Using AI to design proteins is now easy. Making enzymes remains hard.
- Thread starter JournalBot
- Start date
So... whatever is wrong with using fungus? Too slow?
https://www.sciencealert.com/scient...can-break-down-tough-plastic-in-just-140-days
https://www.sciencealert.com/scient...can-break-down-tough-plastic-in-just-140-days
Upvote
39
(40
/
-1)
Upvote
81
(82
/
-1)
This is the type of thing I want to see generative models used for, not chatbots or social media garbage or "art".
Something like this has to take far less energy and input to train because of the specific problem area, why is it so hard to get money flowing to this research?
The thesis lately is that the government should be funding little to no research and that private companies are motivated to do all of this, this seems like an area that even a few hundred million dollars has the potential to significantly impact potential developments and drive actual solutions.
Spending tens of billions of dollars on "AGI" or hypergeneralized models. Seems like it would be better focused and spent on hyperspecific problem spaces.
Maybe I'm wrong, I dunno
Something like this has to take far less energy and input to train because of the specific problem area, why is it so hard to get money flowing to this research?
The thesis lately is that the government should be funding little to no research and that private companies are motivated to do all of this, this seems like an area that even a few hundred million dollars has the potential to significantly impact potential developments and drive actual solutions.
Spending tens of billions of dollars on "AGI" or hypergeneralized models. Seems like it would be better focused and spent on hyperspecific problem spaces.
Maybe I'm wrong, I dunno
Upvote
100
(106
/
-6)
That's the problem of private funding, it generally occurs when there is a profit to be made. And much of this research, that needs to be done and utilized, won't happen without government funding (too much risk, no clear profits for private funding to occur).
Charity can raise only so much, and (probably) inconsistently.
Upvote
53
(53
/
0)
-I- meant all the plastic. Frankly I consider it a win.
I'm with John here - the next obvious step is to encode it into some bacteria and do some forced evolution to really crank up the variations on the novel enzyme. Seems like a great way to get the small optimizations once the gross design is done.
Upvote
-2
(10
/
-12)
ColdWetDog
Ars Legatus Legionis
Nothing wrong with that approach at all. In fact, if the result is repeatable and useful in something other than a lab setting, it may be this AI project to the punch. But they are complementary approaches.
The fungi approach is to throw the entire kitchen at the problem (the organism) and see what sticks. This approach is much more sophisticated (and complicated and slow). It does offer the advantage of being able to do things that natural processes simply cannot or will not.
But, I, for one, welcome our plastic eating fungal overlords. They have to be able to clean things up better than humans are managing it.
Upvote
46
(46
/
0)
Upvote
55
(57
/
-2)
The article doesn't mention who is doing this work. The lead author of the paper is David Baker - joint winner of the 2024 Nobel prize for chemistry, and creator of Rosetta@home which allows you to contribute compute cycles to this stuff: http://boinc.bakerlab.org/rosetta
He won that Nobel prize alongside Deepmind's founder Demis Hassabis (also the co-creator of Theme Park) and John Jumper also from Deepmind.
He won that Nobel prize alongside Deepmind's founder Demis Hassabis (also the co-creator of Theme Park) and John Jumper also from Deepmind.
Last edited:
Upvote
91
(91
/
0)
To be fair, research IS benefiting from everyone chasing chatbots.
In my industry, we're deploying LLMs and building our own transformers, but we're doing it using GPUs that are powerful and cheap, with software written for LLMs, with drivers and middleware that are all built from chasing LLM.
So even though our LLM doesn't even remotely resemble a "chatbot" and our transformers handle data that nothing like language, and we're using them in production on much weaker hardware, we're only able to do that based off the papers, research, advances that were made by other companies pouring billions into the next "chatgpt."
If the transformer didn't exist, a lot of "regular" AI advances that we're also making today would not have been possible.
Upvote
63
(63
/
0)
Thing is we could literally use mushrooms to eat plastic (see my previous post), basically free and it won't cost the trillions invested in Ai.
The only downside us that the plastic eating mushrooms are slower the upside is price and the fact thar under the right weather conditions you can set those plastic eating mushrooms anywhere.
Upvote
0
(8
/
-8)
Curious whether it would work just as well to design multiple enzymes to facilitate that 4-step process. One "do it all" enzyme seems out of reach, but perhaps 2 enzymes can work together similarly? Maybe the amount of work to do that is the same or even more than trying to design the single enzyme...
Upvote
2
(2
/
0)
A few years ago this article would have used ML instead of AI, and in my mind ML would still be a better fit. I'm not quite sure how forcing the AI to learn to go back a workable state is "generative AI" of the sort I normally think of as "Generative AI". Sure, this generates results, but so did any original ML algorithm. Random trees generate outputs, so is it a generative AI? The use of AI rather than ML, and especially generative AI, feels like more buzzword use rather than accurate description (but that's without reading the original paper).
Upvote
17
(21
/
-4)
Ah why can't we just have AI digesting plastic instead of running government or social interface systems... Ah but technology is agnostic, it is the PEBKAC that always messes things up. Sigh. 
Upvote
3
(4
/
-1)
Chemists have been using molecular simulation and screening tools to design enzymes for decades now. The big breakthrough here seems to be that the system can work backwards from the structure to the sequence. That sounds useful, but chemistry is full of surprises as this report shows. I'll call it a great new tool, but it's far from magic. Evolution has had billions of years to come up with something better than rubisco for splitting carbon off of carbon dioxide, and rubisco is slow, imprecise and obviously suboptimal. It's also the basis for photosynthesis on this planet. Chemistry is hard, and not just for scientists.
I'm not surprised funghi can do some impressive biochemistry. Many of them have a specialized genetic like code for producing toxins and likely other chemicals. It uses DNA and RNA, but the translation uses a number of other components along with some amino acids. This lets funghi build some really weird and potent stuff. It's also beyond the scope of existing software. The real breakthrough would involve coming up with an appropriate engine like this as needed. So far, no signs of funghi coming up with something better than rubisco, though.
I'm not surprised funghi can do some impressive biochemistry. Many of them have a specialized genetic like code for producing toxins and likely other chemicals. It uses DNA and RNA, but the translation uses a number of other components along with some amino acids. This lets funghi build some really weird and potent stuff. It's also beyond the scope of existing software. The real breakthrough would involve coming up with an appropriate engine like this as needed. So far, no signs of funghi coming up with something better than rubisco, though.
Upvote
19
(19
/
0)
To do that you'd need to make the precursor toxic to the bacteria, to "incentivize" it to clear it as quickly as possible. That's probably possible to do, but then the precursor wouldn't be polyester any more and the resulting enzyme might not be useful.
Upvote
2
(2
/
0)
Yesterday I watched a "Veritasium" video, "The Most Useful Thing AI Has Done", about using AI to design enzymes and proteins ... and I was both fascinated and scared by what might come of all this.
Upvote
8
(8
/
0)
It's an impressive result, considering how many properties your de novo enzyme must have:
...and more, I'm sure, but that's already asking a lot.
- it must be translatable in bacteria
- it must be stable enough to perform its function
- it must not be toxic to bacteria
- it must catalyze the desired reaction at a useful rate
...and more, I'm sure, but that's already asking a lot.
Upvote
8
(8
/
0)
ColdWetDog
Ars Legatus Legionis
Industrial scale mushrooms are a problem. Not insurmountable, but difficult. Industrial enzymes are also a problem but different ones.
But your logical fallacy is to assume that just because you can solve a problem without AI that you don't need AI to solve other problems or, more succinctly in this environment, to make money.
Upvote
11
(12
/
-1)
There are hundreds, more likely many thousands, of enzymes which do this with a single protein. The basic chemistry is very simple, but the details are complex. The biggest problem for unnatural substrates is having the enzymes actually recognizing the molecules they are supposed to catalyze a reaction of. This is particularly true for plastics, which are long polymers, usually of very low water solubility, and from the perspective of enzymes largely featureless. A multi-enzyme complex would make the chemistry steps more complex, and you would have to get the various component enzymes to recognize one another. There are many such enzyme complexes in nature, but they are usually doing something much more sophisticated than simple ester hydrolysis.
Upvote
9
(9
/
0)
I felt like I was reading an article summarizing this Veritasium video that released 4 days ago.
View: https://youtu.be/P_fHJIYENdI?si=8JKAwXXTeCO_QBax
View: https://youtu.be/P_fHJIYENdI?si=8JKAwXXTeCO_QBax
Upvote
8
(8
/
0)
AI is cool i guess
Ars Centurion
it's so fascinating watching the convergence of multiple AI systems to solve problems that took evolution billions of years. the really interesting pattern here isn't the enzyme, it's what happens when different forms of intelligence learn to collaborate
Upvote
-4
(1
/
-5)
ColdWetDog
Ars Legatus Legionis
Ok, sure. But where are all the enzymes that this 'old' process kicked out? Genuinely curious since Google, at least, has been trying to figure out how to make enzymes on demand for some time. And even if the 'old' system worked, perhaps the LLMs would do better? No idea, not in my wheelhouse. Not even the same boat. But different approaches to the same problem are often useful.
Upvote
3
(3
/
0)
sigmasirrus
Ars Scholae Palatinae
According to that article:
“More than 400 microorganisms have so far been found to degrade plastic naturally, with fungi attracting a fair bit of attention for their versatility and ability to degrade all sorts of synthetic substrates with a powerful concoction of enzymes.”
Yet in the Ars article:
“Unfortunately, there isn't an enzyme for many reactions we would sorely like to catalyze—things like digesting plastics.”
So which is it?
Upvote
9
(10
/
-1)
"polyester got its name due to how many instances of the chemical bond show up in it"
Technically, polyester got its name because the single units (two different monomers) making up the polymer do so by reacting to form ester bonds. It's one of two ways of naming polymers - the other is when a single monomer is joined with itself, and is named for the monomer (such as polyethylene or polyvinylchloride).
Technically, polyester got its name because the single units (two different monomers) making up the polymer do so by reacting to form ester bonds. It's one of two ways of naming polymers - the other is when a single monomer is joined with itself, and is named for the monomer (such as polyethylene or polyvinylchloride).
Upvote
14
(14
/
0)
sigmasirrus
Ars Scholae Palatinae
I just think that machine learning is a better, less loaded term.
Upvote
10
(10
/
0)
A bit of both. Yes, there are organisms that can use their existing enzymes to digest plastics, but that may require using the whole organism in order to retain the activity, which might not be optimal for industrial scale processing. Live bacteria as used in bioreactors, for example, require care and feeding beyond simply chucking in finely ground pop bottles. A single enzyme that doesn't require much else in order to function (maybe a buffer and a controlled temperature) simplifies things considerably, as long as you can produce it in the required quantities.
Upvote
10
(10
/
0)
ColdWetDog
Ars Legatus Legionis
The fungi that can process plastics likely use a multi enzyme pathway, probably attached to a cell membrane to work. And, being fungi, the process is slow (they don't care, they don't have to). Doing a quick search about molecular mechanisms of fungal plastic degradation doesn't yield anything I can either read or access (anybody got some references?). My WAG is that fungal biochemistry isn't well studied* and the particular metabolic pathways even less so.
So it isn't and either / or here. While there always seems to be promise of myco remediation for plastics, fungal biochemistry really hasn't yielded any commercially viable systems. So doing it in vitro (literally in glass, outside of the organism) is certainly a way to go and yes, TFA is correct, we don't have a specific enzyme that does that.
* Actually doing a bit more research, there is a bunch of stuff being done at the molecular level, mostly at the 'neat metabolite' stage (think penicillin, psilocybin, etc.). However, none of it seems easily available and since I retired I don't have access to any institutional libraries and even if I did I'd probably fall asleep reading the stuff so there...
Last edited:
Upvote
9
(9
/
0)
I am mystified by this article. So, it's about digesting plastic using some enzyme(s) as (a) catalyst(s). Fine. 1. But what is the end product? Digested to produce what? One would hope it does not include carbon dioxide or methane. Plastics in landfills produce neither (unless, of course, the landfill catches fire). 2. Where and/or what is/are the proposed application(s)? (I suppose if this article were in a chemist-specialist newsletter these queries wouldn't be required, but it is not.)
Upvote
2
(3
/
-1)
Something like:
We finally decomposed all the hydrocarbons in our landfills!..
And released gigatonnes of CO2 in the process..?
Upvote
10
(10
/
0)
ColdWetDog
Ars Legatus Legionis
The advantage of doing this step by step outside an organism is you can (probably) control it. You don't have to make CO2, maybe CH4 (which is useful) or octanol or butanol or whatever. We're pretty far away from this but if you can control the enzymatic pathway with that degree of precision, the world is your oyster. Or hydrocarbon.
Further, this is really just a test system for designing an enzyme to do some particular, specific, controlled function. Harness this and you have the Diamond Age. Or The Stand.
Upvote
10
(10
/
0)
One of the best YT channels and I look forward to every new episode.
This is a great use of AI. Science data that would have taken years or decades to process can now be done in days or weeks.
Scary and amazing that it can learn at a geometric rate.
Upvote
0
(0
/
0)
The answer is profits.
You get profits selling the AI shit to idiots. Researchers are notoriously under-funded and can't provide the necessary income to make grinding out actually useful things profitable most of the time.
Lots of proofs of concept. But unless it's monetized, it's useless to industry and business - those who implement what researchers discover.
My question here is what comes from the digestion of these plastics? Is it a useful substance, or merely a more biodegradable one than plastic itself? How much money is there in creating biodegradable material from plastic?
I'm thinking little to none, in direct profits.
The problem with humans these days is that we expect all things to lead to "profit" without much, if any, regard to the costs to our health, welfare and safety. creating cheap plastic items generates lots of profit. Finding a way to make that plastic biodegrade solves the problem of disposing of plastic, but it is not a profitable venture unless adequately paid to do it. Dumping plastic into the ground is cheaper on the quarterly statements than paying for a process that biodegrades it.
That's the problem. We have lots of solutions, but no one wants to pay for them. That will happen with this, too, because the people who run companies tend toward psychopathy far more than the rest of the population.
That this process is also complicated complicates the issue.
Upvote
9
(9
/
0)
Machine learning is Artificial Intelligence and Artificial Intelligence is Machine Learning.
Just that you get more interest and money when you talk but Artificial Intelligence.
This work is not a generative model as generative model is a machine learning model designed to create new data that is similar to its training data. Hence, chatGPT is a generative model as it provides sentences like it was trained on.
Rather this is a discriminative model like random forests because it is predicting if the enzyme will function or not.
Upvote
4
(4
/
0)
What doesn't at this point?
Fukushima's been leaking into the pacific for how long now and we still don't have Godzilla?!?
But anyways! Check out this movie:
https://en.m.wikipedia.org/wiki/Crimes_of_the_Future_(2022_film)
One of those possible unintended consequences plays an interesting part in the story.
Changing our bodies to be able to eat plastic. There's also a plot about how random new organs of unknown use are starting to pop up in people. And it gets weirder from there.
Also, it's just a really good and very, very weird movie.
Last edited by a moderator:
Upvote
0
(0
/
0)
What the current US government forgets is that advances like this are built on 10,000s of papers published on protein chemistry and biology over the last 70 years by scientists doing basic research at universities.
Upvote
10
(10
/
0)