Blind Predictions of RNA/Protein Relative Binding Affinities

Abstract

Interactions between RNA and proteins are pervasive in biology, shaping processes such as mRNA translation, localization, and alternative splicing. Developing a predictive understanding of the energetics of these systems would allow us to model biologically relevant mutations of these interactions and ultimately design novel interactions. Despite recent advances in high throughput experimental technologies that measure the energetics of these systems, quantitative computational prediction of relative RNA/protein binding affinities has remained a challenge. This is partly due to the observation that computational binding affinity prediction methods typically break down when the molecules are highly flexible or undergo significant conformational changes, situations that often arise in RNA/protein binding. Here, we present a novel framework within Rosetta for predicting RNA/protein relative binding affinities that begins to address this issue. Specifically, we show that the nearest neighbor energies, which are typically used for RNA secondary structure prediction, can be used to approximate the unbound free energy of the RNA, thus eliminating the need to explicitly account for the flexibility of the unbound RNA or conformational changes of the RNA upon binding. Using this method of calculating the unbound RNA free energy significantly improves the prediction accuracy over a more typical 3D structure-based approach. We optimized this method using a subset of published MS2 coat protein affinities and ultimately made predictions for the system with 1.11-1.28 kcal/mol root mean square (RMS) error. Additionally, we show that this method is able to predict relative binding affinities for four diverse RNA/protein systems with 1.48 kcal/mol RMS error. Finally, to more rigorously assess this method, we independently measured and made blind predictions for PUF3 and PUM2 binding affinities with RMS errors of 1-2 kcal/mol, which is comparable to the accuracy achieved by prediction methods for other types of systems.