Determination of molecular structure is commonly posed as a nonlinear optimization problem. The objective functions rely on a vast amount of structural data. As a result, the objective functions are most often nonconvex, nonsmooth, and possess many local minima. Furthermore, introduction of additional structural data into the objective function creates barriers in finding the global minimum, causes additional computational issues associated with evaluating the function, and makes physical constraint enforcement intractable. To combat the computational problems associated with standard nonlinear optimization formulations, Williams et al. (2001) proposed an atom-based optimization, referred to as GNOMAD, which complements a simple interatomic distance potential with van der Waals (VDW) constraints to provide better quality protein structures. However, the improvement in more detailed structural features such as shape and chirality requires the integration of additional constraint types.

This dissertation builds on the GNOMAD algorithm in using structural data to estimate the three-dimensional structure of a protein. We develop several methods to make GNOMAD capable of effectively and efficiently handling non-distance information including torsional angles and molecular surface data. In specific, we propose a method for using distances to effectively satisfy known torsional information and show that use of this method results in a significant improvement in the quality of α-helices and β-strands within the protein. We also show that molecular surface data in combination with our improved secondary structure estimation method and long-range distance data offer increased accuracy in spatial proximity of α-helices and β-strands within the protein, and thus provide better estimates of tertiary protein structure. Lastly, we show that the enhanced GNOMAD molecular structure estimation framework is effective in predicting protein structures in the context of comparative modeling.