A pioneering research paper from our friends at DeepMind, published in Nature Machine Intelligence, illustrates the state-of-the-art in molecular machine learning. These exciting new advancements were spearheaded by researchers from Carnegie Mellon University, including Assistant Professor Gomes and Ph.D. They employ a novel strategy based on stereoelectronic effects. Ultimately their aim is to improve the accuracy of these models while working with limited data – addressing a formidable barrier for chemists.
For more than ten years, Gomes has been delving into the connection between structure and reactivity at the molecular level. This line of inquiry has brought him to understanding how—and more importantly, why—stereoelectronic effects can dictate molecular interactions and, by extension, chemical fate. Having the opportunity to collaborate with Boiko has deeply enriched our understanding of these principles. We’re just now starting to put them to use with innovative machine learning approaches.
Development of Innovative Models
Boiko and Gomes have developed an intriguing new model which, given a standard molecular graph, quickly produces an enriched representation. These faster calculations replace traditional methods that typically require hours or even days to analyze molecular interactions. Their model produces results in seconds. This new efficiency expedites operational research processes. Even more importantly, it enables chemists to use the model in cases where full quantum chemistry calculations would be too cumbersome.
Their new model encodes this stereoelectronic information through what they refer to as stereoelectronics-infused molecular graphs (SIMGs). Such developments allow for a deeper, richer interpretation of how features in electronic structures control behavior and reactivity on a molecular level.
“This model can be applied when regular quantum chemistry calculations are not possible, like for entire peptides and proteins,” – Daniil A. Boiko
Accessibility Through Web Applications
To further enhance the practical application of their findings, Boiko and Gomes developed a web application designed to analyze stereoelectronic interactions of molecules quickly. This tool brings their methodologies within reach of researchers from nearly any discipline, encouraging the widespread adoption of their innovative techniques.
The web application is simple and intuitive to use. It allows novice and experienced chemists alike to explore intricate stereoelectronic interpretations without fear. Boiko and Gomes want to make the process as easy as possible. They’re interested in moving from cutting-edge machine learning models to integrating with laboratory workflows.
Addressing Challenges in Data Availability
One of the biggest hurdles to applying chemistry in this manner is the sheer lack of available data sets to train machine learning models on. Boiko emphasizes this issue, stating, “In chemistry, we have very small data sets.” Their research does an excellent job to address this limitation. It serves as an example that even with smaller datasets, we can extract important insights without losing overall model effectiveness and explainability.
By matching their innovative modeling techniques with a chemistry-infused approach, new chemical insights emerge that were once completely out of reach. Boiko and Gomes have pioneered the application of stereoelectronic effects. This new approach has opened the door to future research that may result in even more powerful breakthroughs in molecular design and reaction prediction.