Predicting protein complex binding affinity by integrating dynamic features and high-throughput screening data with deep learning

Background.

Understanding the biophysical mechanisms of protein-protein interactions (PPIs) is a fundamental question in biology. A key determinant of PPIs is their binding affinity. Despite the availability of ~100,000 complex structures providing valuable insights into structural mechanisms, the prediction of PPI binding affinity remains challenging to simulation methods as well as machine-learning models (Siebenmorgen, 2020; Guo, 2022).
The binding affinity of a PPI is the result of a complex interplay of dynamical, structural and biophysical / biochemical properties, which are insufficiently captured by currently available structural data. Most importantly, molecular flexibility described by entropy is a key thermodynamic constituent of affinity, which is not captured by static structures.

Aims and Methods.

To address and study these limitations in a defined system, we propose a reductionist approach, where we study the mutational binding landscape of Sso7d, a binding scaffold with a well-defined binding surface composed of nine surface-exposed residues located on a relatively rigid beta-sheet (Traxlmayr, 2016). We will study the mutational impact on binding affinity from two orthogonal perspectives: (i) molecular dynamics (MD) simulations to obtain variant structures and dynamics information and (ii) yeast display-based deep mutational scanning (DMS) (Traxlmayr, 2012; Fowler, 2014) to determine relative binding affinities in a high-throughput format. We will use this coherent and comprehensive dataset using both experimental and simulated data to develop and train cutting-edge deep-learning architectures allowing accurate in silico prediction of binding affinities from PPI structures.

The deep learning algorithm that will be developed will be validated using additional Sso7d-based binders, as well as PPIs based on other surface topologies. Ultimately, we expect that our model will be a highly valuable tool for the design of binding domains tailored for specific applications, as well as for studying the effect of mutations in natural systems (e.g., prediction of drug resistance mutations which disrupt binding to an inhibitor or antibody).

REFERENCES
1. Guo, Z. and R. Yamaguchi (2022) Machine learning methods for protein-protein binding affinity prediction in protein design. Front Bioinform 2, 1065703. doi: 10.3389/fbinf.2022.1065703
2. Fowler, D. M. and S. Fields (2014) Deep mutational scanning: a new style of protein science. Nat Methods 11, 801-807. doi: 10.1038/nmeth.3027
3. Siebenmorgen, T. and M. Zacharias (2020) Computational prediction of protein-protein binding affinities. Wiley Interdisciplinary Reviews-Computational Molecular Science 10, e1448. doi: 10.1002/wcms.1448
4. Traxlmayr, M. W., C. Hasenhindl, M. Hackl, G. Stadlmayr, J. D. Rybka, N. Borth, J. Grillari, F. Ruker and C. Obinger (2012) Construction of a stability landscape of the CH3 domain of human IgG1 by combining directed evolution with high throughput sequencing. J Mol Biol 423, 397-412. doi: 10.1016/j.jmb.2012.07.017
5. Traxlmayr, M. W., J. D. Kiefer, R. R. Srinivas, E. Lobner, A. W. Tisdale, N. K. Mehta, N. J. Yang, B. Tidor and K. D. Wittrup (2016) Strong enrichment of aromatic residues in binding sites from a charge-neutralized hyperthermostable Sso7d scaffold library. J Biol Chem 291, 22496-22508. doi: 10.1074/jbc.M116.741314