Molecular Modeling of TDP-43 Interaction with RNA Oligonucleotides

Abstract

TDP-43 is an RNA-binding protein pathophysiologically implicated in amyotrophic lateral sclerosis (ALS) and frontotemporal dementia (FTD). Experiments by Mann et al. have demonstrated doses of "bait oligonucleotides" with high affinity for TDP-43 can abolish neurotoxicity by dissolving pathological TDP-43 condensates in the cytosol [1]. To increase the odds of translational success for this technology, growing and diversifying the group of sequences known to bind TDP-43 with high affinity is desirable. Computational screening is a promising method for prudent identification of sequences to test experimentally, saving time and money. In this thesis, we explore the use of observables from all-atom molecular dynamics simulations to score the bound poses of TDP-43 to various RNA sequences. Coarse-grained unbinding simulations and AlphaFold uncertainty measurements were also examined. The results indicate that all-atom RMSF correlates the best with experimental Kd (Pearson approx. 0.7). AlphaFold also weakly correlated (Pearson approx. 0.5), while coarse-grained simulations possessed no correlation. Temperature replica exchange simulations suggested that vanilla simulations do not fully sample the range of bound conformations for RNA, which could interfere with correlation of observables and Kd. Notably, vanilla simulations were able to persistently model residue interactions of known experimental significance (pi stacking and hydrogen bonding between TDP-43 and RNA), which may have contributed signal to the RSMF correlation. Our results suggest the RMSF has potential for use in a pose scoring function, but that it is insufficient alone to predict Kd. Due to noise in experimental affinity values and poor sampling in vanilla simulations, more high-quality experimental Kd data is imperative, in order to build more complex MD-based scoring functions while avoiding overfitting.