General rule of thumb is that 35 cycles of PCR is sufficient to amplify even a single molecular of starting template, so Cq values of 30 or more are definitely approaching this territory.

The three main problems of measuring very poorly expressed transcripts are mispriming, stochastic effects, and buffer inhibition.

You've spotted the first, clearly (by running products on gels). This is common: if primers can't find the correct template to bind, then any mispriming that occurs, no matter how rare or unlikely, is subject to the exact same exponential amplification. 

The second manifests in wildly variable results due to stochastic partitioning: if your transcript is present at, say, only 3 molecules per ul on average, then any given 1ul aliquot could actually contain 5, 4, 3, 2, 1 or even zero molecules (or more than 5), giving you Cq values of 32.7, 33, 33.4, 34, 35 and (probably) 40: there's usually a sharp drop from 'one molecule Cq' to 'mispriming Cq'.

The third is a consequence of using too much cDNA in an effort to increase transcripts.well-1. It doesn't always work, as cDNA synthesis buffer can be quite inhibitory. You're already diluting yours, so this should be ok (hopefully).

So. First things first, make a standard curve. Get a PCR product you know for sure to be your actual CYP transcript, amplify it up, gel extract it, clean it, spec it, and work out (based on concentration and template length) the conc. in molecules.ul-1.

Then use this, both alone, AND mixed with a sample that seems wholly CYP-negative, at 100000, 10000, 1000, 100, 10, 1 molecules per well (or similar dilution series). INCLUDE MELT CURVES.

This will tell you if your reaction is still capable of working with very low transcript numbers, AND how your reaction behaves under experimental conditions (I generally find PCRs with pure template have lower Cq values than those mixed with lots of other cDNA).

You should now have a decent idea, by extrapolation, of the Cq value (for your PCR) that corresponds to a single molecule, AND what amplification of 'bugger all' looks like (esp at melt curve level). which should allow you to assess your current data.

If you have any samples that appear to carry workable numbers of transcripts (say, 10+ molecules.ul) then use one of THOSE in a dilution series. In an ideal world, using twice as much cDNA should lower Cq by 1, and half as much should increase it by 1. If you find doubling your cDNA does not affect Cq, you're either amplifying misprime noise (which is concentration independent) or using too much cDNA (such that increase in transcript number is offset by PCR inhibition). Doing this should thus allow you to determine what quantities of cDNA are permissible for your reaction.

Finally, once you've established ballpark figures for your transcripts.ul-1 in your samples, and determined what dilutions are suitable for linear measurement, you're good to go!

Your only remaining problem is stochastic partitioning: if you find your estimates for transcripts.ul-1 are falling within the stochastic range (anything under 20 transcripts.ul-1 or so), simply increase the number of replicates.

"3 transcripts.ul-1 on average" is difficult to accurately quantify using triplicate reactions, but if you run reactions in quintuplicate instead (and average the Cq values) the accuracy increases significantly. I personally find '4 reactions per sample' adequate for 10-20 range, and 5 or more for sub-10 levels. 

General rule of thumb is that 35 cycles of PCR is sufficient to amplify even a single molecular of starting template, so Cq values of 30 or more are definitely approaching this territory.

The three main problems of measuring very poorly expressed transcripts are mispriming, stochastic effects, and buffer inhibition.

You've spotted the first, clearly (by running products on gels). This is common: if primers can't find the correct template to bind, then any mispriming that occurs, no matter how rare or unlikely, is subject to the exact same exponential amplification. 

The second manifests in wildly variable results due to stochastic partitioning: if your transcript is present at, say, only 3 molecules per ul on average, then any given 1ul aliquot could actually contain 5, 4, 3, 2, 1 or even zero molecules (or more than 5), giving you Cq values of 32.7, 33, 33.4, 34, 35 and (probably) 40: there's usually a sharp drop from 'one molecule Cq' to 'mispriming Cq'.

The third is a consequence of using too much cDNA in an effort to increase transcripts.well-1. It doesn't always work, as cDNA synthesis buffer can be quite inhibitory. You're already diluting yours, so this should be ok (hopefully).

So. First things first, make a standard curve. Get a PCR product you know for sure to be your actual CYP transcript, amplify it up, gel extract it, clean it, spec it, and work out (based on concentration and template length) the conc. in molecules.ul-1.

Then use this, both alone, AND mixed with a sample that seems wholly CYP-negative, at 100000, 10000, 1000, 100, 10, 1 molecules per well (or similar dilution series). INCLUDE MELT CURVES.

This will tell you if your reaction is still capable of working with very low transcript numbers, AND how your reaction behaves under experimental conditions (I generally find PCRs with pure template have lower Cq values than those mixed with lots of other cDNA).

You should now have a decent idea, by extrapolation, of the Cq value (for your PCR) that corresponds to a single molecule, AND what amplification of 'bugger all' looks like (esp at melt curve level). which should allow you to assess your current data.

If you have any samples that appear to carry workable numbers of transcripts (say, 10+ molecules.ul) then use one of THOSE in a dilution series. In an ideal world, using twice as much cDNA should lower Cq by 1, and half as much should increase it by 1. If you find doubling your cDNA does not affect Cq, you're either amplifying misprime noise (which is concentration independent) or using too much cDNA (such that increase in transcript number is offset by PCR inhibition). Doing this should thus allow you to determine what quantities of cDNA are permissible for your reaction.

Finally, once you've established ballpark figures for your transcripts.ul-1 in your samples, and determined what dilutions are suitable for linear measurement, you're good to go!

Your only remaining problem is stochastic partitioning: if you find your estimates for transcripts.ul-1 are falling within the stochastic range (anything under 20 transcripts.ul-1 or so), simply increase the number of replicates.

"3 transcripts.ul-1 on average" is difficult to accurately quantify using triplicate reactions, but if you run reactions in quintuplicate instead (and average the Cq values) the accuracy increases significantly. I personally find '4 reactions per sample' adequate for 10-20 range, and 5 or more for sub-10 levels. 

Hi Adam,

Considering that you do a melting curve, I suppose you do a 'standard' SYBR-Green-based qPCR. Instead, you may want to consider using more sensitive (but, unfortunately, also more expensive) TaqMan / Molecular Beacon approaches.

The difference here, that in a SYBR-Green based qPCR the production of any dsDNA is measured, whereas in a TaqMan PCR the fluorescence is only generated when your specific DNA product is generated.

In my hands, we have been able to reliably and reproducibly detect leaky transcription from (extremely low expressed) repressed genes (dynamic range of 1*10^6 times less expression versus highly expressed control genes like GAPDH).

Validated TaqMan primer/probe sets are available for most human and mouse genes, and for many other organisms as well:

https://www.thermofisher.com/fr/fr/home/life-science/pcr/real-time-pcr/real-time-pcr-assays.html

These probes will work on most standard qPCR equipment.

The only other legit way to get a higher signal is to use a mRNA isolation kit (I suppose you are using Trizol now), and then take 1 ug of that RNA for RT. You can put 1-2 ul of RT directly into the RT-PCR reaction mix, there is no need to dilute it. But really if it is low, you can't do anything about it.

If you going to hunt for very low signals, everything in your reaction should be perfect because the lower you go the higher your noise gets. Make sure that your primers are on different exons, absolutely no dimers, optimal Tm, etc. Your primers should perform well in the dilution curve at very high dilutions. Usually it is not worth it as these data are very hard to get and even much harder to reproduce.

Read up on droplet digital PCR as well.  If your project is important enough, and quantifying low level transcripts is central to your mission, this investment may be worth it. Working in this nether-region of the qPCR assay is a touchy sport - for all of the reasons already mentioned by others above. Bear in mind, studies have been done wherein a 70%-efficient PCR reaction can still generate specific product at cycle ~48... (e.g., see: Lockey C, Otto E, Long Z: Real-time fluorescence detection of a single DNA molecule. BioTechniques. 1998; 24:744-746.)

So, if you want to fish, the possibilities are there. Sequencing final (suspected specific/intended) products is always an option.