Molecular Modelling in the Machine Learning Age

by Lukas Müllender

From recommending what movie to watch, to helping diagnose diseases and driving cars, machine learning (ML) and artificial intelligence (AI) continue to change the way we live and work. By uncovering patterns in complex data, these technologies have transformed industries and become essential tools across many fields. Science, of course, is no exception. Researchers are using AI to tackle some of the most challenging questions, opening doors to new discoveries at an unprecedented pace.

I am working in the field of molecular modelling, where we try to understand and predict how molecules behave—a task critical to everything from drug discovery to materials science, and I thought I’d take you on a (non-exhaustive) tour about how AI and ML are changing the way we do science.

Generative Models: Mapping Molecular Possibilities
Generative models, like generative adversarial networks (GANs), normalizing flows or diffusion models, are a class of AI that excels at generating new data by drawing samples from a learned distribution. In molecular modelling, they’re being used to predict equilibrium molecular structures—essentially, the most stable arrangements of atoms in a molecule—and calculate their thermodynamic properties. Think of it like this: rather than painstakingly simulating every possible molecular configuration, generative models learn from existing data to predict which structures are likely to exist and which are thermodynamically favorable.

Last year’s Nobel prize in Chemistry was awarded to Demis Hassabis and John Jumper from Google DeepMind, who in 2020 presented AlphaFold2, which posed a breakthrough in protein structure prediction. While not technically belonging to the class of model architectures mentioned above, it was successfully used by researchers to predict the 3D structure of virtually all proteins found in the human body, with unprecedented accuracy. AlphaFold, and similar models like Boltzmann generators, are helping scientists design new drugs by predicting structures that meet specific requirements, such as stability or binding efficiency.

Predicting Quantum-Mechanical Energies and Forces
In molecular systems, understanding the quantum-mechanical quantities—like energy and forces acting on atoms—is key to predicting how molecules will behave. Traditionally, this requires solving the Schrödinger equation, a process so computationally intensive that it is impractical for anything but small systems. To solve this dilemma, researchers developed AI models trained on quantum-mechanical calculations (such as those from density functional theory), which can predict energies and forces with remarkable accuracy, but at a fraction of the computational cost. Popular models like ANI and SchNet have made it possible to model large and complex molecular systems quickly and efficiently. This breakthrough has far-reaching implications, from accelerating the development of new catalysts to simulating large biological molecules at quantum-mechanical accuracy.

Free Energy Surfaces and Kinetics
Understanding how molecules move and interact requires more than just knowing their structures. Free energy surfaces—maps of the energy landscapes molecules explore during their motions—offer a way to visualize these dynamics. Similarly, studying the kinetics of these motions can reveal critical insights, such as how quickly a drug binds to its target or how proteins fold into functional shapes.

AI is streamlining the process of extracting these free energy surfaces and kinetic properties. Enhanced sampling techniques, often guided by ML, allow researchers to more efficiently explore these landscapes. By reducing the time and computational resources needed to calculate these properties, ML techniques are making it possible to study larger systems over longer timescales.

A New Era of Molecular Science
AI and ML are doing more than just speeding up calculations; they’re redefining how scientists approach molecular modeling. From generative models predicting biomolecular structures to neural networks accelerating quantum calculations, these tools are unlocking new possibilities and providing deeper insights into the molecular world. As AI continues to advance, it will remain central to molecular science, providing sophisticated tools to address fundamental challenges in chemistry, biology, and materials science. Beyond enhancing efficiency, these technologies are set to drive entirely new avenues of scientific discovery.