In the age of AI, powerful new tools are consistently becoming available to researchers, opening up new pathways and entirely new questions. But with these new tools comes a new twist on an age-old issue: If all you can think of is a hammer, everything looks like a nail.
David Sholl, Rice’s executive vice president for research and professor of chemical and biomolecular engineering, recently published an opinion piece in ACS Central Science, where he and co-author Andrew Medford, associate professor of chemical and biomolecular engineering at Georgia Institute of Technology, share their perspective on machine learning tools expected to bring radical changes to the field of computational chemistry.
“We can see the paradigm shift coming in the tools available to analyze and understand the energies defining the atoms within chemical structures,” said Sholl. “This gives our field the unique opportunity to deeply think about how these tools can best be utilized and developed prior to incorporating them in our research.”
Chemists, Sholl and Medford posit, should learn from other research communities who have gone through similar shifts with tools like AlphaFold. Specifically, the field of researchers should start asking what questions these tools can answer and, equally importantly, which questions they can’t.
These machine learning tools will, Sholl explained, be able to analyze energy required for interactions at an atomic level, then use those energies to predict how the atoms will interact with each other in various arrangements. While computational chemists can do this using current tools, they are limited to analyzing a small number of atoms. Machine learning, however, is expected to expand that number to 10,000 atoms or more, bringing a significant change to how computational chemists approach these problems.
For example, a chemist might use the current models to study five potential chemical structures of a potential drug. From that small subset, it is relatively easy for an experienced computational chemist to somewhat intuitively select the best candidate. With machine learning, however, chemists will have thousands more options to consider, far more than they could individually review.
“This exponential increase in information drastically changes how we approach these problems, at every level,” said Sholl. “The questions we can ask both change and exponentially grow, and with that, there becomes a danger of mistaking a useful tool for a universal one.”
Would it, for instance, always be better to consider all 10,000 options, or should chemists use the same machine learning tools to select just a small number for individual review? In both cases, machine learning tools would need to be leveraged to select either a small number or a final option, so how can the researcher know the tools are selecting the desired candidates?
These are not trivial questions. Machine learning tools democratize complex mathematical algorithms by removing the need to understand the mathematical functions guiding the work. In this case, they allow researchers to ask questions about atomic interactions without necessarily needing to understand the complex underlying physics. Soon, non-computational chemists should be able to use these tools to ask computational questions in their own research. Tool developers, therefore, must ensure their tools are not only well-designed, but that they clearly communicate their limitations, in terms accessible to colleagues from a variety of different disciplines.
“By taking a thoughtful pause before these tools become accessible, we can ensure that they become incorporated where and how they are most useful,” Sholl said. “We can learn from the mistakes and accomplishments of other fields that have already benefited from powerful machine learning tools.”