William Sheffler

Joined Program: 2004
Previous Degrees: M.S. Computer Science, Brown University; B.S. Mathematics & Engineering, Brown University
Baker Lab
wsheffle (at) u.washington.edu

Research:
The first goal of my research with Dr. Baker is to determine when rosetta has generated a “correct” candidate structure for a protein or protein interaction. In a recent benchmark of de novo protein structure prediction, rosetta was used to predict the strcuture of 16 small proteins. For 5 of the 16 proteins, rosetta was able to generate correct structures which matched the x-ray crystal structure down to the positioning of individual amino acid side chains. This shows that rosetta can generate correct protein structure predictions in some cases; unfortunately, there is currently no reliable way to decide whether structure prediction in a given case was successful. We will be collaborating with noble lab to develop a machine learning approach to determine the success or failure of a protein structure prediction or protein-protein docking attempt.

A second and related goal is to create local measures of protein structure quality. Many structures generated by rosetta are correct within a region of the protein, but not correct over the entire structure. If correct pieces of a structure could be detected, this information could be used to guide future sampling and help to generate globally correct structures. Detection of correct pieces could used in de novo protein structure prediction as the basis of a divide and conqueror approach, where whole structures are built up out of smaller components previously determined to have likely correct structure. A local quality measure would also be useful in homology modeling and protein design, where it is important to know which regions of a structure are correct and which require further refinement.

Both of my research goals involve analysis of the quality of protein structures. Many measures of structure quality are already available in rosetta in the score functions used in various optimization procedures. These scores used in optimization must be fast to compute and pairwise additive, and previous work has focused on quality measure which satisfy these constraints. I will be developing metrics of protein quality to be used after optimizations are complete, which can thus afford to be more sophisticated and expensive to compute. For example, I have considered patterns of unsatisfied hydrogen bond donor and acceptor groups buried within the core of a protein and inaccessible to solvent, as well as the presence of voids or cavities in a protein core. Such measures complement preexisting scores used in rosetta and should help in the identification of both globally and locally correct protein strucures.