The Work
Schaaf tested whether atom-scale foundation models could inject physics into enzyme design. Core idea: force fields reproduce quantum-chemical energies cheaply, and their learned representations encode real chemistry rather than memorised structures, so they serve as fast physics engines and features for design models.
What Worked
A pretrained interatomic potential collapsed a one-to-two-day transition-state setup to about five minutes, now used routinely in the lab.
What Has Been Achieved
The clearest win: transition-state setup, Schaaf's implementation reduces the need for expert knowledge, takes minutes, and is now in routine use in the lab. The same transferability produced Rem3Di, a chirality-aware 3D descriptor topping or matching 6 benchmarks, and a reactivity filter for ranking designed enzymes, promising but not yet tested at scale.
What Was Learned
A transition-state search used directly as a reactivity filter was too inconsistent to rank designs, ruling out the cheapest option as a standalone filter.
About
Lars is completing a PhD in geometric deep learning at the University of Cambridge. His research centres around developing machine-learning force fields (MLFFs) that speed up atom-scale simulations for reliable property prediction of materials and proteins. His current focus on “designing from the atom up” couples generative models with physics-based constraints to propose realistic, testable candidates.
Lab & Advisors
Aron Walsh
Imperial College London
Chair in Materials Design, Department of Materials
Aron Walsh leads the Materials Design group at Imperial College London, focused on the design and optimisation of materials using high performance computing. His work combines computational materials chemistry, quantum mechanics, machine learning, and multi scale modeling.
