[pmiguel]

No I believe if you read into my argument a little more it boils down to that 36bp are most like too short for YOUR application therefore if you use 36bp I am going to write you off Sure there are applications for 36mers out there maybe siRNA or something.

Notwithstanding the evidence presented by HESmith? I guess you are just trolling.

How about an admission that you overstated your position on 36mers?

I would ask ECO to ban you for trolling. But you have a reasonable point, you are just exaggerating it.

--
Phillip

[rskr]

Notwithstanding the evidence presented by HESmith? I guess you are just trolling.

How about an admission that you overstated your position on 36mers?

I would ask ECO to ban you for trolling. But you have a reasonable point, you are just exaggerating it.

--
Phillip

Nope, I have worked with 36mers, not because I designed the experiment, but because someone who has never seen sequence data designed experiments, and I really don't like the results, especially when they try to compare them with real data. Besides these data-sets have a life of their own. Sure it may work for that one application, but then someone is like, "Hey can't you throw it into an assembler" or "Hey lets map it to the human genome", well if you sequence 36mers no, but people do it anyway, then report the results in journals, as if they were significant.

[frozenlyse]

Nope, I have worked with 36mers, not because I designed the experiment, but because someone who has never seen sequence data designed experiments, and I really don't like the results, especially when they try to compare them with real data. Besides these data-sets have a life of their own. Sure it may work for that one application, but then someone is like, "Hey can't you throw it into an assembler" or "Hey lets map it to the human genome", well if you sequence 36mers no, but people do it anyway, then report the results in journals, as if they were significant.

Seriously, why would I want much longer than 36bp for a ChIP or MeDIP-seq? The return on investment of longer reads is pretty poor.

[rskr]

Seriously, why would I want much longer than 36bp for a ChIP or MeDIP-seq? The return on investment of longer reads is pretty poor.

Just so you can see how bad the data is All I am saying is you can't do much with the 36mer data after it is collected, if your DNA extraction and library construction isn't very good, last I checked those two had very low yields at the extraction step leading to highly clonal coverage, as well as bias towards short inserts. Add on top of it that you won't be able to accurately map many of the reads because you were too ChIeP to pay for the last 64 bases of reagents, and well, you might get lucky on Uncle SAMs dime.

[frozenlyse]

Well at least now it's completely obvious to everyone you have no idea what you're talking about, so we can all safely ignore you.

[rskr]

Well at least now it's completely obvious to everyone you have no idea what you're talking about, so we can all safely ignore you.

What'd I say Is my personality to not sufficiently magnetic to precipitate voluminous amounts of sequence conversation? Don't go blowing my, curt statements out of proportion now.

[chadn737]

I have an older data set of 42bp reads ~7 million reads per sample (back in '09). I have a newer data set of 100bp reads ~35 million reads per sample (recent). They are from the same genetic background and environmental conditions, hence biological replicates of each other.

Looking just at differential expression (all that I am interested in), the results are almost identical. The more recent data set gives me more differentially expressed genes, but this increase can be attributed entirely to the number of reads.

Saying that all the shorter length data out there is worthless is an exaggeration.

[rskr]

I have an older data set of 42bp reads ~7 million reads per sample (back in '09). I have a newer data set of 100bp reads ~35 million reads per sample (recent). They are from the same genetic background and environmental conditions, hence biological replicates of each other.

Looking just at differential expression (all that I am interested in), the results are almost identical. The more recent data set gives me more differentially expressed genes, but this increase can be attributed entirely to the number of reads.

Saying that all the shorter length data out there is worthless is an exaggeration.

Hmm, 42mers didn't give you very many differentially expressed genes, you say? It couldn't be because they were that much more ambiguous, could it?

[chadn737]

Hmm, 42mers didn't give you very many differentially expressed genes, you say? It couldn't be because they were that much more ambiguous, could it?

I didn't say that. 42mers gave me several hundred differentially expressed genes. We originally did the 42mer data as a pilot experiment and it worked so well, supporting previous findings, that we went on to do a more extensive experiment.

Furthermore, if you look at the genes that are called in the 100bp data that aren't in the 42bp data, its pretty obvious that this is because of sequencing depth and increased number of biological replicates. These tend to be genes of 100 or under reads mapping to them. Remember, there is a nearly 5 times as many total reads in the 100mer data as in the 42mer data and for purposes of differential expression, that makes a big difference for lowly expressed genes.

I see very little significant difference in the number of uniquely mapped reads between the two data sets. In fact ~90.9% of the 42mer mapped reads mapped uniquely versus ~89.9% of the 100 mers. A previous poster noted that they only got a ~1% increase in uniquely mapped reads using longer lengths. I don't think it is coincidence that both of us observe only ~1% difference in the unique mapping of long and short read lengths. That 1% difference is not all that significant, certainly not to the extent that you are claiming.

[rskr]

mapped uniquely versus ~89.9% of the 100 mers. A previous poster noted that they only got a ~1% increase in uniquely mapped reads using longer lengths. I don't think it is coincidence that both of us observe only ~1% difference in the unique mapping of long and short read lengths. That 1% difference is not all that significant, certainly not to the extent that you are claiming.

The question left unanswered by the previous poster also casts doubt on your study. Simply unique, does not mean that the alignments were accurate. Sure there maybe a scoring scheme which would yield alignments that could be ranked such that one alignment is better than another, but is there a significant difference between that marginally better alignment. If you have 36mers, the probability that you really can't tell the difference between the next best thing is pretty good. So for example judging a 36mer alignment with a two base gap vs. an alignment with a gap and a mismatch, one will likely score slightly higher and therefore be unique, but the evidence is still underwhelming IMO.

[chadn737]

The question left unanswered by the previous poster also casts doubt on your study. Simply unique, does not mean that the alignments were accurate. Sure there maybe a scoring scheme which would yield alignments that could be ranked such that one alignment is better than another, but is there a significant difference between that marginally better alignment. If you have 36mers, the probability that you really can't tell the difference between the next best thing is pretty good. So for example judging a 36mer alignment with a two base gap vs. an alignment with a gap and a mismatch, one will likely score slightly higher and therefore be unique, but the evidence is still underwhelming IMO.

I don't think anyone is denying that there such problems do exist, but they also exist with 100mers as well. But that doesn't make 36mer data completely worthless. If the alignments for shorter reads were really so significantly worse, as you suggest, then there should be very poor correlation between the results I get with 42mers versus 100mers. But really the two correlate very well, giving me near identical results. I have also validated my results by qRT-PCR and it supports previous findings. If the 42mer data was so pathetically unreliable, then I guess you are saying that its a complete coincidence that it matches everything. Me, I go with the facts of what I have done. The 42mer data works, it can be replicated, it can be validated by other experiments. So saying it is worthless amounts to hyperbole.

[rskr]

what I have done.

Well, what you have done is said 42mers were as good as 100mers, but not refuted the problems with alignments or bench marked your alignments for accuracy, so there is a problem with your reasoning.

It is kind of like in statistics before you can use Anova, you have to prove that the you have enough data to calculate the mean and enough data to calculate the variance.

Here you have claimed that you have mapped 89% of the reads, but you haven't proved that even 1% of the reads were most likely mapped accurately. Last I checked even BWA didn't get over 80% accuracy with 100mers. I didn't bother to check for 36mers, because it can't be any better.

[frozenlyse]

Well, what you have done is said 42mers were as good as 100mers, but not refuted the problems with alignments or bench marked your alignments for accuracy, so there is a problem with your reasoning.

No, he's done better than that, validated the results.

[rskr]

No, he's done better than that, validated the results.

A stopped clock is right twice a day. Just because his analysis pipeline has a bias and yields the same result every time, doesn't mean it is validated. You really need to bench mark the alignment algorithm.

[frozenlyse]

I have also validated my results by qRT-PCR and it supports previous findings.

Obviously reading comprehension isn't your strong suit.

You really need to bench mark the alignment algorithm.

Benchmarking alignment algorithms was an interesting and exciting field in 2008 - in 2011, not so much.