Scientists are using AI to dream up revolutionary new proteins

Hi, I'd love to chat -- could you please email me jordan[at]astro[dot]codes?

I'm gonna check that book, I bet I will learn a lot.

I sometimes wonder how we are going to solve the scaling problems around silicon and energy for this kind of AI applications. Think about this: a life-saving cancer drug may need to be custom-designed for a patient by using a few Giga-joules of energy to feed a data-center.

Would love to contribute, please contact me at variofab(at)icloud(dot)com

Good luck- this has been a long-term interest of mine, but one that has not really led to any forward progress. Lithography continues to be a far more approachable technology.

When you express recombinant proteins in bacteria, you grow them and then let them produce your protein of interest. Then you destroy the bacteria and separate the soluble parts (where you hope your protein stays) from the insoluble ones (protein aggregates, cell membrane components, etc.), usually by ultracentrifugation. The soluble portion (or supernatant) is a clear liquid, while the insoluble fraction (or precipitate) is a whiteish, gunky solid.

So, to answer your question, soluble proteins would look clear. However, to actually check for the presence of your protein of interest in each fraction you load them in a polyacrylamide gel and pass current through it to separate all proteins in the mixture, using a method called electrophoresis[1].

[1]: https://en.wikipedia.org/wiki/Polyacrylamide_gel_electrophor.... In the first picture of the article each band represents one (or several, if you’re unlucky!) protein of a determined molecular weight or size.

Nothing. If they fold well and are water soluble, the solution remains transparent because the protein remains solvated. If they don't fold well, they expose hydrophobic regions that then bind to the hydrophobic regions on neighboring proteins, causing a chain reaction that results in a solid chunk of protein that isn't dissolved.

Not all proteins are water soluble in nature, however. Plenty of membrane proteins and other proteins exist that don't fold well when expressed in a micro-organism.

Indeed, a simple spell for animating an arbitrary number of broomsticks to fetch an amount of water given by argument. [0] Oh, and it was given to us by a demon whose output we don't really understand. What could go wrong?

Thankfully, if I'm understanding the article properly, the broomsticks didn't properly animate, though both the demon and human researchers figured they would. I'm all for demon-assisted magical research, but I very much doubt that the next action taken will be "abandon this avenue of research until we understand the demon's process better". We should probably understand this kind of thing if we're going to be messing with it.

[0] https://www.youtube.com/watch?v=VErKCq1IGIU

Prions don't build new prions form smaller parts, like virus do. Prions just change the shape of a complete protein that is already build. But each prion can change the shape of only one type of protein, not any protein that is floating around.

I think the risk is almost almost almost zero.

Protein design protocols usually favor globular, soluble proteins, where the sort of interactions generating amyloids and prions are absent.

One goal of protein design is to create bespoke enzymes that catalyze reactions of interest, such as the degradation of target compounds. Keep in mind that the vast majority of de novo designed proteins to date lack a defined function and have been used to showcase structure rather than functionality.

Any bio-processes that you can think of. For example, spiders fabricate ultra-strong fibers using a few cubic millimeters. Compare that volume to the smallest Kevlar factory we can set up.

A lot of applications will be about designing or adapting biological processes; so they wouldn't be enabled by just protein design but would need way more "tooling". It's worth noting that the computational viability of much of that tooling is boosted by AI as well. That I think is the main value and the "new thing": we got a new, powerful class of "hacks" that we can use to shortcut and scale in-silico bio-research.

Laundry detergent. Genentech, a biotech company, engineered a version of a protease (a protein that cuts other proteins) to work at high temps, then formed a spinoff with Corning to license that tech for billions to laundry detergent companies. https://en.wikipedia.org/wiki/Genencor

doesn't the article say that the proteins they dreamed up didn't work

What do you mean? It certainly can fold new proteins that were not part of its training set, otherwise it is useless. Maybe you mean that it cannot fold proteins in new ways that were never seen in its training set though ? But new proteins might fold in known ways, right?

Snark aside, this is not AlphaFold per se, but sort of a reverse AlphaFold, the same way protein design can be defined as the inverse protein folding problem.

Tangent: Has AlphaFold and other "ML"/"AI" programs made progress on describing how proteins fold? Or is it solely for final (post-folding) coordinates/dihedral angles? Has there been any progress on understanding folding? Results from folding-at-home? Is it treated as too tough of a problem?

> This can't be true, many bioscience people have commented assuring us that because Deep Fold doesn't literally spit out new proteins, its nothing new, and saying otherwise is misleading

From the article...

> But when they instructed microorganisms to make their creations in the labs, none of the 150 designs worked. “They didn’t fold at all: they were just gunk at the bottom of the test tube,” says Baker.

For the record, David Baker (and other groups) have been successfully designing new proteins just from their own understanding of protein folding, for years. So "Deep Fold" still has yet to catch up to our own limited human understanding.