Some doctors often cheer AI for its potential to advance medicine. Though this seems true, one notes that very few new antibiotics have been developed in the last few decades. This has caused concern some resistant infections may be hard or impossible to treat.
AI has been offered as a way to review potential medications and identify strong candidates. Testing is still expensive and takes time. But I am unaware of a long pipeline of new antibiotics. Still, if it were available it would be strong proof of the benefits of AI in this domain.
So my question is, what similar developments would convince you that AI provides strong benefits commensurate with its unknown risks, in different domains?
It’s probably worth noting here that the “AI” systems that are developing new drugs (or at least, trying to) aren’t particularly related to ChatGPT and its ilk.
This is a good point. Most of the recent usage of the term ‘AI’ seems to relate to Large Language Models (LLMs): basically text-based generative systems trained on an enormous amount of ‘scraped’ material from the internet.
My impression is that these are all generically rather similar?
Does anyone know of a good technical summary that explains the similarities (or differences) between the current generation of these (ChatGpt, Claude, Gemini etc etc…)?
But it ain’t necessarily so: this is certainly the dominant model for ‘AI’ at the moment… but it is probably not a pathway to the Holy Grail of true ‘Strong general AI’?
I’m probably about as dubious about the wonders of AI in its current form as anybody. However, AI hasn’t been around in its current form, or anything like it, for the last few decades. More like the last couple of years, isn’t it, and changing all the time? So why would there be a decades long pipeline of new drugs from it?
Total WAG, but since this is IMHO, one of the low-risk applications of AI in medicine is to offer theories and suggestions that can be validated either by testing and/or cross-reference with other data. Thus ISTM that AI assistance in creating new medications can be valuable in areas like predictive analysis of molecular behaviours and how they may interact with human biology, all subject to exhaustive testing and validation once a potentially new drug has been developed.
This is analogous to an important distinction I’ve made before, the difference between asking a generative AI, on faith, to diagnose a set of medical symptoms, versus using it to suggest appropriate probing questions, like “given these symptoms, what should I ask my medical specialist to elicit the most complete possible information?”.
By the same token, AI can potentially be very useful in pharmacodynamics to answer the question, “what is likely the most productive path forward in developing a drug to treat this particular disease?”.
So the last new class of antibiotics was discovered in the 1980s. Some blame a lack of incentives for this gap, but newer fourth-generation antibiotics are very expensive so this is not a complete explanation.
In his book Super Agers, Topol talks about goweeeo neural networks looked at millions of candidates. For C. Difficile and M. Tuberculosis 107 million possibilities were reduced to eight (Stokes, Cell, Feb. 20, 2020). Impressive, but not much has been heard since then AFAIK. A new drug, lolamicin, has been said to work against bacteria without disturbing friendly microbiome bacteria in the gut. But after the initial hype, again AFAIK not much has been said since.
Sure, AI is new and this research problem is so old that progress in this area could be credited easily to AI. Of course vaccines may help with done things but that is a different kettle of fish.
Expensive isn’t enough. Antibiotics are used selectively and for short periods of time; the market incentive is much less than a new lifestyle medication that will be taken for the remainder of someone’s life.
Then there is the bit you already note: identifying targets is one thing; testing them is still a laborious many years real world process. And the drug has to be in a form that works pharmokinetically. Thus there is the exciting and cautionary example of halicin: discovered but not yet developed:
I don’t think my link is working so here’s another one to a more academically oriented article:
Recent advances in artificial intelligence (AI) and machine learning (ML) have introduced a paradigm shift in drug discovery. These technologies are capable of processing vast chemical and biological datasets, recognizing complex patterns, and predicting bioactivity with unprecedented speed and accuracy. AI-based models can screen millions of molecules in silico, reducing both time and costs while accelerating hit identification [8]. A landmark example of this approach is the discovery of halicin, a compound originally developed as a c-Jun N-terminal kinase (JNK) inhibitor for diabetes and later repurposed as a potent antibiotic through deep learning algorithms developed at MIT [9].
Halicin exhibits a broad-spectrum antibacterial activity distinct from classical antibiotics …
… While our data are promising, several considerations must be addressed before clinical application. First, the pharmacokinetics and pharmacodynamics (PK/PD) of halicin in humans remain poorly understood. Preclinical studies have shown favorable results in mouse infection models, but extrapolation to human physiology requires careful dose optimization and safety evaluation [29,30]. Secondly, halicin’s stability in biological fluids, its interaction with host microbiota, and its potential cytotoxicity must be thoroughly evaluated [31].
It was identified by the MIT deep learning AI over six years ago; the article above published last year. It doesn’t seem that the process of getting from identification to clinical use is being sped up?
There’s a semantic argument to be had, but I would say that this is false overall. Transformers and the way they implement neural networks are pretty broadly shared between different types of deep learning models. ChatGPT (and large language models in general) and Geometric deep learning / graph neural networks are basically two implementations of the same idea. They aren’t separate inventions, they’re just two different branches on the same tree. If you invented one, it’s a relatively small leap to invent the other.
Degree of relatedness does come down to semantics, but at the very least, you can’t just ask ChatGPT “Which of this set of molecules will form a shape similar to this other molecule”, and expect any sort of reasonable answer.
Well, I disagree. I’m guessing your point is something like “having neural networks that help discover new medicines don’t require that we also have chat bots and image generators” but they’re the same genie from the same bottle, just different wishes, and if it’s not that, I’m not sure what your comment was trying to address or add.
It would be hard to not invent planes after the advent of the internal combustion engine and a little bit like saying the planes that fly are not particularly related to the things that drive on the ground. I mean, sure, there’s a distinction there, but they basically both were created as a result of the ICE and both practically inevitable after the invention of the ICE. I would actually say that the geometric deep learning and large language models are more closely related than cars and piston airplanes too.
So if you could magically go back and make LLMs not exist, you take the medicine-designing AI with it.
The technology behind ground vehicles might inevitably have led to air vehicles, but that still doesn’t mean that my Toyota can fly. Similarly, it might have been the same technology that led to both the biomedical AIs and ChatGPT, but that doesn’t mean that ChatGPT can evaluate drugs.
I guess it would help if you cleared up your intent when you tried to specify that medication-designing AI was mostly unrelated to chatgpt. The OP didn’t talk about LLMs or chatgpt specifically and only hinted at the potential risks AI brings. I took your intent to basically say “the risks are mostly brought by things like chatgpt, and we don’t necessarily have to have those risks because it’s not deeply related to the AI systems that bring advantages like designing medicines”
Yeah, I can see why it’s a problematic analogy, especially since I have some unstated premises. Flight was not unique idea on the part of the Wright brothers. Several inventors had basically the right idea. The problem was that metalworking available at the time didn’t allow the production of an engine that had a good enough power to weight ratio to support heavier than air flight. Once it did, several different inventors basically invented flight separately. The ideas were there, they were just waiting on materials science to advance. Some deal with the light bulb – as soon as we could make vacuum sealed glass and work tungsten everyone was basically inventing it at the same time. But I realize this analogy is too complicated to serve really well as an analogy so I regret making it.
My point was this: you basically can’t have the sorts of AIs that are designing new materials and new medications without also having chatgpt. They’re the same technology with different applications, and why would people refrain from exploring all the useful implications of this technology? That’s what I was trying to say with how the advent of the internal combustion engine naturally lead to both the car and airplane.
I don’t think chatgpt and medicine-developing neural networks are meaningfully separable. You’re not going to get one without the other.
Well, mostly, nowadays when someone says “AI”, everyone automatically thinks of LLMs like ChatGPT. If people thought that ChatGPT was helping design medicines, they would be absolutely correct to worry about the outcomes of that process. So I was trying to get ahead of that misconception.
There are also other advanced computer systems that could have been called “AI” when they were developed, but which don’t use neural nets at all, and are written by humans who understand the code they’re writing. Those also have lots of applications (some of which LLMs can do but not as well), and it’s also useful to distinguish those.
I’m not actually sure that’s true. LLMs, at least initially as I understand it, were very focussed on building a neural net ‘weight’ model based on verbal input.
And their success has certainly shown that a lot of human knowledge: more that we might like to admit, even, is encoded in language.
But is that the whole picture? I don’t know, but I doubt it.
So I’m not an expert but I’m an enthusiastic amateur, I hope this answer makes sense and is correct. GNNs (graph neural networks) and transformers are trying to understand the relationships of information by finding the statistical similarities and dissimilarities and building up a model of them in a geometric form of latent space. Basically - they try to turn a massive amount of knowledge into a neural network of relationships. Transformers (what powers chat bots, most images generators and a lot of consumer-facing AI) are basically a subset of GNNs. In GNNs the neurons only care about their neighbors, but in transformers every node cares about every other node. But a transformer is basically a GNN with a sort of larger map of nodes. They’re both doing the same sort of passing massive amounts of data through a networking weighting process to be able to understand the relationships in that data. The data is different. LLMs look at potentially trillions of human documents. Image generators look at billions of pictures with associated word combinations. GNNs use whatever kind of data you want to examine but in the examples of simulating nanomaterials and medicine they build an understanding of molecules, physics, and chemistry.
But it’s basically the same concept underneath all of them being applied to different fields. There’s basically no way we would’ve invented one without inventing the other within a few years at most. And that’s the history of how it happened in the real world.
I’m not really sure, For example, calculus was invented before linear algebra: though, looking back from what we know today, linear algebra seems simpler in principle?
I mean – as a counterfactual that’s hard to answer. I don’t know enough about math to say how closely related or how obvious linear algebra and calculus are related to each other. But we can see in the real world that deep learning neural networks very quickly made people see that the technology and fundamental concepts applied to different domains quickly started exploring those domains. The genie is certainly out of the bottle now - you basically can’t say “let’s focus on this one application from neural network deep learning and try to suppress the rest” – even if one government were on board with that, every other country would just bypass you. And for all the doom and gloom and hatred of AI out there, no one is interested in discussing that kind of ban seriously now, and people are vastly underrating the positive impacts of chatbots, image generators, and other consumer facing deep learning neural networks.