Hacker News Reader

Comment by koeng | Hacker News Reader

The most annoying part with the 200k polymerase prize is that they don't actually provide protocols for validating the polymerase works well. Me and some friends can definitely do the AI-protein design + DNA synthesis + protein synthesis (and would want to), but no way am I gonna figure out idiosyncratic sequencer protocols.

View on HN

The lack of a validation protocol is also an issue for the another way for doing this — generationally mutate the polymerase and let evolution do the work for you. But, this only really works when you have an automated test to know which variations are working better than others. You really want to be able to automate the process to get many (many) generations to effectively explore the search space.

But, the contest isn’t really about length — it’s about polymerase preference for fluorescently labeled dNTPs vs unlabeled.

From: https://454.bio/blog/2024/01/23/polymerase-competition/

> 454 Bio have opened a competition with a prize of $200,000.00 to evolve a DNA polymerase optimized to use Lightning Terminators™ while not incorporating their hydroxymethyl cleavage products with a minimum 1000:1 ratio.

This seems to be an issue with their chosen “One Pot” reaction mixture. When you can’t wash away byproducts after each cycle, with this method, you get a build up of non-labeled molecules. And these are the ones that the polymerase prefers. So, of course they end up with single-digit read lengths… the polymerase is doing what it does best. This seems like figuring out this chemistry should have been fleshed out before trying to build the sequencer.

So, I guess it turns out that there is a validation endpoint, but still not a very convenient way to measure it…

Yeah, now that you mention it, that seems like a major issue. Of course unlabeled nucleotides will outcompete labeled ones, they're smaller so they get less steric hindrance. And I don't understand what can be done about it at the polymerase binding site. Maybe some induced proximity chemistry?

Right? The unlabeled nucleotide will still fit in a binding pocket for the labeled version. And it should always have better binding.

But this is an obvious issue. With this type of photo-reaction, you’ll get a build up of unlabeled nucleotides. There is still some potential for the technology, but it’s really hard for me to see how this works in one tube without some kind of washing (which also increases costs by a lot).

Maybe there is a way to add multiple moieties to the NTPs, such that removal of the fluorophore would cause a bigger hindrance? (Fluorophore + extra-moiety would bind, but extra-moiety alone would have more trouble binding?)

"And it should always have better binding." This is a very strong statement. There are (at least theoretically) a myriad of ways to redesign the binding pocket to preferentially bind the larger labeled nucleotide. Consider many proteases excel at binding the larger substrate peptide and having essentially no detectable binding to the resulting smaller cleaved peptide(s).

Strong, yes. It’s certainly possible to have binding sites that prefer larger molecules. However, I’m not sure you’d be able to do so while still being able to function as a polymerase. That’s the thing I’m thinking about. You have to have an enzyme that will preferentially bind a labeled nucleotide, while still being a good polymerase. And because the resulting enzyme should still be a good polymerase, it would likely also maintain good binding affinity for unlabeled nucleotides.

I’ll mention that your example of a protease is a good counter example — the function of the protein is to bind the larger peptide and cleave it. The evolutionary pressure was to bind the larger molecule and release the smaller (cleaved) ones. Here, the evolutionary pressure was to bind the smaller (unlabeled) nucleotide to build the DNA strand. This is a complex reaction that involves a lot of ligand binding and intermediates. Trying to add a function to preferentially bind a larger fluorophore labeled base seems like it would be overly disruptive or at least reduce the rate of polymerization.

I don’t know enough about the structure of DNA polymerase to say how feasible this really is — but the extra bulk of the fluorophore seems like it would cause issues.

But, is it possible? Sure! This is biology — there is always a way (or three)! I’d expect for there to be a way to generate a polymerase that does this via mutagenesis. But it would be very difficult and probably more expensive than the prize.

If they were able to keep the concentration of fluorescently labeled nucleotides higher, this wouldn’t be an issue. But due to the technology choices (one tube reactions, laser release), they are stuck trying to get a better polymerase.

I certainly wish them luck, I’d love for this to work for more than 5-6 bp.

> AI-protein design

Any recommendations on learning how to do this? As well as the limitations to your approach? I've been interested in engineering site-specific DNA binding proteins for over a decade now, but the tools I've looked at are specific for protein to small-molecule or protein to protein interfaces.

If you’re up for reading a dissertation I have one in mind I can dig up. I will check back here in 12 hours or so or you can contact me from my profile.

Would also be interested in this if you'd care to post publicly.

In case you missed it: https://news.ycombinator.com/item?id=39283253

That sounds great.

It turns out the dissertation is not quite published yet, but this paper [1] will be a part of it. Sam and I were in the same lab and his work has evolved to focus even more on AI-guided protein design. Perhaps the quote from the abstract that made the connection for me to this post:

> We applied the protein G B1 domain (GB1) models to design a sequence that binds to immunoglobulin G with substantially higher affinity than wild-type GB1.

[1] https://www.pnas.org/doi/pdf/10.1073/pnas.2104878118?downloa...

Thank you very much!

Reply on news.ycombinator.com