Towards Energetics-Based de novo Enzyme Design: Elucidating Enzymatic Reactions and Estimating Their Activation Barriers Using Machine-Learned Force Fields

Abstract

Enzymes catalyze nearly every chemical transformation in biology, but natural enzymes evolved for specific cellular substrates and environments. Designing enzymes de novo would enable catalysts for reactions that are poorly served by nature, with applications in selective synthesis, environmental remediation, and sustainable chemical manufacturing. Recent advances in generative protein design have improved the ability to build protein scaffolds around prescribed catalytic motifs, yet many designs that satisfy structural and geometric filters fail to translate into experimentally active enzymes. Catalysis is fundamentally an energetic process, and this gap suggests that new, energetics-based filters could improve the yield of enzyme design campaigns. This thesis investigates whether activation-barrier calculations with state-of-the-art machine learned force fields can provide a missing layer for de novo enzyme design. I use MACE-POLAR, a state-of-the-art machine-learned force field with near-ab initio accuracy, together with metadynamics to estimate reaction free-energy barriers in enzyme active sites. I develop a workflow for preparing active-site crops, assigning protonation and charge states, annotating reactive atoms and collective variables, and analyzing free-energy surfaces. This framework is applied to three enzyme classes for which de novo design has recently produced experimentally active enzymes: cysteine proteases, metalloproteases, and organophosphotriesterases. Across these case studies, the calculations reveal mechanistic information that is not accessible from static structural filters alone. In cysteine proteases, the simulations show that reactivity depends strongly on the initial substrate pose and protonation states. In metalloproteases, computed barriers reproduce activity trends among designed variants but also reveal the sensitivity of free-energy estimates to active-site cropping, solvent exposure, and boundary conditions. In organophosphotriesterases, metadynamics provides a discriminative signal between active and inactive variants through differences in sampling behavior, and solvated simulations reveal a possible solvent-assisted reaction pathway absent from the proposed reaction mechanism. These results support the use of energetics calculations not only as candidate-design scores, but also as mechanistic diagnostics. Activation-barrier calculations can identify why structurally plausible designs fail, reveal missing chemical ingredients, and suggest alternative catalytic geometries or reaction pathways. At the same time, the present workflow is a proof of concept rather than a fully validated high-throughput screening method. It remains sensitive to protonation states, substrate docking, collective-variable choice, simulation convergence, solvent treatment, and the approximation of the full enzyme by an active-site crop. Nevertheless, this work demonstrates that incorporating energetics into the protein design pipeline can help bridge the gap between geometrically plausible de novo enzymes and experimentally active catalysts.