Jakob Uszkoreit:

Basically, it is a way of describing a flavour of medicines or a type of medicines in the broadest sense that is very much akin to software. You generate those medicines by supplying, initially, some kind of declarative or other definition of their behaviour – their intended behaviour.

The hope is that we will eventually get to the point where we can have fairly complex definitions or declarations of intended behaviour that are then successfully compiled into descriptions of molecules that exhibit those behaviours inside cells. This is quite different from programmable medicine, where you start with a base that is customised. Instead, this starts from scratch.

An instance of this that we all know is the mRNA vaccines, whether for COVID-19 or otherwise. The mRNA mostly does one thing – it tells cells to translate a given protein. Now, with the mRNA vaccines, it's evident that that's not the only thing they do.

The idea behind biological software is that you can design molecules based on a description of their behaviour that have, by and large, only the desired effects.

There were three events that happened in rapid succession, within less than three months, in late 2020 – the height of the pandemic. First, my daughter was born. It was an even more special experience to have my daughter around that time; I didn't know if I could actually be there for the birth due to lockdown restrictions. There was lots of reading that I did because it was very much unclear at that stage how [COVID-19] could affect pregnancies and newborns.

Shortly after, the CASP14 results were released. AlphaFold2 wiped the table with every other method for predicting protein structure. CASP14 was effectively when AlphaFold2 made it clear that, yes, this problem is maybe not solved per se, but at least that very particular version of protein structure prediction, albeit contrived and maybe not the most practical, now is effectively done.

This was stunning because AlphaFold2 used deep learning methods I had co-invented. It was moving to see this proof that these methods were ready for prime time in molecular biology, but sobering because I hadn't been directly involved in that work.

A few weeks after that, the first mRNA COVID-19 vaccine efficacy results came out. It was incredible – RNA was effectively saving the world. Yet, there was a clear gap: we couldn't apply deep learning to RNA due to a lack of data.

Those three events left me really questioning what I was doing at Google. It was in so many ways, my dream job. After some months of pondering that and talking to lots of experts in molecular biology, RNA biology, in particular, it became clear that this was almost a moral obligation.

In terms of biological software, we're at the earliest days. We're at a point in time where we can see and maybe give examples for why it should be possible, but where our ability to truly execute on it and really implement it is still nascent at best.

Overall, this is a very typical phenomenon where we want to see revolutions. That's to a large extent because that's how we tell stories, but the truth is that almost all the “revolutions” that I've been a witness to are actually incremental progress that eventually reaches some kind of tipping point.

When it comes to understanding life, I don't think we're there yet. We're seeing that with, say, protein structure prediction, there is the potential to do similar things, but it turns out that protein structure prediction, per se, isn't as general of a problem, if you wish, in something like drug development or drug design that just "solving it" will really break it down.

We're making interesting steps, and we can see light at the end of the tunnel, but it's very unclear how far we have to go. One of the reasons there is that we simply don't have the data yet. The data just doesn't exist that could even conceivably inform a large neural network about all the different mechanisms or pathways of life, even if you reduce it to some of its more simple forms, it's not the case so far that we can truly predict their behaviour given an intervention using deep learning. At least now we have a tool that did something we couldn't do for things like language, and so we can definitely try.

The biggest hurdle is the lack of data. We don't just need observations of life, like individual cells, but also systemic phenomena like immunology.

What we're trying to do is make incremental progress toward designing useful or beneficial interventions with the data we can collect today that may or may not also be useful, in the longer term, as we progress towards this more grand vision of something like biological software. It won't be built overnight – it's a gradual process.

It's interesting, the reality is that it started out being incredibly open and enthusiastic. A good chunk of that was fuelled by the post-COVID mRNA vaccine excitement.

As that subsided, it's been replaced with an incredibly conservative, very methodical, but really also comparatively slow way of making progress. What we're now seeing is that that hype cycle – the en vogue, hype of the day – is now shifting, or has shifted, to a variety of different things.

I guess that's just the way this sector works. and probably for pretty good reasons. Now, could it work in certain areas? Could it work faster? Should it work faster for the benefit of humanity? Absolutely. No question.

I've heard this advice from several people, including one of the AI “godfathers” Geoff Hinton: Never assume that tomorrow will be like today. Sounds trivial, but it's actually really deep. We take so many things not for granted, but as a given. It usually turns out that whatever ways or approaches or directions we have and we pursue might be fine, but they're almost never ideal.

Another quote I like is from Edison: "There's a better way. Find it." That's the essence of innovation – challenging the status quo and always looking for better solutions.

I try to avoid the phrase AI because we don't know what intelligence is. There are many practical definitions, and unless you specify it clearly, I feel it's rather misleading by default.

The biggest misconception today, is that these things “just work”. If you look at what it took to really make it work for language, to the extent that it works today, if you go to ChatGPT or similar tools, yes, sure, there's this core of a large-scale, pre-trained model based on some architecture, and it was trained on a giant data centre full of GPUs. But, the reality is, that's not enough. You need a tonne of human-annotated data, product iterations, tweaks, and little tricks to make this work.

There is absolutely zero reason to believe that this shouldn't be the case in the life sciences or in healthcare. In fact, there are many reasons to believe that you need even more of these kinds of tricks.

We have no instance that is able to understand how even just a single cell works. I think we will never, as individuals, because it's too complex, our cognitive capacity just is not a good fit for this phenomenon.

We have to accept that we will have to implement and devise many such tweaks and hacks in order to make it work even remotely as well because it's a harder problem, and we don't have an instance that we can directly learn from, like we do, for language and vision.

It's exciting and frustrating in equal measure.