Hi,

did you do LFQ quantitation? I had a similar strange volcano shape when using the latest MaxQuant Version. I checked the distribution of all LFQ values for each raw file, normally when everything works they should have a similar distribution. When using the latest MaxQuant Version there were significant changes in the distribution indicating that somehow the LFQ normalization didn´t work. I solved the problem by going back to an older Version of MaxQuant and everything was fine again, LFQ distribution was similiar in all files and the volcano had a normal appearance too.

Hopefully that was helpful,

best

Sebastian

That was very helpful, thank you Sebastian. And yes I DID do LFQ quantification. I will look to see if any older version exists!

Hi Jackie,

can I ask what you mean with invalid values?

I am not sure, but to help you with this problem you need to provide some more info on how you transformed your data in Perseus.

A usual start will begin with log2 transformation of the data and after this you can define your groups and filter for valid values (depending on your number of replicates and whow much replicates with LFQ-Intensities you will at least accept as a valid value (i.e. for 3 replicates (biological) per group I would define at least 2 in at least one group). After this you can replace "missing values" (NaN) by imputation (imputation by normal distribution with downshift or imputation with a constant).

Instead of converting all missing values to a 0, I would recommend to look for the lowest value in the entire dataset (somewhat like the detection limit) and would use this value as a constant to impute all missing values (if you really dont feel well by normal distribution imputation).

Be aware that if you have a missing values for all replicates of one group and the other group has valid values the imputation with a constant to this missing values with lead to no variance in this group whereas the group to compare has a variance. In this case you are not allowed to work with the "usual" t-test which requires equal variances of the groups to compare. If the variances are unequal you will need to do a welch's test.

Best,

Murat

Murat Eravci Thats a great point, I never considered using a non-parametric analysis for this data.

So how I did this analysis was I first log2 transformed my data. Then i filtered on proteins that have 2 values in at least 1 group (I have 6 groups total, but am only performing analysis on a small subset for now, which is why I was using T-tests rather than anova). After this filter, I tried multiple ways of imputation. First I tried imputing based on normal distribution, but found a few instances where significant differences were not identified where they should be, and where is was very obvious based on the raw data (ie one group had no values in any 3 replicates and the other group have higher expression). So i moved to converting all NaN values, that were originally 0 before log transformation, back to 0.

I am not a proteomics expert, nor does my lab really do many proteomics so I cannot confidently say when I am doing is 100 percent optimal.

So any advice will be greatly appreciated.

Thank you for your reply,

Jackie

Jackie Rich For log transformed data you should use a parametric test. Welch's test is actually a parametric test. It is similar to the t-test but allows comparisons of groups with unequal distribution. You can use the welch's test option in Perseus aswell.

The other point is, if your raw data was zero and you have converted the missing values (NaN) after Log2 transformation to 0, it would mean that you have not put them back to 0 instead you have converted them all to 1 (Log2 of 1 = 0). That's the reason why you have NaN values for all zero's after Log transformation because you cannot log transform 0.

Coming back to the imputation by normal distribution, you are also able to increase the downshift of the imputed values to get a larger difference. However I have realized that in some cases (when working with more than 2 groups) the imputation from normal distribution can randomly produce significant results (i.e. from two groups were all values where missing). For this reason I am using more often the imputation with a constant for which I take the lowest value in my present dataset.

Feel free to ask any questions in relation to proteomics and I will try to help you.

Best,

Murat