As volume and complexity of macromolecules increase, theories and algorithms that deal with structure determination at low X-ray resolution are of particular importance. With limited diffraction data in hand, experimentalists rely on advanced computational tools to extract and utilize useful information, seeking to determinate a three dimensional model that best fits the experiment data. Success of further studies on the property and function of a specific molecule - the key to practical applications - is therefore heavily dependent on the validity and accuracy of the solved structure. In this thesis I propose Deformable Complex Network (DCN) and introduce Normal Mode Analysis (NMA), which are designed to model the average coordinates of atoms and associated fluctuations, respectively. Their applications on structure determination target two major branches ? the positional refinement and temperature factor refinement. I demonstrate their remarkable performance in structure improvements based on several criteria, such as the free R value, overfitting effect and Ramachandran Statistics, with tests carried out across a broad range of real systems for generality and consistency.