Using massively parallel simulation to investigate affinity maturation of designed miniprotein binders: Insights from Markov state models and free energy calculations | Poster Board #100

Pioneering work from the Baker lab at University of Washington has resulted in de novo designed miniproteins that bind the H1N1 hemagglutinin stem protein (HA2), exemplifying a promising new avenue for treating infectious disease. This exhaustive binder design and screening process yielded affinity matured (AM) binders of nanomolar affinity. Interestingly, the AM variants have mutations at non-interfacial residues. Additionally, Rosetta site saturation mutagenesis (SSM) predictions differ significantly from experimental SSM. Can all-atom simulation methods better explain enhanced binding affinity in these systems? We are using massively parallel simulation deployed on Folding@home (FAH) to investigate the conformational dynamics of these binding reactions, as well as do higher resolution in silico SSM via the expanded ensemble method (EE).
Our Markov state model analysis of GPU-accelerated binding simulations for three pairs of HA2 binders shows the AM binders to have faster k_ons than the wild type (WT) binders. The AM binders also have access to additional conformational states in complex with HA2, which may explain some of the enhanced affinity as well. To calculate k_offs for these systems we are using simulations with a repulsive gaussian bias between the binders and HA2, both unbound and in complex. We are using multiensemble Markov models to analyze both our unbiased and biased simulations. Preliminary results from these models show that the k_offs do not change much between the binders, but this work is ongoing.
Using ~42000 EE simulations, we performed in silico SSM on the WT binders. For comparison, we inferred experimental binding affinities from the reported convoluted FACS SSM using a Bayesian method. We also recalculated the Rosetta SSM using the more recent Flex ddG method. From this dataset, we are able to identify sources of uncertainty in the EE method. Further, we determine that EE more accurately predicts significant mutation effects and average residue position fitness compared to Flex ddG. Attempting to rationalize multiple mutation effects, we ran simulated tempering simulations in the experimental SSM sequence space for each binder. Shannon entropy mutability profiles built from these results show AM mutant positions having higher Shannon entropy. These studies move towards a framework for in silico protein design with atomic resolution understanding of miniprotein binding reactions and how mutation effects those reactions.