Structural biology hides many puzzles that seem very difficult to solve by using pure statistical physics approaches. For example, the question of how a protein sequence adopts its 3D shape arose more than a half-century ago, with practical solutions being just discovered thanks to the advances in multiple disciplines, most notably in machine learning. Most often, proteins do not act alone and function via interactions with other molecules. Sometimes they form stable complexes, which can be homo- or heteromeric. They can even organize in higher-order assemblies. And very often, these assemblies follow strict symmetrical principles. Thus, our understanding of these principles will help us understand the physics of life and hopefully will pave the way to designing new macromolecular machines. Protein structures under physiological conditions are neither rigid - indeed, proteins often perform their function by changing conformational states or regulating the amplitude of fluctuations upon binding. Describing and predicting their internal motions will be the next frontier of structural biology and bioinformatics. My thesis presents studies on modeling protein structure, interactions, and flexibility. I will describe several new methods to analyze and predict symmetrical protein assemblies. I will also introduce a novel approach to modeling protein motions. I will then explain our procedure to compute small-angle scattering profiles and accordingly optimize molecular shapes. Finally, I will review our developments on machine and deep learning applied to protein structures and interactions and give a future perspective.