The data looks pretty reasonable.  Dilution range between 10x and 1000x is really wide for ELISA.  It is pushing the limit of this technology.  Those data points around the middle of the curve are close to each other indicating the reliable range of this ELISA assay.  Those data points at the extreme of both sides, such as below 1.5 and above 2.5, with higher variability indicated that the concentrations at those dilutions are at or near the assay plateau of either high or low end.  The dilution result provided a reference to adjust assay range with acceptable accuracy and precision.  The lower readings at low dilution factor is also an indication of hook effect, suggesting that the minimum required dilution should be increased to alleviate matrix effect. A parallelism test between recombinant calibrator and matrix with endogenous marker is the best way to demonstrate the similarity between the 2 molecues and valid the application of calibrator.

As you stated this is commonly called the "matrix" effect.  Which really means there is something going on and we are not quite sure what it is, but it is reasonably reproducible under set conditions.  The reproducibility allows us to create useful clinical assays.  The unknown part will drive you crazy.  Sometimes it is possible to minimize or alter the effect you are seeing by manipulating your matrix (salts, proteins, detergents, etc.).  Also, as you dilute one binding partner of an antibody/antigen pair things are not always linear.  Antibodies have two binding sites and have an avidity effect.  The avidity of the antibody is affected by the relative concentration and spatial geometry of the antigen.  Also, the concentration of the antibody itself may alter the binding characteristics of the assay.  At high antibody concentrations the antibody may only be able to bind one site (self-competition), which will reduce the avidity and you will see this as weaker binding.  This is why most clinical assays have a range of use qualifier.  This way the sample is only measured where the "matrix" effect is minimal.

Thank you, James for you comment!

I decided to dilute my samples in such a way to obtain maximal antigen levels. In most cases this approach works pretty well. The measurements are reproducible and I obtain sensible results. In some samples antigen content estimations were checked by independent method (western blot), which confirmed that they were correct.

However, there are some samples which produce different results. Estimations get higher and higher while samples get more and more diluted. In these cases optical densities in wells happen to be low (0.1-0.2) but more or less constant despite progressive dilution. Some colleagues of mine had reported about this phenomenon too.

These cases make me think twice before I choose which value include into subsequent analysis. Sometimes there is a temptation to pick up higher values because matrix effect becomes less pronounced in highly diluted sample. But maybe I just want it to be higher while this is really a method mistake.

These thoughts make me return to the basics: why does this dependency exists?

It is important to "dilute" the matrix effect. You may have a look at the concept of "spike recovery" and the "linearity-of-dilution ".

These two articles below well explained the two concepts above. Have fun.

https://assets.fishersci.com/TFS-Assets/LSG/Application-Notes/TR0058-Spike-and-Recovery.pdf

https://cygnustechnologies.com/images/poor-dilutional-linearity.pdf 

Thank you, Li, for the comment! I'll have a look...

Dongyang, nice summaries on assay linearity and some good suggestions.

Ilya, your comment that you are seeing this issue at O.D. of 0.1-0.2 suggests some things to me.  You need to calculate limits of detection for your assay, and not go outside these limits.  The reason is that the variability of the calculated values will quickly rise outside these limits.  At O.D. values this low you are probably outside the limits of detection of your assay.  The actual limits are fairly arbitrary but generally not more than a 20% coefficient of variation in calculated values.  The spike and recovery experiments will help you find your limits. 

Thank you, James!

OD of 0.1-0.2 are still above limit of blank. In fact in negative control (i.e. dilution buffer) I usually have values less then 0.05. Samples depleted by specific MAbs are not distinguishable from the negative control. So, I suggest 0.1-0.2 are valid values and could be used for quantification.

I should apologize because I didn't specify the type of samples I had 0.1-0.2 values. They are urine. The figure above was drawn based on blood plasma samples which contain much more antigen compared to urine.

In urine of healthy volunteers I usually have values in range of 0.1-0.3. In samples of cancer patients I sometimes have higher values but in some cases I have that low ODs, but if I dilute them more they demonstrate the behavior I described above. Values are still within the range 0.1-0.25 but antigen content estimations rise because of higher dilution factor.

That's why I'm thinking about what values I should include into the analysis.

What are your dilution factors for urine samples? If the sample is highly diluted, matrix effects become less ad less important.If dilution factors are low, many strange effects can be expected.

 If some samples consistently produce the same absorbance (in the low range 0.1 and 0.2) at different dilutions  (resulting in increasing calculated concentratons when increasing the dilution factor), its clear that this absorbance range is not useful for concentration determination. You really need to estimate the useful range of your assay (that is not everything above blank) using several dilutions of samples with low, medium or high concentration. All dilutions of the same sample should produce similar concentrations values. If very low or very high dilutions clearly deviate from this behaviour, these dilutions are out of the useful range of the assay and should not be taken into account. Probably these aborbance values correspond to a high plateau or low plateau of the standard curve.

Thank you, Gertrudis, for your comment!

I'm very sorry I didn't write you back immediately. Just didn't have an opportunity to do this.

Urine samples are diluted 4 and more times.

I used the strategy you described before for plasma samples. Dilution 1:3 was optimal for normal urine. 1:1 and 1:9 gave pour results. So, 1:3 was optimal.

In this particular situation I decided to conduct systematic study and test a set of samples in different dilutions and explore different strategies of antigen concentration  calculation. Estimation of sensitivity and specificity for each approach will help to select one I need. That seems for me the most adequate way of solving this problem.

1/4 is a very low dilution factor. Matrix effects can be really important when using this procedure.  But you said that the samples were diluted hundreds of times.

You say that 1/3 is an optimal dilution. But what is important is to have an optimal range (several dilutions producing very similar values of calculated concentration). Otherwise, how do you know that the result at 1/3 is the best one? Do you have samples of known concentration to validate your method?

Thank you, Gertrudis, for your response!

I usually dilute blood plasma 1:100 or more (see the figure). The antigen of my interest is much less abundant in urine, so the dilution is 1:3.

On the figure you can see the dependence between dilution factor and estimated antigen content in blood plasma. Similar figures could be drawn for urine titration but they are less clear due to low antigen content.

I spiked antigen depleted blood plasma by the recombinant calibrator and detected 80-110% of antigen.

To be perfectly clear I attached a table with my last experiment. Two urine samples were tested in different dilutions. Antigen content was calculated.

So, the question is which values would you take into analysis?

This is difficult to say. In principle, it seems that the method is not good at all to quantify the antigen in urine, because there is something like a low plateau when diluting urine samples. Im sure you would need a more sensitive assay to properly study samples like the ones you  showed.

As there is a plateau, the less diluted samples should be more reliable, but matrix effects should also be more evident at such low dilutions.

In my opinion you need some urine samples with kown concentrations of the antigen (determined by other method or made my adding the antigen to antigen-depleted urine )to see wheher any dilution is giving you the right results.

Anyway, the best advice is to improve the sensitivity of the ELISA, for instance using an amplification system or hiher affinity antibodies...your current method is quite bad for urine samples.

Thank you, Gertrudis, for your thoughts!

I have to say that there is no better method then I use to quantify this antigen in urine. I have compared several ELISA systems and that I use is the best one. Some of ELISA kits do not detect the antigen in urine at all.

The antigen I work with is very unusual. It's molecules seems to be highly heterogeneous. So, different ELISA systems detect various amounts of the antigen in all types of samples (blood plasma, urine, culture media, etc.). The most pronounced differences (~100-fold) were obtained when blood plasma samples were tested. Using western blot analysis we showed that estimations made in the system I use were much closer to the total antigen content.

We now sure that the antigen presents in urine. However, it is much more difficult task to make an independent antigen content estimation in urine using western blot due to its very low concentration.

Surely, this is the right thing to add recombinant calibrator to antigen depleted urine and test it.

The data looks pretty reasonable.  Dilution range between 10x and 1000x is really wide for ELISA.  It is pushing the limit of this technology.  Those data points around the middle of the curve are close to each other indicating the reliable range of this ELISA assay.  Those data points at the extreme of both sides, such as below 1.5 and above 2.5, with higher variability indicated that the concentrations at those dilutions are at or near the assay plateau of either high or low end.  The dilution result provided a reference to adjust assay range with acceptable accuracy and precision.  The lower readings at low dilution factor is also an indication of hook effect, suggesting that the minimum required dilution should be increased to alleviate matrix effect. A parallelism test between recombinant calibrator and matrix with endogenous marker is the best way to demonstrate the similarity between the 2 molecues and valid the application of calibrator.

Thank you, Chien, for your recommendations!

I think your results are very typical. The less-diluted samples (higher concentrations) have depressed signals likely because of interference from other serum components and the possibility that large aggregates get washed-off the plate. The most-diluted samples (lower concentrations) have depressed signals because of mass-action considerations.

Thank you, Don, for you comment!

Dear Ilya,

I have often the same problem. And besides explanation we need solutions of course. I try to focus on that :-)

This matrix phenomenon can play an extreme role, depending on the antibody. When I test for antibodies to a protein on a microarray (with many other proteins) or in ELISA (with only one) I get sometimes completely different results. even if the target protein is really the same. That is of course most dramatic for autoantibodies, or targets with a high degree of tolerance in the mouse. One can try a polyadsorption step or include other proteins, so one might get a direction whether this might be relevant. but basically i never found a way to solve this. I fully agree with James Drummond on that, it drives you crazy.

Another important problem which can be solved is the "blocking effect" of the serum sample. If you block your plates with tween, albumin etc., sometimes they are not so well blocked. Addition of the samples serum will improve this incomplete blocking (because of lipids etc. in the serum). Therefore if you now dilute the sample just in PBS or even blocking buffer, this additional blocking will be reduced. and you get an increase in non-specific binding. The readings for diluted samples will therefore be higher as expected.

You can solve this by using different Block and dilution buffers. i had to play around with compounds for ages to finde solutions. Addition of a few % of Low MW PVP helps, also sugars like mannitol can be good. Higher ionic strength can reduce the non-specific binding, e.g. increase salt concentration...

With your urine samples this might be the case. They are lower as the blank, so maybe they block more.

another problem is with denaturation or changes of your target protein or antibody. An antibody is more stable, but at least in the labs where i have been, denaturated protein plays a major role, but most people are not aware of this.

if you use old samples, which have been stored at -80, often they are ok for maybe 2-3 years (pure serum). That is my experience. The more Chaos is in the -80, the longer people search stuff and leave the door open, which leads to an often increase in temperature. Like jumping from -80 to -70 or above. Most people think this doesnt matter if no sampley are thawing, but it does.

The tempature changes lead somehow to a change in the order of the crystallization of the sample (you need to ask a chemist, i dont understand why this happesn in detail). And this change denaturates proteins. Its comparable, but not as severe, as freeze thaw cycles. However, if you monitor it as I did, about 20-40% of the proteins get denaturated after 20 increases in temp over 10 °C. With a normal lab -80 which is opened every day, after 2 years some proteins and antibodies are in large parts denaturated, others are more stable. This has dramatic effects on the ELISA. The denaturated samples bind to anything, you get a high background and wells can be much lower as the blank due to this blocking effect. If you check a range of samples and blot age vs. signal, sometimes you get a good linear correlation. this depends on the ELISA u use of course.

my last important point is about the substrate and enzyme.

Sometimes the substrate gets used up very fast in the strong positive wells and with increased time, the lower conc. wells also give a stronger signal as they should. And with absorption there is a myriad of other possible problems. What is an idea, is to use once a reader with photomultiplier and use a luminescence or fluorescence substrate/ conjugate. This often solves problems with the old substrates for me...

Dear Oliver, thank you for increadibly detailed answer! Discussing this problem with collegues I realized how common this problem is and how low attention people pay to it. And this phenomenon could have a serious influence on decisons physicians make.

At least for now we accepted a rule of thumb that we test the same sample in at least in two different dilutions and estimate an agreement of the measurements. In the case of disagreement we retest the sample using higher dilutions. This procedure helps to identify samples with high antigen content. However, it may give incorrect results when sample has low antigen level (i.e. concentration gets higher when OD approaches to the level of negative control).

You are totally right that ELISA system needs fine tunnig for particular type of samples. In our case there are four types of samples (blood plasma, urine, cerebrospinal fluid, and cell culture media). So, I will reconsider your advice when I tune the system for each type of samples.

Once again, thank you for you advice!

Oliver Schmetzer and Ilya Smirnov . I am facing a similar problem and it is kind of a relief that people face a similar problem in terms of dilutions and platforms. Would you care to add if you know anything about a decent dilution range for LPS-stimulated whole blood where plasma will be used to detect IL-6 levels in regular ELISA to be comparable to a Luminex dataset. Redirecting to question:

https://www.researchgate.net/post/What_could_be_the_possible_reasons_for_low_correlation_between_ELISA_and_Luminex_for_the_same_cytokine_measures_in_human_plasma

Thank You in advance!

Dear Kalpita R Karan ,

This is quite a common problem for ELISA and other techniques. I was surprised to know that people do not pay much attention to it.

I have figure out a way how to determine appropriate dilution factor range in ELISA. I make a series of dilutions of sample and estimate antigen concentration in each well according to a standard. Then I draw a plot of dilution factor versus estimated concentration in log scales. Finally, I fit a regression line with slope of -1 and estimated intercept. Slope of -1 says that if you dilute sample twice, for example, antigen content also decreases twice. You can easily see which points are aligned with the regression line and which points are not (see the figure). Based on this plot you may easily figure out what dilution range is better for your samples. Usually this method works quite well, but some other factors could have a significant influence. So, more changes should be done before test will work appropriately.

Best wishes,

Ilya

P.S. Apologies for untidy figure. I just quickly got it our from one of my reports. Anyway, the idea could be clearly seen.

PPS. As I figured out from your question you try to compare results of two methods measuring the same factor. It is a common task and there is a bunch of statistical literature explaining how to do this. Just type "agreement of two methods statistics" in Google and you will find a lot of articles. This problem is worked out very well. You may find an example how to describe such a relationship in our most recent article entitled "New highly sensitive ELISA system for soluble endoglin quantification in different biological fluids".

Ilya Smirnov Thanks a lot for the dilution paradigm. I shall look at it.

My question was " reasons of disagreement" between two platforms concerning the same sample. I understand to some extent why and it is a mixture of reasons including dilutions in the right range. But I shall look at the "agreement methods".

Also, I donot access your article in university.

Dear Kalpita R Karan

Agreement and disagreement between two methods could be two sides of the same coin.

Statistical methods I mentioned could be used for assessing disagreement as well.

In out recent study we assessed agreement/disagreement between two ELISA methods measuring solube endoglin. It turned out that estimations made in out ELISA were 100 times higher compared to commercial kits. Nevertheless we found out estimates made in both systems agreed to each other but the scales were different.

Ilya Smirnov I will try agreement analysis once I have data on Luminex/ELISA. In my case Luminex measures were >100 fold higher in Luminex than a commercial ELISA (Duo set) and the scales were different as well as the systems disagree between both measures. Hence needs a proper dilution and a re-run.