On a recent SGU live streaming discussion someone in the chat asked – aren’t frivolous AI applications just toys without any useful output? The question was meant to downplay recent advances in generative AI. I pointed out that the question is a bit circular – aren’t frivolous applications frivolous? But what about the non-frivolous applications?

Recent generative AI applications are a powerful tool. They leverage the power and scale of current data centers with the massive training data provided by the internet, using large language model AI tools that are able to find patterns and generate new (although highly derivative) content. Most people are likely familiar with this tech through applications like ChatGPT, which uses this AI process to generate natural-language responses to open ended “prompts”. The result is a really good chat bot, but also a useful interface for searching the web for information.

This same technology can generate output other than text. It can generate images, video, and music. The results are technically impressive (if far from perfect), but in my experience not genuinely creative. I think these are the fun applications the questioner was referring to.

But there are many serious applications of this technology in development as well. An app like ChatGPT can make an excellent expert system, searching through tons of data to produce useful information. This can have many practical applications, from generating lists of potential diagnoses for doctors to consider, to writing first-draft legal contracts. There are still kinks to be worked out, but the potential is clearly amazing.

Perhaps most amazing, however, is the potential for AI in general, including these new generative AI applications, to assist in scientific research. This is already happening. As someone who reads dozens of science press releases a week, it is clear that the number of research studies leveraging AI is growing rapidly. The goal is to use AI to essentially complete months of research in mere hours. A recent such study caught my attention as a particularly powerful example.

The researchers used generative AI (an application called SyntheMol) to design potential antibiotics. Again, AI aided drug development is not new, but this looks like a significant advance. The idea is to use a large language model AI to generate not text but chemical structures. This is feasible because we already have a large library of known drug-like chemicals, their structure, their chemistry, the chemical reactions that make them, and their biological activity. The AI was trained on 130,000 chemical building blocks. This is a type of chemical language, and the AI can be used to generate new iterations with predicted properties.

This is essentially what traditional drug design does, but AI just does it much faster. It is estimated, for example that there are 10^60 potential drug-like chemical structures that could exist. That is an impossibly large space to explore with conventional methods. The AI used in the current study explored a “chemical space” of 30 billion new compounds. That is still a small slice of all possible drug molecules, but this subset had parameters. They were looking for chemicals that could have potential antibacterial activity against Acinetobacter baumannii, a Gram-negative bacterial pathogen. This also has been done before – looking for antibiotics – but one problem was that many of the resulting chemicals were hard to synthesize. So this time they included another parameter – only make molecules that are easy to synthesize, and include the chemical reaction steps necessary to make them.

In just 9 hours SyntheMol generated  25,000 potential new drugs. The researchers then filtered this list looking for the most novel compounds, to avoid current resistance to existing antibiotics. They chose 70 of the most promises chemicals and handed them off, including the recipe of chemical reactions to synthesize them, to a Ukrainian chemical company. They were able to synthesize 58 of them. The researchers then tested them as antibiotics and found that six of them represented structurally unique molecules with antibacterial activity against A. baumannii.

These results would have been impossible in this time frame without the use of generative AI. I would call that a non-frivolous outcome.

Drug candidates resulting from this process still need to be tested clinically, and may fail for a variety of reasons. But chemists who develop drugs know the parameters that make a successful drug. It has to have good bioavailability, a reasonable half-life, and a relative lack of toxicity (among others). These are all features that can be somewhat predicted based upon known chemical structures. These can all become parameters that SyntheMol or a similar application can use when generating potential molecules.

The goal, of course, is to do as much of the selection and filtering as possible digitally, so that when you get to in-vitro testing, animals testing, and eventually human testing, the probability of a successful drug has already been maximized. The potential for saving money, time, and suffering is massive.

This is only one specific example of how this new generative AI technology can supercharge scientific research. This is a quiet revolution that is already happening. In spaces where this kind of technology can be effectively leveraged, the pace of scientific progress may increase by orders of magnitude. Fans of the Singularity might argue that this is the beginning – a time when the pace of scientific and technology progress becomes so rapid that society cannot keep up, and the horizon of future predictability narrows to insignificance. The Singularity refers more to a time when general AI takes over human civilization, technology and research. But even with these narrow generative AI tools we are starting to see the real potential here. It’s both exciting and frightening.