Koch Lab:Publications/Proof of principle for shotgun DNA mapping by unzipping

January 19th rough draft

• Document available on Scribd
• Comments about draft also on our Koch Lab blog and Steve Koch's Research Blog.

Comments

Richard C. Yeh 20:44, 20 January 2009 (EST) First, congratulations on your new paper draft! Steve Koch 17:10, 21 January 2009 (EST): Thank you! I'm going to respond to some of your comments below, and then I expect eventually they will be moved to sub-pages, since they are really good ideas warranting discussion and further data.

My understanding of your paper is that you have shown that you can identify a naked dsDNA sequence by comparing its unzipping force-index curve against a library of simulated curves. Very cool, and I want to thank your team for writing the paper in a way that I can understand. Here are my comments:

1. I think the major contribution of the paper is that you have an accurate simulation and matching method, which can easily account for elasticity of the tethering construct, viscosity, etc. I'd like to see more details (or at least know that you have the answers in your back pocket) on the information-recall side of the problem. See terms at the bottom of < http://en.wikipedia.org/wiki/Sensitivity_and_sp... >. I'm no expert here, but here are my thoughts (and most of the answers don't necessarily belong in the paper):

You matched experimental F-j curve vs simulated F-j curve. Why not interpret your experimental F-j curve into a sequence (with wild-card characters) and then do the matching in ASCII space? Is there a rough map between your match score and a BLAST similarity score? (But, since I'm no BLAST expert, I would need to do lots more research before talking about BLAST in the paper.)

In the "Future Improvements" section, you discuss the general sequencing-by-unzipping problem. Imagine simulating all possible sequences from j=1200 to j=1700. In this space of 4^500 sequences, what are the characteristics of those close to your experimental sequence, and what is the minimum similarity that can be resolved? From the other side, in your matching problem, how much easier has the problem been made by considering only the 2700 restriction fragments instead of the 4^500 general genome? What if you don't restrict yourself to those restriction fragments, but allow any contiguous 500-bp sequence from yeast --- then how many false positives do you get? (In Figure 4, what is the outline that you would get by simulating random fragments from the 4^500 or yeast-sequence universe?) A related question is characterizing the sequences in the overlap of the Gaussians.

Suppose when you unzip chromatin, you get only a naked-DNA signal over the linker DNA regions, and an un-simulated signal over the non-linker regions. Would you speculate on how that (change in signal) will affect your recall?

Suppose you simulate all 2700 restriction fragments. What if you make another version of figure 4 by scoring all simulated sequences versus each other (sim vs sim instead of expt vs sim)? (Also, by symmetry, what is the spectrum of scores obtained for individual OT runs --- both match and mismatch --- against each other (expt vs expt, instead of expt vs sim)?) Is the difference in distance between your blue and red peaks roughly equal to the difference in distance between correct and incorrect peaks for the 2700 x 2700 test? (If not, then you will have to explain what's special about your 32 sequences.) Do the 2700 fragments include the reverse-complement sequences?

2. In the excerpt of the match score formula, I see kT / C / stdev(force difference). Here are the implications that I read from the formula: a. for a given match score, a 1-unit force difference on 4 indices (neighboring or separate) is worth a 2-unit force difference on 1 index; and b. for a given match score, the standard deviation of the force difference scales with the temperature of the environment.

On (a), I recognize that this is a draft and see that you may not want or need to refine the formula.

On (b), are you making a statistical-mechanical statement (and I see no support for it elsewhere in the paper), or are you just dressing up the standard deviation of force in statistical-mechanical clothing, by adding a kT and making it unitless? For the purposes of your experiment, kT/C is constant, so that portion of the relationship cannot be tested. My thought is, if you're not making a statistical-mechanical statement, then your match score should be as simple as possible:

m2 = stdev(force difference/scale force)

where the scale force is kT/2/C. Do you actually get a better Gaussian when you take exp(-1/m2)?

My own biases would reverse your scale, suggesting numbers close to 1 as a good match, and numbers close to zero as a poor match --- something that can be interpreted as a probability, like exp(-m2) / sum_{all sequences} exp(-m2). On second thought, perhaps it is good that the score does not change as we consider a different universe of sequences.

3. The beginning of the paper reads to me like "peeing on the tree". In particular, I think mentioning Pol II is overreaching.

4. It really bothers me in figure 2 that the colors are swapped between A and B. Also, next to figure 4, in the last paragraph before "Future Improvements", you use the word "fell" instead of "felt". Also, I have not worked on the language very much, but the last line on page 8 is a little strange.

5. I'd like to be able to estimate what the 30-point smoothing window represents (in nm or bp) when multiplied with your stretch rate.

6. You mention viscosity a few times and say that it's significant, but it seems to be absent from the "Creation of Simulation Library" description. However, later, you say that the viscosity is not important because it would just impose the same shift on the scores of both the matches and the mismatches (by affecting the OT data but not the simulated data). It's not clear to me how the GC content affects your estimation --- is it that the higher force needed to unzip improves the signal compared to the viscosity offset? The nature of standard deviation is such that if you had a mismatched sequence modeled with the correct force scale, it might have a better score than a matched sequence with the incorrect force scale. Alternatively, if you colored the points in figure 3 by GC content, would all those with the highest GC content have the lowest (most similar) match scores?

Initially, I felt that there was a contradiction here, but now I think that the argument needs to be revised: you depend on the scale of the force to be correct so that you can use the standard-deviation-of-difference in your match score. But if your simulated force doesn't include viscosity, then we cannot trust the energy calculation in your simulation.

Richard C. Yeh 20:44, 20 January 2009 (EST)