METHODS FOR COMPARISON OF PROTEIN THREE DIMENSIONAL STRUCTURE

我们已经开发了蛋白质三维比较的算法， 结构. “VAST”算法，即“矢量比对搜索工具”， 通过比较类型快速识别子结构相似性， 连接性和相对取向二级结构元件。 工作集中在三个方面：（1）界定门槛 相似性统计，2）灵敏度和特异性验证，以及 3)制定“最佳”逐个残基的统计标准 对齐。 VAST通过偶然出现对子结构相似性进行排序 可能性 这被计算为概率的乘积， 独立的元素对叠加残差将被观察到， 机会，根据经验分布确定， 随机抽取结构数据库中的元素对。 阈值 统计量被给出为最大子结构似然的乘积 以及可能的子结构选择的数量，由 包含离散元素个数的组合公式 每一种蛋白质 如果很小，则该值可直接解释为 显著相似性检验统计量或p值。 VAST已经 通过对已知结构的全面分析进行测试。 灵敏度 与BLAST比对的“命中率”为99.5%。 测试用例的特异性 文献似乎很完美，与现有的非- 统计方法表明，命中率较低，遗漏率较低 边缘相似性 残基-残基比对现在构建 通过吉布抽样方案，这证明比大都会更有效 取样. 评价了替代残基比对的意义 通过参考叠加残差的经验分布 随机比对，找出最令人惊讶的一个。 分析 经验随机元自比较密度函数 残差被证明上级子结构特定的空模型， 再现文献专家比对，并且另外更快。 的 这项工作的意义将是扩大视野， 进化关系可以通过使用结构 比较代替单独的序列比较。 结果令人 生物学家可以将其作为“结构邻居”， 浏览器.

We have developed algorithms for comparison of protein three dimensional structures. The "VAST" algorithm, for "vector alignment search tool," identifies substructures similarities rapidly by comparing the types, connectivity, and relative orientations secondary structure elements. Work has focused in three areas: 1) definition of the threshold similarity statistic, 2) validation of sensitivity and specificity, and 3) development of statistical criteria for "optimal" residue-by-residue alignment. VAST ranks substructure similarities by chance-occurrence likelihood. This is calculated as the product of the probabilities that independent element-pair superposition residuals would be observed by chance, as determined by reference to an empirical distribution based on random draws of element pairs in the structural database. A threshold statistic is given as the product of the greatest substructure likelihood and the number of possible substructure choices, as determined by a combinatorial formula involving the number of discrete elements present in each protein. If small, this value is interpretable directly as a test statistic or p-value for significant similarity. VAST has been tested by an all-against-all analysis of known structures. Sensitivity with respect to BLAST "hits" is 99.5%. Specificity in test cases from the literature appears perfect, and comparison to existing non- statistical methods indicates fewer hits, with omissions being low-score marginal similarities. Residue-residue alignments are now constructed by a Gibb's sampling scheme, which proved more efficient than Metropolis sampling. Significance of alternative residue alignments is evaluated by reference to an empirical distribution of superposition residuals in random alignments, to identify the most surprising one. An analytical density function fitted to empirical random-element self comparison residuals proved superior to a substructure-specific null model in reproducing literature expert alignments, and is in addition faster. The significance of this work will be in extending the horizon to which evolutionary relationship may be detected, by employing structure comparison in place of sequence comparison alone. The results have been made available to biologists as "structural neighbors" in the "Entrez" browser.