Final Report Summary - SOUPINMYCRYSTAL (How can we improve our models of biological macromolecules to reproduce experimental crystallographic X-ray intensities better?)
The fellow was awarded the Marie Curie Career Development Fellowship so that she could move from Germany to the group of Dr. Garib Murshudov at the MRC Laboratory of Molecular Biology, Cambridge, in order to conduct research targeting the question:
How can we improve our models of biological macromolecules in order to better reproduce experimental crystallographic X-ray intensities?
Today, we have models for approximately 110,000 macromolecular structures, most of which were derived from X-ray diffraction experiments. In X-ray structure determination, the R-factor reports how well a given model agrees with experimental data (with a lower percentage indicating a model more consistent with observed data). In small molecule crystallography, R-factors of around 3% are routinely reached. However, for biological macromolecules, R-factors are normally around 20%, even when data extend to atomic resolution. Evidently, our current models of macromolecular crystal structures have severe shortcomings. In this project, the reasons for this discrepancy were researched, with the objective of improving parameterisation thus allowing model quality to be routinely improved. By doing so, not only would every known structure potentially benefit, but it may also enable the solution of borderline cases and “problematic” data sets.
Structure factors, which are Fourier coefficients of the electron density map calculated from an atomic model, are complex numbers with a amplitude and phase. The amplitudes are measured almost directly; the square of an amplitude is proportional to the observed X-ray intensity measured on the detector. The phase, however, cannot generally be experimentally observed. Phase estimates are usually gained by assuming certain properties of the crystal structure, and hence phase estimation may obscure any details we do not know.
In this project, differences between models and corresponding data were examined in two ways:
1) Gaining phases without implying a model.
2) Looking at the differences between observed and calculated amplitudes, ignoring phases.
For the first objective, a suitable way to do this was discovered during the course of the project. The necessary collaborations were initiated, a measurement strategy established, and the first set of phases have been collected. However, the procedure still needs to be improved, as only 78% of the current model-free phases are estimated to be sufficiently accurate for the purpose.
As for the second objective, we observed a resolution-dependent behaviour of the difference between calculated and observed structure factor amplitudes. We were able to exclude a number of potential sources, including: the specific program used to refine the model; the detector used to record the X-ray data; and the slight disorder inherent to crystallized macromolecules. The behaviour is observed to be very similar in structures of different size, symmetry, resolution, and also present for both proteins and nucleic acids. However, it seems not to occur for structures of smaller molecules. It could be speculated that the scattering of the solvent inherent in macromolecular crystals – a disordered “soup” of water, ions, cryo protectants and other additives – might be responsible for this effect. Results indicate that, in some cases, modelling this behaviour could lead to R-factor improvements of over 4% in macromolecules.
In addition, we are also revising validity of using R-factors as quality indicators. Whilst they are used throughout all crystallographic calculations, simple inspection of the relevant equations shows that R-factors are strongly dependent on some overall properties of crystals, such as the overall B value. We are analysing the robustness of alternative quality indicators, such as average correlation, which has less dependence on overall crystal properties.
In addition to the scientific objectives mentioned above, the purpose of this fellowship was also to allow the fellow to become an independent researcher, able to build a team, and to further knowledge exchange across national borders. The fellow has gained proficiency in several computing languages, crystallographic software development, and learned new statistical approaches. In order to conduct this research, she also set up collaborations spanning Australia, Germany, UK, the United States, Korea and France.
The fellow, who has reached professional maturity and independence in the two years, was able to secure a position as Investigator Scientist at the MRC Laboratory of Molecular Biology after the fellowship, but plans to eventually set up her own group back in Germany.
How can we improve our models of biological macromolecules in order to better reproduce experimental crystallographic X-ray intensities?
Today, we have models for approximately 110,000 macromolecular structures, most of which were derived from X-ray diffraction experiments. In X-ray structure determination, the R-factor reports how well a given model agrees with experimental data (with a lower percentage indicating a model more consistent with observed data). In small molecule crystallography, R-factors of around 3% are routinely reached. However, for biological macromolecules, R-factors are normally around 20%, even when data extend to atomic resolution. Evidently, our current models of macromolecular crystal structures have severe shortcomings. In this project, the reasons for this discrepancy were researched, with the objective of improving parameterisation thus allowing model quality to be routinely improved. By doing so, not only would every known structure potentially benefit, but it may also enable the solution of borderline cases and “problematic” data sets.
Structure factors, which are Fourier coefficients of the electron density map calculated from an atomic model, are complex numbers with a amplitude and phase. The amplitudes are measured almost directly; the square of an amplitude is proportional to the observed X-ray intensity measured on the detector. The phase, however, cannot generally be experimentally observed. Phase estimates are usually gained by assuming certain properties of the crystal structure, and hence phase estimation may obscure any details we do not know.
In this project, differences between models and corresponding data were examined in two ways:
1) Gaining phases without implying a model.
2) Looking at the differences between observed and calculated amplitudes, ignoring phases.
For the first objective, a suitable way to do this was discovered during the course of the project. The necessary collaborations were initiated, a measurement strategy established, and the first set of phases have been collected. However, the procedure still needs to be improved, as only 78% of the current model-free phases are estimated to be sufficiently accurate for the purpose.
As for the second objective, we observed a resolution-dependent behaviour of the difference between calculated and observed structure factor amplitudes. We were able to exclude a number of potential sources, including: the specific program used to refine the model; the detector used to record the X-ray data; and the slight disorder inherent to crystallized macromolecules. The behaviour is observed to be very similar in structures of different size, symmetry, resolution, and also present for both proteins and nucleic acids. However, it seems not to occur for structures of smaller molecules. It could be speculated that the scattering of the solvent inherent in macromolecular crystals – a disordered “soup” of water, ions, cryo protectants and other additives – might be responsible for this effect. Results indicate that, in some cases, modelling this behaviour could lead to R-factor improvements of over 4% in macromolecules.
In addition, we are also revising validity of using R-factors as quality indicators. Whilst they are used throughout all crystallographic calculations, simple inspection of the relevant equations shows that R-factors are strongly dependent on some overall properties of crystals, such as the overall B value. We are analysing the robustness of alternative quality indicators, such as average correlation, which has less dependence on overall crystal properties.
In addition to the scientific objectives mentioned above, the purpose of this fellowship was also to allow the fellow to become an independent researcher, able to build a team, and to further knowledge exchange across national borders. The fellow has gained proficiency in several computing languages, crystallographic software development, and learned new statistical approaches. In order to conduct this research, she also set up collaborations spanning Australia, Germany, UK, the United States, Korea and France.
The fellow, who has reached professional maturity and independence in the two years, was able to secure a position as Investigator Scientist at the MRC Laboratory of Molecular Biology after the fellowship, but plans to eventually set up her own group back in Germany.