Hi Iddo, Rana, 

I'll give my long winded perspective on this.

For starters Iddo - this is a question directly related to your experimental objectives and should be considered as a core part of your experimental design. The short answer is that if you already have the infrastructure to perform qPCR (which I'm assuming you do at the Weizmann) and you are strictly interested in a single or just a couple of genes, and if you already have the reagents, then order some primers and give qPCR a go. Honestly, this tech is on out, but you should do a quick cost analysis considering your objectives. Calculate the cost of all the reagents considering your replicates, screw-ups, training, how many genes you're interested in. You'll need to consider as well both the cost of the tissue prep, the cost of the reverse-transcriptase (and kit), an the actual qPCR reagents. Chances are, after you get up past maybe 5 genes, you'll start to get close to the cost of performing RNA-seq, which you may be surprised to find is actually quite inexpensive if you outsource it to a major institute. There are facilities in China (BGI), australia, the US, etc that are worth googling for.

In your situation, I would forget qPCR and go straight for RNA-seq. You're interested in a lot of genes - though its not clear whether or not you are trying to compare samples in your experiments or if you are looking perform a low level descriptive analysis. A pertinent question is: What have you done to normalize your samples? Without more information, I wonder if this is where your difference is coming from. I'll explain below, but before I do, I'll qualify all of this by saying - as long as you are consistent in your preparation of the data, its probably not that big of an issue. The differences in the amounts between your sample prep kits doesn't seem to be that great and you're still detecting a great deal of miRNAs.

As for the normalization step and the use of RNA-seq for absolute and relative quantification: 

With RNA-seq,there are essentially two kinds of normalization - within sample (CPM, TPM, RPKM - for single read data, FPKM - for paired end data, etc..)  - and between sample (TMM, etc... probably don't even really need to mention any others). The thing to remember is that within sample normalization (preferably CPM...) will provide you with a measure of absolute quantification. In this sense, Rana is correct.

Within sample normalization corrects for a problem that arises from the fact that there is a linear relationship between the number of transcripts detected during sequencing and the actual length of a gene. I.e. For two genes A and B with the absolute ground truth of expressing the same number of transcripts, if A is twice as long as B, you can expect to detect twice as many transcripts. Within sample normalization therefore allows you to compare expression values of genes within a given sample library, so if you wanted to compare two miRNAs from the same sample. This measure does NOT allow you to compare expression values between libraries, even though you'll see a lot  of presentations where people do make the comparison. (You should raise your hand and question them on this if this comparison is made.)

This is because there is always a sequencing bias effect where the subspace available for a  given gene or group of genes within the sample is limited by the quantity of other RNAs being sequenced. This is where between-sample normalization comes in. When comparing across samples, you need to correct for this bias - this is the prime function of TMM normalization. This method allows you to directly compare values of the SAME gene across multiple different samples. In this method, the length of the gene isn't corrected for since the length of the gene is constant across all samples. But... you lose the ability to compare different genes across samples. The idea of coupling within sample normalization to between sample normalization is an area of active research. 

So in summary, if you are comparing expression from a single sample, Rana is correct - you can use within sample normalization to get absolute quantification values. If you are comparing between sample libraries (e.g. libraries produced from different kits), then Rana is incorrect - you should use between sample normalization looking at relative amounts between samples. Again, don't mix the two - subspace sequencing bias precludes the direct comparison of absolute quantification of expression between samples. Its almost like the uncertainty principle. You can measure 'absolute' quantification within a sample without knowing where it sits next to a different sample, or you can measure relative quatification across samples without knowing where that gene sits relative to another gene.

With all of this in mind back to your original question - is the  difference you see in your measurements explainable by normalization? Well, the problem is that you can't really directly compare all of the parameters that you'd like to compare across your experiment. So the best thing to do is to just use the kit that gives reasonable and consistent proportions of miRNA's - are you detecting the miRNAs that you are meant to detect using the kit? If you want to try and validate this way using qPCR, then you should  aim to use a qPCR kit that is specifically designed for the type of genes you're interested. 

Don't bother trying to correlate qPCR measurements to RNA-seq. There is just no way to do this reasonably. Rana is right - this is strictly a relative quantification. If you are interested in absolute quantification by PCR, then look in to digital droplet PCR (ddPCR). This is a relatively new tech, and its probably still pricey - especially if you don't own or have access to the machine yet. Which points you right back to outsourcing RNA-seq.

You  should be able to reason about your questions with all of this information.

What if I have just one sample that I want to compare between two methods? Shall I look at differential expression (DE) between miRNAs rather than experimental groups?

Maybe, if you mean between kit libraries, but first consider your other question -

Which one of the profiles is more in agreement with what is reported in the literature about serum miRNAs?

This is domain specific knowledge that you need to invest time in if you're performing research on the topic. With this information, you can have a better sense of whether or not you  are seeing reasonable proportions of genes when looking at within-sample absolute quantification measurements in a given sample. 

The DE measurements within a sample aren't informative - you  are better off making the same pie charts you've made. The DE measurements (using between sample TMM normalization) will only give you an idea of how different your kits are.

Overall, this is why is best to just pick on kit that matches your domain knowledge, is easy to use, is reliable and consistent, and use that to perform your analyses.

I hope this was helpful and stimulated some thought. There may be more up to date information on some of this, so if you find updated information on any of this, please share it!

If anything needs clarification (on your end... or mine!) don't hesitate to continue the conversation.

Cheers,

Paul

Hi,

You have always to keep in mind that RNA-Seq data represents absolute acount, while qPCR data represent relative gene amount to your housekeeping gene. So, with that in mind the correlation would be just about being upregulated, downregulated, or no effect without comparing numbers. 

but what if I have just one sample that I want to compare between two methods? shall I look at differential expression between miRNAs rather than experimental groups?

Another question: which one of the profiles is more in agreement with what is reported in the literature about serum miRNAs?

Hi Iddo, Rana, 

I'll give my long winded perspective on this.

For starters Iddo - this is a question directly related to your experimental objectives and should be considered as a core part of your experimental design. The short answer is that if you already have the infrastructure to perform qPCR (which I'm assuming you do at the Weizmann) and you are strictly interested in a single or just a couple of genes, and if you already have the reagents, then order some primers and give qPCR a go. Honestly, this tech is on out, but you should do a quick cost analysis considering your objectives. Calculate the cost of all the reagents considering your replicates, screw-ups, training, how many genes you're interested in. You'll need to consider as well both the cost of the tissue prep, the cost of the reverse-transcriptase (and kit), an the actual qPCR reagents. Chances are, after you get up past maybe 5 genes, you'll start to get close to the cost of performing RNA-seq, which you may be surprised to find is actually quite inexpensive if you outsource it to a major institute. There are facilities in China (BGI), australia, the US, etc that are worth googling for.

In your situation, I would forget qPCR and go straight for RNA-seq. You're interested in a lot of genes - though its not clear whether or not you are trying to compare samples in your experiments or if you are looking perform a low level descriptive analysis. A pertinent question is: What have you done to normalize your samples? Without more information, I wonder if this is where your difference is coming from. I'll explain below, but before I do, I'll qualify all of this by saying - as long as you are consistent in your preparation of the data, its probably not that big of an issue. The differences in the amounts between your sample prep kits doesn't seem to be that great and you're still detecting a great deal of miRNAs.

As for the normalization step and the use of RNA-seq for absolute and relative quantification: 

With RNA-seq,there are essentially two kinds of normalization - within sample (CPM, TPM, RPKM - for single read data, FPKM - for paired end data, etc..)  - and between sample (TMM, etc... probably don't even really need to mention any others). The thing to remember is that within sample normalization (preferably CPM...) will provide you with a measure of absolute quantification. In this sense, Rana is correct.

Within sample normalization corrects for a problem that arises from the fact that there is a linear relationship between the number of transcripts detected during sequencing and the actual length of a gene. I.e. For two genes A and B with the absolute ground truth of expressing the same number of transcripts, if A is twice as long as B, you can expect to detect twice as many transcripts. Within sample normalization therefore allows you to compare expression values of genes within a given sample library, so if you wanted to compare two miRNAs from the same sample. This measure does NOT allow you to compare expression values between libraries, even though you'll see a lot  of presentations where people do make the comparison. (You should raise your hand and question them on this if this comparison is made.)

This is because there is always a sequencing bias effect where the subspace available for a  given gene or group of genes within the sample is limited by the quantity of other RNAs being sequenced. This is where between-sample normalization comes in. When comparing across samples, you need to correct for this bias - this is the prime function of TMM normalization. This method allows you to directly compare values of the SAME gene across multiple different samples. In this method, the length of the gene isn't corrected for since the length of the gene is constant across all samples. But... you lose the ability to compare different genes across samples. The idea of coupling within sample normalization to between sample normalization is an area of active research. 

So in summary, if you are comparing expression from a single sample, Rana is correct - you can use within sample normalization to get absolute quantification values. If you are comparing between sample libraries (e.g. libraries produced from different kits), then Rana is incorrect - you should use between sample normalization looking at relative amounts between samples. Again, don't mix the two - subspace sequencing bias precludes the direct comparison of absolute quantification of expression between samples. Its almost like the uncertainty principle. You can measure 'absolute' quantification within a sample without knowing where it sits next to a different sample, or you can measure relative quatification across samples without knowing where that gene sits relative to another gene.

With all of this in mind back to your original question - is the  difference you see in your measurements explainable by normalization? Well, the problem is that you can't really directly compare all of the parameters that you'd like to compare across your experiment. So the best thing to do is to just use the kit that gives reasonable and consistent proportions of miRNA's - are you detecting the miRNAs that you are meant to detect using the kit? If you want to try and validate this way using qPCR, then you should  aim to use a qPCR kit that is specifically designed for the type of genes you're interested. 

Don't bother trying to correlate qPCR measurements to RNA-seq. There is just no way to do this reasonably. Rana is right - this is strictly a relative quantification. If you are interested in absolute quantification by PCR, then look in to digital droplet PCR (ddPCR). This is a relatively new tech, and its probably still pricey - especially if you don't own or have access to the machine yet. Which points you right back to outsourcing RNA-seq.

You  should be able to reason about your questions with all of this information.

What if I have just one sample that I want to compare between two methods? Shall I look at differential expression (DE) between miRNAs rather than experimental groups?

Maybe, if you mean between kit libraries, but first consider your other question -

Which one of the profiles is more in agreement with what is reported in the literature about serum miRNAs?

This is domain specific knowledge that you need to invest time in if you're performing research on the topic. With this information, you can have a better sense of whether or not you  are seeing reasonable proportions of genes when looking at within-sample absolute quantification measurements in a given sample. 

The DE measurements within a sample aren't informative - you  are better off making the same pie charts you've made. The DE measurements (using between sample TMM normalization) will only give you an idea of how different your kits are.

Overall, this is why is best to just pick on kit that matches your domain knowledge, is easy to use, is reliable and consistent, and use that to perform your analyses.

I hope this was helpful and stimulated some thought. There may be more up to date information on some of this, so if you find updated information on any of this, please share it!

If anything needs clarification (on your end... or mine!) don't hesitate to continue the conversation.

Cheers,

Paul