There’s been a lot of AI doom and gloom of late. The newest generation of Large Language Models – OpenAI’s ChatGPT, Google’s Bard, and many others – have advanced so quickly that many knowledgeable experts are worried that they could eventually get out of control and do serious harm.

So much so that experts, including the heads of OpenAI and Google Deepmind warned this week that artificial intelligence could lead to the extinction of humanity.

But last week came a big media splash that pushed the conversation in the opposite direction. “New superbug-killing antibiotic discovered using AI”, said the BBC. “Scientists use AI to discover new antibiotic to treat deadly superbug”, said The Guardian. And so on, across dozens of news outlets with flashy, exciting headlines.

It was hard not to link these up with all the discussions about ChatGPT and the rest: it seemed as if the new advances in computing have already ushered in an age of AI drug discovery, where we can use super-advanced technology to guide our medical research in amazing new directions to save lives.

And it’ll all be so easy! The news stories about the new study noted that it took just an-hour-and-a-half for the AI to discover this powerful new drug.

But let’s take a deep breath. The new study did indeed use AI, and it did indeed discover a new antibiotic. But the general excitement about AI has, in my view, led to some over-hyping of this new discovery. Let’s try to put the whole thing in the proper context.

First of all, you can see very clearly that this isn’t a 2023-specific advance by looking at some older stories. In 2020, the press office at MIT announced that “artificial intelligence yields new antibiotic”; in 2021 Vox told us that “artificial intelligence can now design new antibiotics in a matter of days”. Obviously these happened before the current ChatGPT models, and before the whole world was talking about AI.

Indeed, in recent months “AI” has come to be almost synonymous with Large Language Models like ChatGPT. But it’s a much bigger concept than that. The term “AI” is used by different researchers to mean very different things, from quite basic statistical models all the way up to complex powerhouses like ChatGPT where even the designers can’t say they truly understand what the algorithms are doing.

“AI” is a very general term, but the vast majority of the current discussion is about a subset of AI called “machine learning”. Here, you take a set of data and feed it into a machine to “train” it to recognise patterns. Once it knows the patterns, it can predict what will likely happen in future.

This isn’t new: it’s been at the root of some major cultural trends for decades. Take Netflix as an example. Its machine-learning algorithm has learnt the sorts of films that people watch after watching, say, The Irishman. So if you happen to watch that film, it’ll give you suggestions of what to watch next that are probably pretty close to the kinds of things you like (other gangster films; other films by Martin Scorsese; and so on). If you think the “Watch Next” suggestions are impressive – or the same for other AI suggestions, like Spotify’s algorithmic playlists – that’s down to machine learning working well.

Scientists have also been using machine learning algorithms in studies large and small for a very long time (though with varying degrees of success). For example, in neuroscience research scientists have used data on the features, shapes, and sizes of parts of people’s brains to try and predict their likelihood of getting dementia in future. In genetics, too, complex algorithms are used to improve the prediction of various diseases and traits just from features of our DNA.

The research in the newest antibiotic paper – and indeed the algorithm used in ChatGPT – is an even more complex subset of machine learning called “deep learning”. In deep learning, it’s not that you tell the algorithm which features of a dataset you think are important and it learns them. In deep learning, the algorithm has a great deal more complexity, and can learn without so much human intervention, deciding for itself what the most important features of a dataset are.

That’s why it can be applied to very detailed datasets like pictures, videos, and texts (the latter of which is what ChatGPT is learning from: it’s been fed vast amounts of human-produced text, and uses all that learning as the basis of its uncanny ability to produce human-like language). Advances in both software and hardware – deep learning requires an awful lot of computing power – mean that this kind of model, which has been around as a concept for decades, can be applied in more and more useful contexts.

That’s where the new study comes in. The data in this case weren’t texts or pictures, but molecules that were known to slow the growth of one particularly nasty bacterium, Acinetobacter baumanii. Even though we knew about these antibiotic molecules already, we urgently need to find other ones. That’s because A. baumanii has the disturbing ability to develop resistance, meaning that we might in future run out of drugs that work to kill it.

The deep learning model learned the shapes and structures of these antibiotic molecules – features that, crucially, humans didn’t or couldn’t necessarily pick out themselves. It was then applied to nearly 7,000 different, newer molecules – and it picked out one with the potential to kill A. baumanii that we didn’t know about before.

There’s reason for caution – this hasn’t been tested in a clinical application yet. In the lab, the new molecule, called “abaucin”, seemed to reduce the growth of the bacteria in wounds in mice – so there’s every reason to be optimistic. But these were very small-scale, artificial tests; we’ll need a good deal more testing (and importantly, testing in humans) before we know whether abaucin is the antibiotic we’ve been waiting for.

So the “AI” used in this new study is deep learning, which has been around for a long time; it has a common underlying statistical model with systems like ChatGPT, but – just to be clear – would’ve come about in a world where ChatGPT was never invented.

Nevertheless, the scientists ensured that their research was plugged directly into the current AI zeitgeist: whereas in previous years they might’ve used the terms “deep learning” or “machine learning” (as they do many times in the study itself), for their press release they hit on the genius idea of using the term “AI” (a term not used at all in the study). That, I suspect, explains a big chunk of why this study got quite so much attention last week.

That doesn’t by any means imply that this isn’t an exciting discovery, or that we shouldn’t be excited more generally about the ability of these more complex statistical models to discover life-saving new drugs. We should be: deep learning was totally unavailable to previous generations of scientists and it’s very likely that it’ll lead to huge numbers of discoveries in future.

But there’s a final important caveat. As I noted above, for this study they had to feed the model information we already had on which molecules killed the A. baumanii bacteria. And that’s the fundamental issue with this kind of “AI” model: the model is only as good as the data we have. If we weren’t lucky enough to have good data on the previously-known molecules, there would be nothing to learn from: even the smartest AI model couldn’t help us.

Recently we’ve seen an AI drug discovery company laying off staff because one of its AI-developed drugs failed in clinical trials. The most likely reason for that was that the AI wasn’t trained on enough data to be fully accurate. Also this month, we’ve seen chemists lamenting the lack of good data they can use to train AI models to discover new substances.

So although it’s easy to get excited about AI’s scientific potential, it’s not always easy to feed the models with what they want most of all, and what makes them truly work for us: data. For ChatGPT, the data is everywhere: humans produce endless amounts of language that the model can learn from. For AIs that help with scientific research, though, much of the necessary data is either locked away on the computers of scientists who haven’t shared it with the community at large, or it simply doesn’t exist yet.

When it comes to discovering new drugs, or making all sorts of other scientific advances, it might be the case that data – not the AI models themselves – is what really holds us back.

By admin