An AI that learns about chemical reactions and designs a procedure to make them | Hacker News Reader

An AI that learns about chemical reactions and designs a procedure to make them

new.nsf.gov

141 pointsgeox46 comments

This is identical to what I've been building and using my own gpt (via ChatGPT) with API integration to automate full genome base-calling, assembly, annotation and prediction pipelines. Using this approach, me and my team can use it to debug pipeline failures, interpret the results, draw inferences and create context-dependent experiments. There's no leap in the tech beyond GPT facilitating the interaction with APIs, but it is a completely new way of working. This is what I can currently do:

User: Show me my genomes

ChatGPT: Here's a list of your genomes ... [API call to GET /genomes]

User: Annotate genome #nnnn

ChatGPT: Kicking off annotation [API call to POST /annotate]

User: Show me what genes are related to growth regulation for the genome

ChatGPT: The genome appears to have several genes related to growth regulation. Here's a list [API call to GET /genome/X/annotations] ...

User: Design an experiment to knock-out this gene using CRISPR in Aspergillus niger, include the guideRNA specific to this gene.

ChatGPT: Here you go...

Really interesting, what would be your real-world use cases?

This is the real-world use-case! We're using to accelerate the discovery and elucidation of novel or unknown gene function in previously unstudied fungal strains.

I should have meant, have you already found customers? I am also working on a project on genomics, which is very orthogonal to yours, and finding customers in this field the difficult part

We're pursuing grant-funded research at the moment. We have some hunches on some specific fungal genetic functions but they need to be better characterized before we think about commercialization.

Why do you need ChatGPT for the API calls?

As an example, asking questions like "show me which genes from my genome are related to growth regulation" requires an API call to give ChatGPT a list of annotated genes from the genome we're studying so it can then filter them into growth regulation related genes. Leveraging ChatGPT's extensive knowledge of genes lets us find new insights into our newly assembled genomes much faster.

Is this doing anything more than using ChatGPT as a frontend? I thought based on your comment ChatGPT would run all of that, not just call an API that does the actual work.

The parent article isn't doing anything more than using ChatGPT as a frontend either. The API does all the _work_ (pipeline running, robot control, etc) while ChatGPT uses it's vast knowledge to interpret, analyze and iterate on future commands sent to the API.

Oh God this is going to be really bad, there are so so so many misannotated genes, even well known ones that have more or less "wrong names"[0], that a lot of lazy scientists are going to either be led astray or totally miss targets as a result from an uncritically science-believing LLM

[0] do you really really think ferritin is best described as an "iron carrier"?

For anyone interested in the applications of ML to chemistry, PostEra has been doing some pretty cool stuff in medicinal chemistry for the last several years: https://postera.ai/

I have a friend that works on Azure Quantum Elements and they said the Copilot integration is pretty impressive. I imagine that's a very specific niche though so likely not interesting if you're not into material science or computational chemistry.

Copied text from Wikipedia somehow seems different from deriving the nobel prize winning process from a dataset that didn't have the answer in it.

And downvotes. The article presents this as a magic box capable of novel work, and some distance down says it found the answers on Wikipedia. Those are different things, whether you're excited about LLMs or not.

Definitely different things but I suppose “machine that finds related prior research for me effortlessly” is still quite useful.

Yes, missing prior work at best can impact your review, and at worst can mean you waste a significant amount of research time

> In less than four minutes, Coscientist had designed an accurate procedure for producing the required reactions using chemicals provided by the team. When it sought to carry out its procedure in the physical world with robots, it made a mistake in the code it wrote to control a device that heats and shakes liquid samples. Without prompting from humans, Coscientist spotted the problem, referred back to the technical manual for the device, corrected its code and tried again.

What does it matter that it looked up the reactions on Wikipedia? They're building a "coscientist", not an AI-based Nobel prize winning researcher.

> What does it matter that it looked up the reactions on Wikipedia? They're building a "coscientist", not an AI-based Nobel prize winning researcher.

Well, quite. This is attempting to automate the boring, time-consuming and dangerous part of chemical research.

We should be applauding.

You got it! That’s part of what we did, but so much more… Coscientist democratizes access to fields that one is not a deep specialist in, allowing for people to merge different ideas with ease. Biologists can do cutting-edge chemistry, chemists can try the latest in materials science, and so on.

"a magic box capable of novel work"

or a box whose content can be cleanly swapped out with future iterations which already demonstrate the potential for novel discovery:

https://thenextweb.com/news/deepminds-ai-finds-solution-to-d...

"Researchers claim it is the first time an LLM has made a novel scientific discovery"

> "Researchers claim it is the first time an LLM has made a novel scientific discovery"

LLM didn't make discovery on its own in that work, it played role in one step of unknown importance, other steps were lots of manual coding and lots of CPUs to brute force solution.

I think the point was that the application includes the source attribute of the data.

so crediting prior work is the novel thing done by the AI?

You should read the actual paper before dismissing it. Fig 6 and the surrounding discussion shows that their system is able to iteratively optimize a reaction by adding steps and changing reaction conditions. Based on how poorly the model does on its first try, it seems unlikely that it had already just memorized the best reaction. So, they show that the model can try things, observe results, and try again.

Also their model converts the "wikipedia answer" into steps that they run on a robot, which hasn't been done before for this specific reaction.

Novel work can often be accomplished by taking two existing methods and implementing parts of both, it doesn't have to be a eureka moment that stems from nothing. In this case, seeing what went wrong in one experiment, seeing what went right in another, combining the best of both, and re-implementing...is novel work.

This framework of retrieving information , analyzing it with program code as an output , programming machines to do something , retrieve information about the output of the machines , analyzing it to see if it matches the expectations and if not repeat and rinse until you get the perfect outcome ....

seems useful in general for all kind of AI agents. Is this framework open source yet?

It seems... unlikely, that a LLM built from scraped webpages, would be able to predict the results of novel chemical reactions. I don't really understand what information in the model's source would lead to new insights into chemistry.

Serious question: Given that AI's may (apparently) not own patents, what would be the implication if an AI (presumably) independently comes up with Nobel Laurette level chemistry/[science] - What happens to such body of work WRT the greater humanity? Who owns such (whatever the product is?)

If a company has built upon whatever AI GOLLUM service, and then either the GOLLUM comes up with some epic thing, is it all for the corp that owns the GOLLUM? Does the AI provider get some slice?

Is this already a solved problem and I dont know what to call it?

Yeah science! Yeah Mr. GPT!

Is there something along these lines for physics experiments?

Looking forward to having my own personal Jarvis

“Consider a future device … in which an individual stores all his books, records, and communications, and which is mechanized so that it may be consulted with exceeding speed and flexibility. It is an enlarged intimate supplement to his memory.”

By Vannevar Bush

The Memex

A great idea until Big Brother subpoenas you and decides that pleading the fifth does not apply for electronic devices.

It doesn’t apply to paper documents either. Though it does apply to your passcode :)

"Say, that's a nice 'democratic' governmental organization you have there....would be a real shame if something was to happen to it."

I would assume a device that records data is already fair game for evidence by legal means. If you want to get technocratic about it, you'd need to start arguing that this kind of tool is an extension of your mind/consciousness of which you are currently entitled privacy to (probably more due to lack of tech than strong ethics). But there's no tech the government does not want to regulate, so your version is probably right.

I wonder if that memory thing Dumbledore had was open to Ministry subpeonas...

Smartphones are basically already "The Memex", didnt really expect this would start te current thread aha

Read/discussed "As We May Think" in a grad school HCI class - highly recommend it to anyone interested in the history of computing.

It is available behind a paywall on The Atlantic website, but you can see the original magazine layout here (which is even cooler IMO) - https://archive.org/details/as-we-may-think

Building 13 @ MIT is named after him.

Cool to think this tech will be there soon :)

I thought we already had that. We humans are never happy with the current stuff

Wait, it works cheaper than Post Docs? I am somewhat serious because I am a musician. In Dallas my view is Bench Science is woefully underpaid and over performs at an unsustainable dynamic.

> I am somewhat serious because I am a musician.

Saying the post docs might have it worse than musicians.

musicians/artists are the archetypal poster children for intense training with underpaid work. That a musician is calling out science work as the same, is notable as the musician would have the experience to know.

Typical post-docs may be the worst systemic ratio of pay to skills and experience anywhere. Probably related: it's otherwise often a pretty amazing gig. Not universal obviously.

Reply on news.ycombinator.com