Bone, R., Fujishige, A., Kettner, C.A., and Agard, 
D.A. 1991. Structural basis for broad specificity 
in a-Iytic protease mutants. Biochemistry 30: 
10388-10398. 
Fujishige, A., Smith, K.R., Silen, J.L., and Agard, 
D.A. 1992. Correct folding of a-lytic protease is 
required for its extracellular secretion from Esch- 
erichia coli.J Cell Biol 1 18:33-42. 
Holmes, T.J., Liu, Y.-H., Khosla, D., and Agard, 
D.A. 1991. Increased depth of field and stereo 
pairs of fluorescence micrographs via inverse fil- 
tering and maximum-likelihood estimation. /Af^- 
croscopy 164:217-237. 
Paddy, M.R., Agard, D.A., and Sedat, J.W. 1992. 
An extended view of nuclear lamin structure, 
function, and dynamics. Semin Cell Biol 3:255- 
266. 
Wilson, C, and Agard, D.A. 1991. Engineering 
substrate specificity. Curr Opin Struct Biol 
1:617-623. 
COMPUTER SIMULATION OF BIOLOGICAL MACROMOLECULES 
AxelT. Brunger, Ph.D., Assistant Investigator 
Dr. Briinger's research focuses at the interface be- 
tween theory and experiment in structural biophys- 
ics. The current elforts center on studies of macro- 
molecular structure, interaction and energetics, and 
methodological developments in x-ray crystallogra- 
phy and nuclear magnetic resonance (NMR) spec- 
troscopy. The research tools are a variety of simu- 
lation methods for macromolecules, including 
molecular dynamics, conformational searching, and 
simulated annealing. 
Crystallography: Assessment of Atomic 
Models and Phases 
In the past decade macromolecular crystallogra- 
phy has undergone major advances in crystalliza- 
tion, data collection by synchrotron x-ray sources 
and area detectors, and data analysis by high- 
performance computers and new computational 
techniques. In addition, recombinant gene technol- 
ogy often allows the expression of large amounts of 
protein. This has resulted in an unprecedented 
increase in the number of protein structures 
elucidated. 
Despite these successes, the fundamental prob- 
lem in x-ray crystallography, the phase problem, re- 
mains unchanged. That is, from a monochromatic 
diffraction of a single crystal, it is possible to obtain 
the amplitudes but not the phases of the reflections, 
whereas construction of the electron density by 
Fourier transformation requires both components of 
the complex structure factors. Phase information 
has to be obtained through experimental ap- 
proaches, most commonly multiple isomorphous re- 
placement or knowledge-based procedures referred 
to as Patterson search or molecular replacement. 
Phase information obtained through these tech- 
niques is usually of limited accuracy and resolu- 
tion, making it difficult sometimes to interpret elec- 
tron density maps in certain regions of the mole- 
cule. Furthermore, macromolecular crystals usually 
diffract to less than atomic resolution, compromis- 
ing the process of fitting an atomic model to the 
observed intensities. 
Clearly, it is imperative to obtain crystal struc- 
tures with maximum correctness and accuracy. This 
applies to their interpretation in terms of macromo- 
lecular function and interaction, to their incorpora- 
tion into databases that are the foundation for struc- 
ture prediction and other theoretical studies, and to 
their use in designing drugs on the basis of struc- 
tural information. 
Cross-validation, a tool of modern statistics, can 
be used to assess the quality of a model that is fitted 
against observed data. The idea behind cross- 
validation is deceptively simple: it consists of omit- 
ting a certain fraction of the data while the atomic 
model is fitted to the remaining data. The agreement 
between the model and the omitted data serves as an 
unbiased measure of the quality of the fit. Dr. 
Brunger realized that cross-validation could be ap- 
plied to the traditional measure of the fit between an 
atomic model and the diffraction data: the /? value. 
Despite stereochemical restraints, it is possible to 
overfit, or misfit, the diffraction data: an incorrect 
model can be refined to fairly good conventional R 
values, as several publications of (partially) incor- 
rect crystal structures have shown. The "free" R 
value is defined as the agreement between observed 
and computed structure-factor amplitudes for the 
omitted test reflections. 
Examples showed that the free R value is a much 
better measure than the conventional R value for 
distinguishing between correct and incorrect 
atomic models. Furthermore, the free 7? value can be 
used to assess the improvement in the fit of an 
atomic model to the observed diffraction data by 
STRUCTURAL BIOLOGY 463 
