Deep learning models like RoseTTAFold and AlphaFold2 allow highly accurate protein structure prediction, but large protein assemblies remain hard to predict because of their size and complex subunit interactions. In a study published in Nature Methods, Shor and Scheidman-Duhovny introduce CombFold, a combinatorial and hierarchical assembly algorithm that predicts structures of large protein complexes by making use of pairwise interactions between subunits predicted by AlphaFold2.

The CombFold assembly algorithm contains three stages: generating pairwise interactions with AlphaFold2, creating a unified representation of subunits and interactions, and combinatorially assembling subunits. The authors validated the approach on two large heteromeric benchmark datasets. CombFold accurately modeled 72% of the complexes among the top 10 and 62% among the top 1 predictions.

RoseTTAFold 和 AlphaFold2 等深度学习模型可以实现高度准确的蛋白质结构预测，但大型蛋白质组装体由于其大小和复杂的亚基相互作用仍然难以预测。在《自然方法》上发表的一项研究中，Shor 和 Scheidman-Duhovny 介绍了 CombFold，这是一种组合和分层组装算法，可通过利用 AlphaFold2 预测的亚基之间的成对相互作用来预测大型蛋白质复合物的结构。

CombFold 组装算法包含三个阶段：与 AlphaFold2 生成成对相互作用、创建子单元和相互作用的统一表示以及组合组装子单元。作者在两个大型异聚基准数据集上验证了该方法。CombFold 准确地模拟了前 10 个预测中 72% 的复合体和前 1 个预测中的 62%。