer cells, 10,000 per well in 100 µl of medium, 72 hours at 37 °C in 5 percent CO2, then an MTS assay read at 490 nm. Two brands of plate, VWR and Greiner. And one detail I want to flag as excellent experimental hygiene: to mimic a normal working lab rather than an idealized one, they opened the incubator door ten times during the incubation.
In the VWR plates, compared against the four central wells, metabolic activity was down:
Corner wells: 34 ± 2 percent lower. Outer wells: 35 ± 3 percent. Second row in: 25 ± 5 percent. Third row in: 10 ± 5 percent. All four reductions were statistically significant, the first three at p < 0.0001 and the third row at p = 0.015.
Read the third-row figure again, because that is the finding. The received wisdom is that the edge is one ring wide. In these plates it was three rows deep. The authors put it flatly in their conclusion: in some plates "the edge effect extends well beyond the edge."
Now do the accounting on what that costs. Excluding the outer row leaves 60 of 96 wells usable. Excluding the outer two rows leaves 32. The paper works this out in plain economics: those exclusions raise your plate cost by roughly 50 percent and 200 percent respectively. That is the real reason the practice is inconsistent across labs. It is not ignorance, it is budget.
The Greiner plates behaved substantially better under identical handling: corners 26 ± 4 percent low, outer wells 16 ± 8 percent, and the second and third rows not significantly different from center at all (7 ± 7 and 1 ± 6 percent, p = 0.66 and 1.00).
Same cells, same incubator, same person, same door-opening protocol. Different consumable, different experiment. If you have ever switched plate suppliers mid-project because purchasing found a better price, that sentence should land somewhere uncomfortable.
The mitigations they tested are also worth knowing, mostly because the popular one barely works. Putting the plate back inside its original packaging during incubation, the classic bench trick, helped the VWR plates a little (corners to 26 ± 5 percent, outer to 23 ± 4 percent) but left the outer two rows still significantly depressed. For the Greiner plates it did almost nothing.
What worked best was filling the spaces between the wells with sterile PBS. With all interstitial spaces filled, the corner and outer reductions fell to 13 ± 8 and 10 ± 9 percent and neither remained statistically significant. The authors still say they would not rely on data from the outer wells. And there is a catch they report honestly: adding buffer lowered the overall MTS signal, possibly because the extra liquid mass changes how fast the plate returns to temperature after handling.
Everything above describes an effect with a consistent direction: outer wells read low. If that were universal you could at least reason about it.
It is not universal. A 2020 Scientific Reports study from the University of Gothenburg, examining sources of variability in cancer drug sensitivity screens, found the opposite sign. Working with a resazurin readout on breast cancer lines, and using culture microplates specifically designed to minimize evaporation, they observed an edge effect in which elevated absorbance values were measured for cells in the perimeter wells, for both DMSO controls and bortezomib-treated wells. Their optimization table lands on the same instruction from the other direction: avoid perimeter wells, rows A and H and columns 1 and 12.
The Bahrain authors ran into the same contradiction in the literature and were careful about it. They note a prior group that saw MTS readout increase as volume decreased with no change in cell number, walk through the Beer-Lambert argument for why concentration and pathlength changes should partly cancel, and then say straightforwardly that they cannot explain the other group's result.
That unresolved disagreement is the practical takeaway. The edge effect is not a constant you can subtract. Its magnitude depends on the plate, and its sign depends on the readout chemistry, the duration, and what evaporation does to your particular detection reaction. Any protocol that says "correct for the edge effect" without measuring it in that lab, in that incubator, with that plate and that assay, is applying a fudge factor of unknown direction.
The reason this matters beyond tidiness is dose-response.
When medium evaporates from a well, the drug in it becomes more concentrated. At the same time the salts become more concentrated, which independently slows the cells down. So a perimeter well in a cytotoxicity screen is simultaneously receiving a higher effective dose and suffering a nonspecific growth penalty, and both effects land on the same viability number, at the same well position, in the same direction of "looks more killed."
If your layout assigns concentration by column, which most do, then position and dose are confounded by construction. Your IC50 shifts, and nothing in the output tells you it shifted.
There is a second, quieter finding in the Bahrain paper that I think is the most transferable thing in it.
They state that their analyses "consider each 96-well plate to represent n = 1," and then observe that this appears not to have been the case in some earlier publications, where each well seems to have been treated as an individual data point.
That is the difference between n = 6 and n = 576, and it is not a rounding error in a p-value, it is the whole p-value. Wells on one plate share an incubator position, a medium batch, a seeding session and an evaporation gradient. They are not independent replicates of anything. Treating them as independent is textbook pseudoreplication, and it manufactures significance out of a shared physical environment.
This connects to two things I have written before: what error bars actually mean, and the difference between reproducible and replicable. A result that survives only because 96 correlated wells were counted as 96 experiments is not reproducible in any sense worth the word. It will fail the moment someone else runs it on a different brand of plate.
Four things, all cheap.
Map your own plate before you trust a layout. Run one plate with uniform conditions in every well, on your bench, in your incubator, with the consumable you actually buy, for the duration you actually use. You now know your own gradient, its depth and its sign. This is one plate, once per plate type.
Do not assign treatment by position. If concentration runs along columns and evaporation runs along columns, you cannot separate them afterward. Randomize or block treatments across positions so a positional artifact becomes noise rather than a trend.
If you exclude wells, say so in methods, and count the cost. "Outer two rows excluded" is a legitimate design choice and it means 32 usable wells out of 96. Undisclosed exclusion is a different thing entirely.
Treat the plate as the experimental unit. Wells are technical replicates. Plates run on different days are biological replicates. Report n accordingly, even when the number gets embarrassingly small.
The Bahrain study is a single laboratory, one cell line, one readout, with n = 6 or n = 8 plates per condition. The authors say so, and explicitly advise caution because some of their observed trends did not reach significance at that sample size. The specific percentages here belong to SW480 cells in those two plate brands under those conditions. Do not carry the numbers to your own bench. Carry the method.
The plate-cost figures of roughly 50 and 200 percent are the authors' arithmetic, reported as they give it, and they assume you discard the unused wells rather than repurpose them.
The Gothenburg study is cited here only for the direction of its perimeter-well observation and its resulting protocol recommendation. It was not designed as a head-to-head comparison with the Bahrain work, and I am not presenting it as a refutation, only as evidence that the sign is not fixed.
One source I wanted and could not open: a 2016 Journal of Biomolecular Screening paper on medium evaporation and spheroid reproducibility in 384-well plates, cited by the Bahrain authors. It is paywalled, so nothing here rests on it.
Three things to carry out of this.
First, the edge effect is not a one-ring problem. In at least one commercially sold plate it was measurable three rows in, which means a protocol that skips only row A and row H can still be reading a gradient across everything it kept.
Second, a consumable is an experimental variable. Two plate brands, handled identically, produced materially different data. Lot and supplier changes belong in the methods section next to the antibody catalogue number, and for the same reason.
Third, the general form. Physical position in an apparatus is a variable whether or not your analysis has a term for it. That is true of wells in an incubator, samples in an autosampler tray, cells in a flow cell, and detectors in an array. The instrument does not know your experimental design, and it will happily let a temperature gradient masquerade as your effect.
U.S. Army combat veteran (Counter-IED / Electronic Warfare), peer-reviewed researcher in microwave spectroscopy, and founder & CEO of Shroombiosis. Consults on laboratory operations, AI, and supplement formulation.