From Atoms to Algorithms, AI is Having a Big Impact in Chemistry

As the Lindau sun was scorching outside, the Inselhalle was having its own stellar event. Six researchers, including some of the world’s most influential voices in chemistry, gathered to answer a deceptively simple question: how is artificial intelligence changing chemistry, and science as a whole?

AI tools like AlphaFold are already having a big impact on the scientific community. However, the panelists were quick to point out that such models are only part of the story.

The panel, which was moderated by Derek Muller of the Veritasium Youtube Channel, turned into a spirited discussion. From the promise of faster drug development to warnings about over-reliance on algorithms, the discussion showed just how transformative AI has been for the chemistry community (and science in general).

AI Is Not Just About Data

Right from the start, when panelists were prompted to give their own definition on what AI is, you could see the different visions emerging. John Jumper, who was awarded the Nobel Prize in 2024 for this work on AlphaFold, offered a clear, practical framing: “I tend to take an expansive definition and say AI is machine learning, I really don’t understand the people who draw all these distinctions, at least in chemistry. It’s whenever basically data plus programme comes together to make a function.”

Michael Levitt, who was awarded the Nobel Prize in 2013 for his pioneering work on computation biology, quipped that even a calculator could qualify as AI, depending on your definition. Meanwhile, Animashree Anandkumar, Professor of Computing at California Institute of Technology, emphasized that it’s not about the data, but about adaptability: “intelligence is the ability to learn and adapt to surroundings,” she noted. This is particularly true for AI used in scientific discovery. Her approach is to incorporate physical laws and domain knowledge into the AI, essentially installing “guardrails of physics” to make good use of the data.

“Especially for scientific discovery, a purely data driven method will fail, because if the whole point of discovery is something new that has never been seen before, so you cannot expect to see it in training data… you’re also embedding the knowledge of physics within the algorithm itself.” A good example of this is weather forecasting. Traditionally, this has been done in a bottom-up way, where you start from things like Stokes equations and thermal conductivity and so on, and you try to forecast what the weather will be like in the future.

“Instead, we were able to train AI models based on the historical weather data, as well as some of the knowledge of physics together, and be able to speed up forecasting tens of thousands of times.”

The panel all agreed that higher computation power supported the revolution of AI in science, but Jumper emphasized that this is only half of the reason for the progress. The other half comes from better data and better algorithms used to train and direct AI. “We aren’t just running the algorithms that we ran before,” the laureate mentioned.

AI Is a Tool That Should Be Used Wisely

Another point of agreement was that AI should be treated as a tool and as an accelerator. Joachim Frank, who was awarded the Nobel Prize in 2017 for his work in cryo-electron microscopy (cryo-EM), says he’s “absolutely enthusiastic” on using AI to guide and accelerate experimental analysis. Frank describes how AI predictions with high-confidence regions can be aligned with experimental cryo-EM density maps to locate and test structural elements, a painstaking “detective work” that AI can greatly accelerate.

However, this doesn’t mean that experiments or human supervision is no longer needed, Frank emphasizes. Even for applications for which AI is excellently suited, like drug discovery, you can’t overlook the complexity of real-world chemistry and biology and you need human input to make sense of the findings. Danielle Belgrave, Vice President of AI and Machine Learning at GSK, and also a Lindau alumnus, further emphasized this idea. She explained that even with AI, drug development isn’t a simple pipeline. “When we use AI in the drug development process, what we see in reality is a lot of heterogeneity in patient response,” she said. “So it’s very much a measurement science, trying to iterate what are the things we need to measure in order to encapsulate patient’s heterogeneity.”

Although large language models (LLMs) aren’t at the forefront of scientific disocovery, they were also discussed. Levitt said he is an avid user and asked LLMSs some 50,000 questions, noting that it’s “sometimes really stupid and sometimes incredibly clever.” Used wisely, LLMs could be used in the complementary, often bureaucratic aspects of science, such as preparing drug applications or patent submissions.

However, with all AI uses, things can go easily wrong. Frank mentioned his concerns about using AI to both analyze and generate data, which can go in an uncontrolled direction and hijack the scientific process. Jumper also emphasized the importance of using AI tools responsibly in science. In particular, he underlined that it’s important to know when AI does something wrong.

“An AI tool where the user misunderstands the reliability of the tool can be really challenging and may not accelerate science… It’s better for your wrong answers to look really, really wrong than for all your answers to be structurally plausible, but some of them to be very wrong.”

From Finding Molecules to Romeo and Juliet

If there was still any doubt that AI is just getting started in chemistry, you just needed to make your way to the Stadttheater (Lindau Theater). There wasn’t any dramatic play happening, but nine young researchers presented their work mixing AI and chemistry.

To understand why AI is reshaping chemistry, consider one of the field’s most essential tasks: figuring out what an unknown molecule actually is. This process, known as structure elucidation, is at the heart of many scientific discoveries. One of the most powerful tools for doing this is mass spectrometry. It works by breaking molecules into fragments and measuring their mass-to-charge ratios. These measurements produce a kind of fingerprint, a pattern of peaks that can be compared to reference spectra in a database. If there’s a match, the identity of the molecule can often be confirmed.

But There’s a Catch: The Chemical Universe Is Vast

“For the chemical space itself, and for small molecules only, it ranges up to 10 to the power of 16,” explains Magdalena Lederbauer, one of the young researchers from ETH Zurich. “That’s equivalent to taking all grains of sand on planet Earth and multiplying it by the number of stars in the entire observable universe.” Databases typically contain spectra for around 100,000 known compounds.

If your molecule isn’t in the database, you’re stuck running extra experiments and burning through time and resources, says Lederbauer. This is where AI can make a dramatic difference by learning how to predict what the mass spectrum should look like, even for molecules that have never been measured before.

“We developed a physics-constrained hybrid neural network,” she said. The model breaks down a molecule into fragments and predicts the strength of each signal. This process is not a black box. It works in steps, mimicking how the molecule would behave in the real machine, and enabling an understanding of how the algorithm is reaching its conclusions. Each predicted fragment of the model can be linked to a peak and vice versa, which can help identify and verify structural hypotheses. Incorporating physical laws also makes the entire process much faster, just as Anandkumar explained.

Meanwhile, Lucas Franco, from the Federal University of Rio de Janeiro, is tackling a different kind of unknown: the vast space of drug-like molecules that haven’t been tested yet. His team uses unsupervised learning, a type of AI that finds patterns without being told what to look for. They mapped out clusters of molecules to understand where new drugs might be hiding. Using this approach, Franco’s group ran virtual screening for kinase inhibitors, a class of drugs explored in treating some neurodegenerative diseases, finding several promising candidates.

Some researchers are using AI not just to identify molecules, but to design entirely new ones — especially materials that can capture and convert carbon dioxide. Mohmmad Faizan, from the National Institute of Technology Warangal in India, is tackling this problem using a clever metaphor drawn from Shakespeare.

“Imagine a situation where Romeo and Juliet can see each other, talk to each other, but cannot meet,” he said. “Now imagine the amount of frustration they have.” 

This, he explained, is how Frustrated Lewis Pairs (FLPs) work. These are pairs of chemical groups, one acidic and one basic, that “want” to react with each other but are kept just far enough apart to prevent it. In their “frustrated” state, they become unusually reactive to other molecules like carbon dioxide. Essentially, they can activate and break carbon dioxide, which is normally a very stable molecule.

This process could play a role in our climate efforts; one of the key approaches (in addition to reducing our emissions) is finding ways to take out CO₂ from the air; using it for useful products would be an important bonus, and FLPs could offer significant help in this regard.

The lectures went on, covering remarkably diverse topics. Wenchao Duan, from CSIRO in Australia, is using AI to detect antibiotic pollution in water, while Giulia Frigerio is studying molecular dynamics that could be used to deliver more targeted cancer treatment.

Although the sessions were diverse, they all pointed to a common thread: AI is no longer a future tool, it’s a present force. Yet AI alone is not enough. It needs to be grounded in physical laws, guided by human insight, and most of all, used responsibly. When applied thoughtfully, it can unlock parts of the chemical universe we’ve never seen before.