Supplementary file 1_Successful prediction of LC8 binding to intrinsically disordered proteins sheds light on AlphaFold’s black box.docx

IntroductionLC8 is a hub protein involved in many processes from tumor suppression and cell cycle regulation to neurotransmission and viral infection. Despite recent progress, prediction of binding sites for LC8 is plagued by motif variability and a multitude of weakly binding motifs, especially when binding depends on multivalency. Our binding site prediction algorithm, LC8Pred has proven useful for uncovering new LC8 binders, but is insufficient for finding all LC8 binding sites.
MethodsTo address this, we probed the ability of a general structure predictor, AlphaFold, to predict whether a given sequence binds to LC8. Certain combinations of in-built AlphaFold scores were extracted and distributions of scores of binders were compared to scores of nonbinders.
ResultsAlphaFold successfully places proteins at the correct interface of LC8. A set of threshold values of built-in AlphaFold scores enables differentiation between known binders and nonbinders with minimal false positive (8%) and acceptable false negative rates (20%). This cutoff, along with a more inclusive cutoff, was used to predict elusive LC8 binding sites in proteins known to bind LC8.
DiscussionCorrelations between binding affinities and AlphaFold scores provide insight into the black box and indicate that AlphaFold learned an inaccurate energy function that nevertheless is useful for making inferences and conclusions about physical systems. Binding sites predicted by this method can be prioritized for investigation by comparing to result by LC8Pred, local structure, and evolutionary conservation.

引言：LC8是一种枢纽蛋白（hub protein），参与从肿瘤抑制、细胞周期调控到神经传递以及病毒感染等诸多生物学过程。尽管近年来研究取得一定进展，但LC8结合位点的预测仍受困于基序变异性与大量弱结合基序的问题，尤其是当结合依赖多价相互作用时。本团队开发的结合位点预测算法LC8Pred虽已被证实可有效发掘新型LC8结合蛋白，但仍无法完整识别所有LC8结合位点。
方法：为解决上述问题，我们探究了通用结构预测工具阿尔法折叠（AlphaFold）预测特定序列是否能与LC8结合的能力。我们提取了阿尔法折叠内置的若干评分组合，并对比了结合蛋白与非结合蛋白的评分分布。
结果：阿尔法折叠可成功将蛋白定位至LC8的正确结合界面。通过设置一组阿尔法折叠内置评分的阈值，能够以极低的假阳性率（8%）与可接受的假阴性率（20%）区分已知的结合蛋白与非结合蛋白。我们利用该阈值及更宽松的阈值，在已证实可结合LC8的蛋白中预测出了难以被发现的LC8结合位点。
讨论：结合亲和力与阿尔法折叠评分之间的相关性为解析这一黑箱模型提供了思路，结果表明阿尔法折叠学习得到的能量函数虽不够精准，但仍可用于对物理系统进行推断与总结。通过将本方法预测的结合位点与LC8Pred的结果、局部结构以及进化保守性进行对比，可以优先选择这些位点开展后续研究。