Hacker News Reader

Comment by dekhn | Hacker News Reader

It's powerful, and a definite step up over crystallography in some ways, but personally I think structural biology has repeatedly been oversold as a method to understand the function of proteins at the molecular level. I say that as somebody who obtained a PhD in NMR spectroscopy and worked for 15 years on structural biology problems.

I've come to see SB as a visualization tool that helps form hypotheses, but not an actual method by which most protein structural/function problems can be directly solved.

The techniques which seem to provide the most utility are fluorescent imaging with light microscopy, often at low magnification. These are more amenable to scale-up and machine learning, both of which are likely to be necessary to unlock the next level of understanding biology.

View on HN

I think your criticism of structural biology can be applied to any methods-focused investigation, as opposed to biology-focused. I've certainly been down the "okay we have the data, what does it mean?" path more times than I care to remember. Unfortunately I think this is a natural consequence of working on cutting-edge methods - structural biology has traditionally required a degree of over-specialization.

Sharing the OP’s position, and having worked in a biochemical technology development lab in grad school, I don’t think that’s quite right. Structural biology is disproportionately seductive as a data source because we are such visual creatures. You “see” it so it must be true. I’ve definitely been guilty of it while looking at protein structures, I doubt anyone in the field hasn’t. Hell there used to be 3D rigs you could buy with IR activated LED shutter glasses so you could see structures in 3D on your monitor and “better understand” the structure.

It’s easy to get misled, proteins are incredibly dynamic in solution. Binding pockets appear and disappear, change shape, add and remove water. Given a structure of an unknown enzyme, with little to no biochemical characterization, you can’t tell what the catalytic mechanism or rate is. You probably can’t tell what the crucial residues are for binding at all. You may be able to tell the binding pocket, and... generate some hypotheses!

That’s not to say protein structures aren’t helpful, they are! Especially things like co-crystals with drugs can help study that specific interaction. But I think the OP is totally right, we should back off strong conclusions on structure alone. It’s a great hypothesis generator, not a final answer in and of itself.

The awesome part of cryoem at atomic scale is that we can now get MORE hypotheses generated faster with tougher proteins than with X-ray crystallography and other methods.

I won't lie, I was thrilled the first time I saw the holes in aromatic rings in electron density from my experiment. But until the very end of last century, X-ray crystallography was still difficult enough to be a full-time occupation and a single crystal structure was a PhD thesis. In my experience that kind of narrow focus becomes self-propagating whether or not the method generates beautiful figures.

Oh yeah, X-ray crystallography used to be crazy hard, and still is for many many things. I can imagine those holes must have been so satisfying :)

> I've come to see SB as a visualization tool that helps form hypotheses, but not an actual method by which most protein structural/function problems can be directly solved.

This seems to understate the challenge and power of generating good hypotheses.

But structural biology can be very inefficient.

Let's say you want to know which residues in a protein are important for a particular function, so you can tweak them to improve function.

You could

- spend months or years solving a structure of the complex then look at the structure and look at the direct contacts - then mutate to various other resides to see if they affect function ( finding out that even with a structure of the complex your chosen list is far from 100% correct )

- look at sequence alignments and then make a library of millions of variants and screen in a couple of months.

The first feels better as the glorious scientist is doing 'rational design', however it can be much more hit and miss and slower and more expensive.

As I understand it, one of the things about Cryo-EM is it's potentially much faster/easier than previous methods - but even so, looking at the structure doesn't get you to the answer direct - it's still very much hit and miss.

Brute-forcing the problem with a variant library and high-throughput screening could easily end up being far more inefficient. At least with crystallography it's easy to set up the experiment, whether or not it actually works.

You can even buy robots that do grid searches for crystallography (IE, if you're desperate, can't find conditions to make a crystal, and you've exhausted your postdocs, the robot will be able to try more possibilities in shorter time). Unfortunately they don't work very well and end up being a real pain to clean.

“Easy” is rarely a word I’ve heard associated with crystallography. Hence cryo-em! Easier and faster sample prep.

In the very best case, it is indeed easy! I've seen the entire process take less than a month for a completely novel structure - those were extremely lucky cases of course, but my point is that it costs very little to set up some screens and see what happens. (Small globular proteins of course, not ribosomes or ion channels.)

Nice! I always assumed unless there was a known crystal, you’d be in for pain. “Good and globular” is the way to go I guess.

Exactly.

There are obviously famous exceptions - like the structure of DNA revealing key functional aspects.

But for the most part it's like claiming photography is the key to understanding how animals really behave.

Single pictures in isolation are typically not very helpful in understand the biology.

Even the promise of having a structure and therefore being able to calculate physio-chemical properties has been much harder and slower than the hype.

Wouldn't a single image of an animal greatly improve understanding of animal behavior if previously we were unable to see the animal at all?

Yes, hopelessly extending this analogy, see the change in the way artists represented galloping horses before and after photography.

https://www.amusingplanet.com/2019/06/the-galloping-horse-pr...

And yet that was a snapshot of a dynamic process. Seeing an image of the horse didn’t actually reveal how it galloped, we had plenty of those and got it wrong. Seeing many images of the dynamics did, that’s a subtle but important difference. Movies are way more helpful than photographs. A crystal (or cryoem) structure is a snapshot of a horse standing still. We rarely care about that, though maybe we can guess that it uses its hind legs intensely because it’s very muscular. We do care about how it gallops, and don’t have a clue how with a single picture.

Put another way, we have “pictures” of dinosaurs, but have limited understanding of how they socialized, interacted, moved. We have ideas and guesses, sure, most based on descendants and modern context. Hypotheses, but few answers. Now if we had a movie of a dinosaur...

If we have its scat and see footprints we can learn a lot: where it lives, what it eats, what's related to. The fluorescence methods+knockout/expression can get you that sort of thing.

We need to know the elemental sequence of every human protein at some point, so we might as well start now.

Deepmind has made some interesting progress on modeling molecular structure from first principles using advanced ML architectures. Hopefully years from now these types of tools will let us understand the in vivo structural dynamics of these amazing molecular machines that give rise to life.

A few things here- first, we know the sequence of most proteins already. This is about structure, which is very different. It is entirely unclear that we "need" to know every structure to solve most important problems in biology (either basic research or applied).

As for Deepmind... they made an incremental improvement over the existing field (and everybody will be doing the same thing in 2 years). This ignores and underweights the decades of contributions from a wide range of people to bring us to this point in protein structure prediction and design.

I'm just an interested layman but it seems like this technology might be useful in developing therapies that either suppress or amplify immune response (e.g. diabetes, vaccines, cancer). For example, it may be possible to optimize the targeting of PD-1 inhibitors by comparing the PD-L1 ligands of healthy and diseased tissues in a cancer patient (edit: found something [2]). There was also a story [1] about how more advanced imaging was used to help identify antibodies that would be effective against SARS-CoV-2.

[1]

https://www.diagnosticimaging.com/view/role-emerges-imaging-...

https://www.nih.gov/news-events/nih-research-matters/novel-c...

[2]

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5050107/

It's absolutely useful for drug development and pharma companies appear to be getting heavily invested in the technique. The problem is that they've been trying to use structural biology this way for close to four decades now, and there are still many obstacles to this that have nothing to do with the structural biology methods themselves. Biology is messy, human biology particularly so. There are success stories of course but it's never been a magical solution.

Ah ok. That type of problem has to be orders of magnitude more difficult, especially when this still doesn't give us a great view into how everything interacts in a living creature.

Doesn't your disclaimer kind of also explain a bias against SB, as NMR is mostly used for protein dynamics elucidation (vs. the more rigid structures of Cryo-EM)? Or do you actually mean all fields of SB?

I mean all fields of SB. NMR is one of the worst- my PhD is in NMR structural biology. NMR gives you some dynamics information, but it's not great, hard to deconvolve and it's limited to very small molecules. of all the techniques, cryo is probably the best right now in terms of ease of sample prep and utility of data, but my point is much more that we don't even really need to know the structure of many things to make useful progress- fluorescence microscopy is far easier than any structural biology technique and provides far more useful signal.

Interesting!

> the most utility are fluorescent imaging with light microscopy, often at low magnification. These are more amenable to scale-up and machine learning

Any particular problems there, that you think can be tackled well by machine learning? I just recently started my first ML project in a fluorescence microscopy lab (biochem undergrad with previous software engineering experience), and while I see a lot of things that could benefit from ML for incremental improvements (which often need to be really tailored for each project), as is the case with most disciplines, I haven't really seen anything where I thought that there was potential for something groundbreaking.

there's good stuff here https://www.nature.com/collections/cfcdjceech if you have access to Nature.

Reply on news.ycombinator.com