Artificial Intelligence Meets Real Problems

Artificial Intelligence is a hot topic in many fields of research, and chemistry is no exception. A quick search reveals over a hundred thousand papers published in this field, and the number is only growing year after year.

AI is being used more and more by chemists for various tasks. Initially, AI was applied to accelerate drug discovery and predict functional properties, as well as reduce the costs associated with computation and experimentation. But can AI truly bring a revolution in chemistry, or has its promise been exaggerated? The topic was addressed by a stellar panel and an Agora Talk at #LINO22.

Brute Power

AI has already made a tremendous impact on chemistry, says moderator Gunnar Schröder, a Professor at Universität Düsseldorf whose work focuses on computational and experimental structural biology. “Specifically in chemistry,” says Schröder, “there’s been a lot of progress in the past years.”

But much of this progress doesn’t necessarily come from better algorithms or smarter methods, but rather from an increase in processing power, says Michael Levitt.

“I think that all of AI is much more continuous than you realize. AI was very popular in the 1960s, the big difference has been the massive increase in computer power. So basically, what has been really important for AI, in general, is the speed of modern computers.”

Levitt was awarded the Nobel Prize in Chemistry in 2013, and much of his work included using rudimentary computers to simulate virtual molecules. Someone else who was working on something similar at the time was Arieh Warshel. Warshel was trying to simulate small molecules, while Levitt wanted to look at larger molecules. Ultimately, this work would bring them both a share of the 2013 Nobel Prize.

Warshel also seems slightly disenchanted with AI in chemistry – not because you can’t do useful things, but because it doesn’t always help you understand things, and that’s what he’s most interested in.

“I never believed in AI and only recently I kind of was forced to use some of it. The questions that interest me forever are how enzymes work,” the Laureate said at an agora talk where he presented some of his past and current work.

The problem raised by Warshel has also been addressed multiple times by researchers in various fields where AI is making an impact. Since AI is a bit of a black box, even for its makers, can we truly trust AI if we don’t understand the decisions it makes? However, Warshel also concedes that sometimes, understanding isn’t everything. If you have a cancerous mutation and you use AI to find a compound that is effective against it, that’s great news, although it may not truly advance our understanding of the process.

“To me, AI is something like brute force and it remains to be seen just how much we can learn from it,” Warshel adds.

Open Science

Warshel also opens up another concerning point: that AI is opening up a wider gap between chemists and the public.

“We had major difficulty to explain it to the public, partially because the public does not learn chemical physics, and AI is one of the reasons why people learn it even less.”

But panelist Neo Neng Kai Nigel National of the University of Singapore sees an opportunity of democratising science and making it more open. While some years ago, AI algorithms were restricted to big companies and research groups, now there’s a myriad of algorithms that can be used by anyone, which makes the entire field more democratic. Asked by the audience what was the most disruptive part of AI in chemistry, the researcher said:

“To me, the answer would be the availability of open-source chemical informatics.” A couple of minutes later, he also added an important aspect: a lot of these machine learning models are rooted in the open-source community, which harbors transparency and democratization of science. It’s the big companies who typically produce the most un-transparent science, he adds.

Paulina Paiz of the University of California sees a similar advantage. Although the hype can work against some computational biologists, AI in chemistry can make bring make science a more open endeavor. “Traditionally, science is gated, but with these efforts, you get a lot of citizen science,” Paiz says, which not only draws more people to science but also contributes to the system of checks and balances that is so important to science, and suggests that some AI could even be used to explain what other AIs are doing, opening up the “black box”.

Folding Toward the Future

Of course, the panel couldn’t avoid AlphaFold in this discussion. AlphaFold, an AI iteration from DeepMind, performs predictions of protein structure, a problem computational researchers have struggled with for years. The 3D structure of proteins is crucial to understanding their function; similarly, if you want a protein that can perform a certain task, you need to have a good idea about what shape it should have.

In 2021, AlphaFold proved its worth by predicting 3D models of protein structures with remarkable accuracy, something which its creators say has the potential to accelerate research in every field of biology. But the panel has mixed opinions on AlphaFold, and without minimizing the accomplishment, they suggest that once again, it may be the extra computation power that is making a difference.

“Since AlphaFold there have been other groups that say they have performed marginally better but a lot of that can be attributed to more data,” says Paulina Paiz. Neo Neng Kai Nigel highlights another problem: DeepMind hasn’t exactly been very open about how its algorithm works. “AlphaFold itself wasn’t really transparent in terms of how the model was creating the design, so I think more transparency by AlphaFold will definitely be better.”

But ultimately, AlphaFold is a great example of what AI can bring to chemistry: it promises a lot (maybe even too much), it’s a bit mysterious, but it offers a lot of valuable information to the scientific community. Perhaps herein lies the key to AI in chemistry – and to a greater extent, in science in general. It works best when used in conjunction with human intuition and knowledge, and when the data is provided with competence and transparency. AI isn’t here to replace human researchers – it’s here to complement them.

“A lot of people talk about AI replacing humans, but that seems unlikely because AI plus humans just works better,” concludes Levitt.