Perhaps I'm missing the point, but I'm just not that excited about this work, mutating a GFP in silico and as a result getting a GFP (which was a constraint) with an entirely different aa sequence.... www.science.org/doi/10.1126/... - ThreadSky

joachimgoedhart.bsky.social • 40 days ago

Perhaps I'm missing the point, but I'm just not that excited about this work, mutating a GFP in silico and as a result getting a GFP (which was a constraint) with an entirely different aa sequence....
https://www.science.org/doi/10.1126/science.ads0018

Comments

joachimgoedhart.bsky.social•40 days ago

Maybe I'm just disappointed that the model did not (yet?) answer the unreasonable demand for a (much) smaller GFP...

mads100tist.bsky.social•40 days ago

Because the model is not trained to do that, regardless of what people try to sell

schwessinger.bsky.social•40 days ago

LLMs also don't 'reason' and they did not simulate evolution.

mads100tist.bsky.social•40 days ago

Correct

docdellaire.bsky.social•40 days ago

Iterate - regurggitate and enshittify. It is all about the quality and quantity of training data. LLMs are only as good as the data space and only “innovate” by hallucinating, often with ridiculous results. They have a place as a smart Encyclopedia though… kind of like a certain fictional guide

johndoench.bsky.social•40 days ago

prasad.bsky.social•40 days ago

"You don't understand...I coulda been a contender. I coulda been somebody, instead of a bum, which is what I am, let's face it."

https://youtu.be/7yz8kJnTRug?si=Lmvjn-QY7PTPEp5c&t=257

brooklynkid53bskys.bsky.social•39 days ago

I thought it was impressive cause they wound up with a protein so far away in seq space
but I'm really not an expert in this field

jdhl18.bsky.social•35 days ago

Seems very interesting and potentially important to me. Maybe I'm easily pleased 😀
What do we think of '58% sequence identity'?
If I copied somebody else's work to that extent I'd be in trouble for plagiarism 🤷‍♂️

joachimgoedhart.bsky.social•35 days ago

This is also an interesting observation:
https://bsky.app/profile/ddelalamo.bsky.social/post/3lfypeyjiw22z

cen0ten.bsky.social•40 days ago

To me it is mainly proof that they can design some proteins, but the design in itself does not have much of a use. And they make the effort to characterize the designs in vitro. Which not all design model paper does.

barakrotblat.bsky.social•40 days ago

perhaps we are expecting too much from one paper? sure, if they evolved a totally new enzyme in silico and tested it in vitro that would be grand. imo training llm to simulate evolution and show that it can change seq and keep function is interesting 🤷‍♀️

laurencepearl.bsky.social•40 days ago

It’s cute - sometimes that’s enough to make satisfying science

begrudginglyyeast.bsky.social•39 days ago

Does anybody have a system that can, for example, take an extant protein and remove its tendency to multimerize?

joachimgoedhart.bsky.social•39 days ago

Cool suggestion - for GFP-like proteins, or just any protein?

begrudginglyyeast.bsky.social•39 days ago

I mean I would love it for any protein but GFP would be a nice first step

labliston.bsky.social•39 days ago

Do you mean remove aggregation-prone seed sequences?

begrudginglyyeast.bsky.social•39 days ago

I'll just state what I am doing.

I am dealing with expressing a membrane protein. The protein tends to oligomerize into a hexagonal lattice in the membrane. I'd love for the protein to stop sticking to itself and making flat membrane plates but be otherwise functionally the same.

labliston.bsky.social•39 days ago

If the binding site is known, then wouldn’t it be as simple as creating a conflict in oligomerisation? A cation-cation repulsion should do it. If that site is occupied by oligos it shouldn’t have other functions. As long as you don’t do something to change overall structure should be straightforward

begrudginglyyeast.bsky.social•39 days ago

Yeah, I just would love something that could automatically make sure I am not changing overall structure.

labliston.bsky.social•39 days ago

Design a dozen mutations, test them in alpha fold, and then empirically test

begrudginglyyeast.bsky.social•39 days ago

Seems like just the sort of thing that you could build with one of these models

labliston.bsky.social•39 days ago

Might be overkill! Removing function is generally an easy problem to solve

senger.bsky.social•35 days ago

Hum so, the GFP could have existed with a different AA sequence

gemistosplethon.bsky.social•39 days ago

I agree plus the title is over the top.

vivekdna.bsky.social•38 days ago

This is a $142,000,000 paper. Hype is at its core.

timisstuck.bsky.social•40 days ago

I'm with you, I'm not seeing the big deal here. It seems like generating a new GFP was a demonstration of protein evolution in silico, rather than actually improving GFP. If that is the strongest/most exciting case they have for designing new proteins, it's not very exciting.

aydoganlab.bsky.social•40 days ago

To play the devil's advocate, isn't it quite exciting that these colleagues are building a large language model to potentially infer protein function from sequences? It seems a big deal to me. Most authors seem to be from a company that offers this as service. You can try it for free, it looks like!

timisstuck.bsky.social•35 days ago

Great point about the accessibility of this as a free service. As with any paper in Science, the bar is very high for amazingness. I agree it is a big deal, but is it a big enough deal for Science... I am just not *so* excited personally.

Comments

Posting Rules

Reply