A knowledge-based system for macromolecular model building and evaluation

This dissertation focuses on the development and application of automated methods to help biologists model the structure of macromolecular complexes from experimentally derived structural information and evaluate proposed models. In particular, we assert that general purpose knowledge-based systems can support the construction and evaluation of models. For RIBOWEB, the prototype system, we use the macromolecular complex of the ribosome as the domain.; In this document, we describe the RIBOWEB system and present our work applying RIBOWEB functionality to the construction of tRNA models in the ribosome and the evaluation of several ribosomal models and crystal structures with respect to published experimental structural data. We show that a knowledge-based system can indeed be used to support macromolecular modeling and evaluation.; We propose a model of A-site and P-site tRNAs in the ribosome based solely on experimental structural information. We found that valuable structural information for modeling can be generated by tethered hydroxyl radical probing experiments. Our tRNA model was created before the high-resolution crystal structure was available and is in agreement with it though created from totally independent data.; We also provide an evaluation of the applicability and reliability of experimental structural information for low resolution ribosomal modeling. We performed comparisons of ribosomal models with experimental structural information and identified several characteristics of experimental data that could be useful for future modeling. These characteristics include large amounts of variation in structural information from crosslinking experiments and large distance ranges represented by structural information from footprinting experiments. We also performed comparisons of ribosomal crystal structures with experimental structural information. We found that most distances in the ribosome derived from experimental structural information were small, but that the distribution of distances was quite large, implying that loose distance assignments to experimental data is important when modeling with experimental data from multiple sources.