AlphaMissense is better correlated with functional assays of missense impact than earlier prediction algorithms.
AlphaMissense is better correlated with functional assays of missense impact than earlier prediction algorithms.
复制标题
与早期的预测算法相比，AlphaMissense 与错义影响的功能分析具有更好的相关性。
DOI：
10.1101/2023.10.24.562294
复制
发表时间：
2023
期刊：
影响因子：
--
通讯作者：
Sanders,StephanJ
中科院分区：
文献类型：
--
作者：
Ljungdahl,Alicia;Kohani,Sayeh;Page,NicholasF;Wells,EloiseS;Wigdor,EmilieM;Dong,Shan;Sanders,StephanJ
改变编码蛋白质中单个氨基酸的错义变体导致许多人类疾病，但在解释方面提出了实质性挑战。虽然这些变异体可以通过测序可靠地鉴定，但区分临床上显著的变异体仍然很困难，使得“未知意义的变异体”超过被分类为“致病性”或“可能致病性”的变异体。已经开发了许多计算机方法来预测错义变体的功能影响，以告知临床解释，最新的是AlphaMissense，它使用在预测的蛋白质结构上训练的人工智能方法。为了独立评估AlphaMissense和其他38个错义严重性预测因子的性能，我们将预测结果与来自变异效应多重测定（MAVE）的数据进行了比较。MAVE实验产生基因中几乎所有可能的单个氨基酸变化，并使用高通量测定法测量其功能影响。通过评估5个基因（DDX 3X、MSH 2、PTEN、KCNQ 4和BRCA 1）的17,696个变异，我们发现AlphaMissense始终是基于功能影响相关性的前5大算法之一，并且是两个基因的最佳相关算法。我们的结论是，AlphaMissense代表了当前最好的同类预测器，但是，与其他算法相比，改进是适度的。我们注意到，包括AlphaMissense在内的多个错义预测因子似乎将变异过度称为致病性，尽管功能影响最小，并且需要更高质量的训练数据，包括一致分析的患者队列和MAVE分析，以提高准确性。
Missense variants that alter a single amino acid in the encoded protein contribute to many human disorders but pose a substantial challenge in interpretation. Though these variants can be reliably identified through sequencing, distinguishing the clinically significant ones remains difficult, such that “Variants of Unknown Significance” outnumber those classified as “Pathogenic” or “Likely Pathogenic.” Numerous in silico approaches have been developed to predict the functional impact of missense variants to inform clinical interpretation, the latest being AlphaMissense, which uses artificial intelligence methods trained on predicted protein structure. To independently assess the performance of AlphaMissense and 38 other predictors of missense severity, we compared predictions to data from multiplexed assays of variant effect (MAVE). MAVE experiments generate almost every possible individual amino acid change in a gene and measure their functional impact using a high-throughput assay. Assessing 17,696 variants across five genes (DDX3X, MSH2, PTEN, KCNQ4, and BRCA1), we find that AlphaMissense is consistently one of the top five algorithms based on correlation with functional impact and is the best-correlated algorithm for two genes. We conclude that AlphaMissense represents the current best-in-class predictor by this metric; however, the improvement over other algorithms is modest. We note that multiple missense predictors, including AlphaMissense, appear to overcall variants as pathogenic despite minimal functional impact and that substantially more high-quality training data, including consistently analyzed patient cohorts and MAVE analyses, are required to improve accuracy.