Korban Svetlana A, Mikhailovskii Oleg, Gurzhiy Vladislav V, Podkorytov Ivan S, Skrynnikov Nikolai R
Laboratory of Biomolecular NMR, St Petersburg State University, St Petersburg, 199034, Russian Federation.
Utrecht, The Netherlands.
IUCrJ. 2025 Jul 1;12(Pt 4):488-501. doi: 10.1107/S2052252525005123.
In this report, we describe a set of structures of the engineered protein LCB2 that has been solved starting from different computer-predicted molecular replacement (MR) models. We found that AlphaFold3, AlphaFold2, MultiFOLD, Rosetta, RoseTTAFold and trRosetta all produced successful MR models for this three-helix bundle 58-residue protein, while some of the older predictors failed. To assign B factors in the MR models we used the predictor-generated confidence scores or, as a convenient alternative, the accessible surface area (ASA) values. The process of multi-start structure determination using Coot and Phenix demonstrated good convergence, leading to six structures within 0.25 Å (all-atom RMSD) of each other. Of note, structural differences between the computer-predicted MR models and the final structures can be largely attributed to a single specific crystal contact. Comparing the six structural solutions, we observe that a number of surface side chains have been solved with different conformations. Interestingly, for each individual structure the electron density is consistent with a single rotameric state and offers no direct evidence of conformational heterogeneity. Strictly speaking, this behavior constitutes a case of model bias; we argue, however, that it represents a benign side of model bias. Specifically, when we use a model where the side-chain conformation corresponds to one of the actual (significantly populated) rotameric states, this leads to an enhancement of the electron density for this particular conformation. Conversely, when we use a model with an irrelevant (low-population) side-chain conformation, it fails to produce the matching electron density. We thus conclude that the six LCB2 structures obtained in this study can be grouped into a multiconformer ensemble, where structural variations are representative of protein's conformational dynamics. Indeed, using this six-member ensemble leads to a significant drop in R and R compared with the individual solutions. This interpretation was also supported by our MD simulations of the LCB2 crystal.
在本报告中，我们描述了从不同的计算机预测分子置换（MR）模型出发解析得到的工程蛋白LCB2的一组结构。我们发现，AlphaFold3、AlphaFold2、MultiFOLD、Rosetta、RoseTTAFold和trRosetta都为这个由58个残基组成的三螺旋束蛋白生成了成功的MR模型，而一些较旧的预测工具则失败了。为了在MR模型中分配B因子，我们使用了预测工具生成的置信度分数，或者作为一种方便的替代方法，使用可及表面积（ASA）值。使用Coot和Phenix进行多起始结构测定的过程显示出良好的收敛性，得到了六个相互之间全原子均方根偏差（RMSD）在0.25 Å以内的结构。值得注意的是，计算机预测的MR模型与最终结构之间的结构差异在很大程度上可归因于一个特定的晶体接触。比较这六个结构解决方案，我们观察到一些表面侧链以不同的构象被解析出来。有趣的是，对于每个单独的结构，电子密度与单一的旋转异构体状态一致，并且没有提供构象异质性的直接证据。严格来说，这种行为构成了模型偏差的一个例子；然而，我们认为这代表了模型偏差的一个良性方面。具体而言，当我们使用一个侧链构象对应于实际（大量存在）旋转异构体状态之一的模型时，这会导致该特定构象的电子密度增强。相反，当我们使用一个具有不相关（低丰度）侧链构象的模型时，它无法产生匹配的电子密度。因此，我们得出结论，本研究中获得的六个LCB2结构可以归为一个多构象集合，其中的结构变化代表了蛋白质的构象动力学。事实上，与单个结构相比，使用这个由六个结构组成的集合会导致R和R显著下降。我们对LCB2晶体的分子动力学（MD）模拟也支持了这一解释。
