As a science communicator with a skeptical brand, I often have to walk a fine line. New scientific and technological developments can be amazing, but they are often surrounded by hype. I want to encourage enthusiasm for science, and I want to share the amazement and joy I experience following the latest discoveries. But it is very important to separate hype from reality, to temper our enthusiasm with realism, and to not get ahead of the science. It can be a very narrowly calibrated sweet-spot, one I have to consciously pay attention to.

I have found it particularly challenging to hit this sweet spot with artificial intelligence (AI), especially recently. The past few years in particular have seen some amazing advances in AI applications, and in 2022, applications like DALLE-2, Midjourney, and ChatGPT showcased the power and disruptive potential of AI to the public. It’s hard to oversell how powerful these narrow AI applications can be, but at the same time it is easy to overhype them in other ways. I think it often comes down to this – people generally (even experts, historically) underestimate the potential of narrow AI (AI applications that have a specific function and are not conscious or have general intelligence). At the same time they tend to overestimate how soon we will see general AI, or they extrapolate linearly current AI advances into the future.

For example, with applications like Midjourney and ChatGPT, how far will applications like these go with extrapolations of current technology? They certainly will be better in 5 and 10 years, but will they transform into something truly creative? And does that even matter? Will these applications run into limits, or will they advance to be indistinguishable from human creators?

Meanwhile, programmers continue to show off the power and potential of current narrow AI technology, fueled by massive data sets and ever more powerful computers. Perhaps the most transformational applications are those running in the background of public consciousness, working to accelerate research and technological development. One area where AI is becoming particularly powerful is drug development.

Drug development is ripe for AI applications by its very nature. There are astronomical numbers of potential chemical compounds, which can potential interact in numerous ways with complex biological systems. That is a vast “space” in which drug development can function. Traditional methods of finding new drugs in this unbelievably vast conceptual space include building on naturally-occurring chemicals that have a known effects in animals or people (essentially, taking advantage of all the trial and error already done by evolution). Drug developers also have developed catalogues of chemical structures and their known effects, to help them predict what a new chemical compound might do. But then their predictions have to be tested in cells and animals before ultimately in people. For years already drug developers have been using computer simulations to help make the initial predictions of chemical structure and function, to optimize the chance of testing successfully in biological models.

And now drug developers can use AI to run simulations, crunch massive data sets, and make predictions about drug structure and function. This is perhaps an ideal application for AI – exploring a vast conceptual space to find needles in massive haystacks. This process can replace a lot of costly and time-consuming trial and error in the biological research phase. A recent study validates at least one application of AI to drug development – in the design of polymeric long-acting injectables (LAI).

The idea of LAI is to improve compliance, reduce the daily burden of drug-taking, improve efficacy and reduce side effects. Drugs have a half-life in the body, and this is often a challenge in drug development. If the half-life of a drug is too short, even if it is otherwise a perfect drug, that can render it useless. Drug developers have created methods for extending the half-life of drugs, including extended release technology, so that they only need to be taken once or twice a day. This has a significant effect on compliance. In one review:

Compliance rates averaged 76% during 3428 days observed: 87% of the once daily, 81% of the twice daily, 77% of the three times a day, and 39% of the four times a day dosages were taken as prescribed.

Four or more times per day is essentially unworkable. But even for drugs that need to be taken three times a day patients missed almost a quarter of their doses. For some pharmaceutical applications, steady blood levels is critical. So sometimes drugs need to be taken more frequently in order to level off the resulting blood level. Another way to look at this is – the less frequently you take a drug the higher the peak and lower the trough (lowest) blood levels. This can translate into more side effects during peak levels and lower effectiveness during trough levels.

LAIs are one solution to all of this. Imagine getting an injection once a month and then not having to worry about taking a pill one or more times per day, always having it with you when you travel, keeping the drugs secure and not losing or spilling them, and having optimal steady blood levels over time. But LAIs can be challenging to develop, specifically engineering a long acting steady release of active medication. That is where AI comes in. The recent study looked at published results of using AI to design polymeric LAIs. They found:

Here we show that machine learning algorithms can be used to predict experimental drug release from these advanced drug delivery systems. We also demonstrate that these trained models can be used to guide the design of new long acting injectables.

Basically, the AI-driven process works. This can lead to significant reduction in the cost of trial-and-error experimentation, and increase the speed of development. This kind of advance is unlikely to grab the public attention (like an art-creating program like Midjourney can) but it is more profound. It works in the background, and you have to know a little something about medicine and drug development to appreciate what a revolution it is.

What I see in my clinic is that patients are happy to receive the benefits of the latest medical technological development, drug or otherwise, but have little appreciation for the long scientific process that lead to their treatment. This is not a judgement – why would they? Unless you are in a scientific field or follow this kind of news closely, it’s mostly a black box. An expert could write a book about the 30 year saga of multiples lines of scientific research and technological develop that lead to the latest monoclonal antibody treatments for migraine, for example. In the near future such books would include the long history of AI technology and how they contributed to the process. The public mostly sees only the end result.

I do think it is good for the general public to have a greater appreciation for the process, and least conceptually if not in detail. It is necessary for the public’s continued support for the institutions of science and research. The latest treatments they enjoy do not come from nowhere. Also, perhaps it will make people generally less susceptible to the latest medical pseudoscience. If they have some appreciation for the process perhaps they will be less likely to believe stories of miracle breakthroughs by lone geniuses outside the system.

Meanwhile, technologies like AI will continue to rapidly transform our world, operating mostly in the background.