利用统计知情图网络精确预测蛋白质功能-动脉网

Accurate prediction of protein function using statistics-informed graph networks

Nature 等信源发布 2024-08-04 13:26



可切换为仅中文







AbstractUnderstanding protein function is pivotal in comprehending the intricate mechanisms that underlie many crucial biological activities, with far-reaching implications in the fields of medicine, biotechnology, and drug development. However, more than 200 million proteins remain uncharacterized, and computational efforts heavily rely on protein structural information to predict annotations of varying quality.

摘要了解蛋白质功能对于理解许多关键生物活动的复杂机制至关重要，在医学，生物技术和药物开发领域具有深远的意义。。

Here, we present a method that utilizes statistics-informed graph networks to predict protein functions solely from its sequence. Our method inherently characterizes evolutionary signatures, allowing for a quantitative assessment of the significance of residues that carry out specific functions. PhiGnet not only demonstrates superior performance compared to alternative approaches but also narrows the sequence-function gap, even in the absence of structural information.

在这里，我们提出了一种利用统计信息的图形网络仅从其序列预测蛋白质功能的方法。我们的方法固有地表征了进化特征，从而可以定量评估执行特定功能的残基的重要性。与替代方法相比，PhiGnet不仅表现出优异的性能，而且即使在没有结构信息的情况下也缩小了序列功能差距。

Our findings indicate that applying deep learning to evolutionary data can highlight functional sites at the residue level, providing valuable support for interpreting both existing properties and new functionalities of proteins in research and biomedicine..

我们的研究结果表明，将深度学习应用于进化数据可以突出残基水平的功能位点，为解释研究和生物医学中蛋白质的现有特性和新功能提供有价值的支持。。

IntroductionProteins bind to other molecules to facilitate nearly all essential biological activities. Consequently, understanding protein function is of paramount importance for comprehending health, disease, evolution, and the functioning of living organisms at the molecular level1,2,3. The primary sequence of a protein contains all the essential information required to fold up into a particular three-dimensional shape, thereby determining its activities within cells 4,5.

简介蛋白质与其他分子结合以促进几乎所有必需的生物活性。因此，了解蛋白质功能对于在分子水平上理解健康，疾病，进化和生物体的功能至关重要1,2,3。。

The evolutionary information in massive protein sequences that are gleaned from extensive genome sequencing efforts has significantly contributed to recent advances in protein structure prediction6,7,8,9. This evolutionary data, especially the couplings between pairwise residues, has also been utilized to characterize protein functional sites10,11.

从广泛的基因组测序工作中收集到的大量蛋白质序列的进化信息为蛋白质结构预测的最新进展做出了重大贡献6,7,8,9。这种进化数据，特别是成对残基之间的偶联，也被用于表征蛋白质功能位点10,11。

The evolutionary couplings have been utilized to pinpoint functional sites in proteins, capturing interactions between residues that contribute to specific functions5,12. Indeed, the analysis of evolutionary information has allowed the identification of allosteric mechanisms in proteins13,14, disease variants15, and metamorphism in proteins that undergo reversible switches between distinct folds, often accompanied by different functions16.To date, more than 356 million proteins in the UniProt database17 (6/2023) have been sequenced and the vast majority (~80%) of these have no known functional annotations (e.g., enzyme commission numbers and gene ontology terms).

进化偶联已被用于精确定位蛋白质中的功能位点，捕获有助于特定功能的残基之间的相互作用5,12。事实上，对进化信息的分析已经允许鉴定蛋白质13,14，疾病变体15和蛋白质中的变构机制，这些蛋白质在不同的折叠之间经历可逆的转换，通常伴随着不同的功能16。迄今为止，UniProt数据库17（6/2023）中超过3.56亿个蛋白质已经测序，其中绝大多数（〜80%）没有已知的功能注释（例如，酶委员会编号和基因本体术语）。

Classical methods for annotating protein functions have been constrained by the extensive sizes of sequences, and the majority of function annotations are assigned at the protein level rather than the residue level18,19. As an alternative to these classical methods, computational approaches have b.

注释蛋白质功能的经典方法受到序列大小的限制，大多数功能注释是在蛋白质水平而不是残基水平上分配的18,19。作为这些经典方法的替代方法，计算方法具有b。

(1)

where H(k) and W(k) are the representation of residues and weights of the kth layer, respectively, and σ( ⋅ ) non-linear activation functions. In the present study, we implemented a normalized form over GCN and essentially arrive at the propagation rule55:$$f\left({{{{\bf{H}}}}}^{(k+1)},\, {{{\bf{A}}}}\right)=\sigma \left({\hat{{{{\bf{D}}}}}}^{-\frac{1}{2}}\hat{{{{\bf{A}}}}}{\hat{{{{\bf{D}}}}}}^{-\frac{1}{2}}{{{{\bf{H}}}}}^{(k)}{{{{\bf{W}}}}}^{(k)}\right),$$.

其中H（k）和W（k）分别表示第k层的残基和权重，以及σ（⋅）非线性激活函数。在本研究中，我们在GCN上实现了一个归一化形式，并基本上得出了传播规则55：$$f左（{{{{bf{H}}}}}^{（k+1）}，\，{{{{bf{a}}}\右）=\ sigma \左（{{bf{D}}}}}}^{-\ frac{1}{2}}\ hat{{{{{bf}}a}}}}}{\hat{{{{\bf{D}}}}}}}}{-\frac{1}{2}}{{{\bf{H}}}}}}^{（k）}{{{\bf{W}}}}}^{（k）}\右），$$。

(2)

with $\hat{{{{\bf{A}}}}}={{{\bf{A}}}}+{{{\bf{I}}}}$, where I is an identity matrix and $\hat{{{{\bf{D}}}}}$ is the diagonal node degree matrix of $\hat{{{{\bf{A}}}}}$.There are three blocks of GCN layer that are used in each channel of PhiGnet, and the number of hidden units in each GCN layer is set to 512.

其中I是一个单位矩阵，（{{{{bf{A}}}}}={{{bf{A}}+{{{{bf{I}}}}），其中I是一个单位矩阵，（{{{{bf{D}}}}）是“（{{{bf{A}}}}}”的对角节点度矩阵）。PhiGnet的每个通道中使用了三个GCN层块，每个GCN层中的隐藏单元数设置为512。

Information extracted by different channels, using either EVCs or RCs, can promote PhiGnet to learn features at two levels (Supplementary Figs. S9–S11). The outputs of the GCNs are concatenated into a tensor of dimensions L × D, where L represents the number of nodes in the graphs. To consolidate the information across the L dimension, we apply a SumPooling layer, reducing L to 1 while preserving the other dimension.

使用EVCs或RCs通过不同通道提取的信息可以促进PhiGnet在两个层面上学习特征（补充图S9-S11）。GCN的输出被连接成尺寸为L×D的张量，其中L表示图中的节点数。为了整合L维上的信息，我们应用了SumPooling层，将L减少到1，同时保留了另一个维度。

This aggregated tensor of size 1 × D is forwarded to the FC layers for predicting protein functions.Hyper-parameter tuning and PhiGnet trainingThe present PhiGnet allows us to directly learn information from a sequence alone (without using any structural knowledge) to significantly explore functional sites at the residue level.

这个大小为1×D的聚集张量被转发到FC层以预测蛋白质功能。超参数调整和PhiGnet训练目前的PhiGnet允许我们直接从序列中学习信息（不使用任何结构知识），以显着探索残基水平的功能位点。

To achieve an optimized model, we have to tune and choose values of the hyper-parameters in our method, e.g., thresholds for filtering EVCs/RCs (Supplementary Fig. S8). This tuning of parameters is crucial to guarantee both the stability and performance of PhiGnet.With the pre-defined hyper-parameters, we implemented a cross-entropy loss function to balance the abilities of learning and generalization.

为了实现优化的模型，我们必须在我们的方法中调整和选择超参数的值，例如过滤EVC/RC的阈值（补充图S8）。参数的这种调整对于保证PhiGnet的稳定性和性能至关重要。通过预定义的超参数，我们实现了一个交叉熵损失函数，以平衡学习和泛化能力。

The loss function is defined as follows,$${{{\mathcal{L}}}}=-\frac{1}{N}{\sum }_{i=1}^{N}\mathop{\sum }_{j=1}^{F}\left[{y}_{ij}\log ({\hat{y}}_{ij})+(1-{y}_{ij})\log (1-{\hat{y}}_{ij})\right],$$.

损失函数的定义如下：$${{{\mathcal{L}}}=-\frac{1}{N}{\sum}\ui=1}^{N}\mathop{\sum}\uj=1}^{F}\left[{y}_-{y}_{ij}）\log（1-{\hat{y}}\uij}）\right]，$$。

(3)

where N is the number of data samples, and F is the number of function classes in EC numbers/GO terms. yij is to label the ground truth to 1 if the ith sample is in the jth function class, otherwise, it is 0. Similarly, ${\hat{y}}_{ij}$ is a label for the prediction.PhiGnet was trained with batch size of 64 for maximum 500 epochs using early-stopping criterion over the defined cross-entropy loss (Eq.

其中N是数据样本的数量，F是EC数字/GO术语中的函数类数量。如果第i个样本在第j个函数类中，则yij将基本真值标记为1，否则为0。类似地，\（{\ hat{y}}uij}）是预测的标签。在定义的交叉熵损失（Eq.）上，使用提前停止标准，对PhiGnet进行了最大500个时期的批处理大小为64的训练。

(3)). During training, we leveraged the Adam optimizer56 with a learning rate of 2 × 10−4, β1 = 0.9, β2 = 0.999, ϵ = 1 × 10−6, and L2 weight decay of 2 × 10−5. To avoid over-fitting, we implemented a dropout of 0.3 for the second fully connected layer. Accordingly, we achieved fine-trained models of PhiGnet that are leveraged to predict the probability of assigning EC numbers/GO terms to a given protein by learning from sequence embedding under constraints of evolutionary couplings and couplings intra residue communities.Function annotations at the residue levelTo quantitatively evaluate the importance of residues, we implemented the gradient-weighted class activation map method (that localizes the most important regions in images relevant for making correct classification decisions in computer vision)32 for a specific function annotation to compute scores that are assigned to each residue in a protein.

（3）））。在训练期间，我们利用Adam优化器56，学习率为2×10-4，β1×0.9，β2×0.999，β1×10-6，L2权重衰减为2×10-5。为了避免过度拟合，我们对第二个完全连接的层实施了0.3的辍学。因此，我们实现了经过精细训练的PhiGnet模型，该模型可用于通过在进化偶联和残基内群落偶联的约束下从序列嵌入中学习，预测将EC数/GO项分配给给定蛋白质的概率。。

In the grad-CAM method, the gradient information of a given layer is used to compute localization map ${{{{\bf{M}}}}}^{c}\in {{\mathbb{R}}}^{u\times v}$ with width u and height v, and it is used to characterize the importance of every single element of the input for a specific class c. Given a feature map Fk, the activation value ${{{{\mathcal{S}}}}}^{c}$ for scoring the class c is computed to measure the importance of neurons, ${\alpha }_{k}^{c}$, as.

在grad-CAM方法中，给定层的梯度信息用于计算宽度为u和高度为v的定位图\（{{{{{bf{M}}}}^{c}在{{mathbb{R}}}^{u \ times v}\），并用于表征特定类c输入的每个元素的重要性。给定特征图Fk，激活值\（{{{{mathcal{S}}}}}^{c}计算用于评分c类的\），以测量神经元的重要性\（{\ alpha}{k}^{c}\），如。

(4)

$${\alpha }_{k}^{c}=\frac{1}{L}{\sum }_{i}^{L}\frac{\partial {Y}^{c}}{\partial {{{{\bf{F}}}}}_{i}^{k}},$$

$${\alpha}}{k}}{c}=\frac{1}{L}{\sum}}u{i}^{L}\frac{\partial{Y}^{c}}{\partial{{{{{\bf}}}}}}\u{i}^{k}}$$

(5)

where ReLU( ⋅ ) is a non-linear activation function, holding a positive effect for function class c, and L is the number of elements in the input.In the present method, we evaluated the importance of the ith amino acid in the feature map Fk obtained from the layer concatenated from the two channels in PhiGnet, and the gradient $\frac{\partial {{{{\bf{Y}}}}}^{c}}{\partial {{{{\bf{F}}}}}_{i}^{k}}$ is calculated by the derivative of the function annotation c with predicted score Yc, with respect to the feature map ${{{{\bf{F}}}}}_{i}^{k}$ in sequence of length L.Comparison with existing approachesIn the present study, we compared our method to eight methods, including BLAST18, FunFams40, DeepGO25, DeepFRI21, ProteInfer43 ATGO45, SPROF-GO44, and CLEAN46 in details.

其中ReLU（⋅）是一个非线性激活函数，对函数类c有积极影响，L是输入中的元素数。在本方法中，我们评估了从PhiGnet中两个通道连接的层获得的特征图Fk中第i个氨基酸的重要性，并且梯度\（\ frac{\ partial{{{{{\ bf{Y}}}}^{c}}{\ partial{{{\ bf}}}}{i}^{k}）是通过函数注释c相对于特征图\（{{{{{\ bf{F}}}}}{i}^{k}}的长度顺序。与现有方法相比，在本研究中，我们将我们的方法与八种方法进行了比较，包括BLAST18，FunFams40，DeepGO25，DeepFRI21，ProteInfer43 ATGO45，SPROF-GO44和CLEAN46。

Moreover, our method was compared to predictions collected from two web-servers, DeepGOWeb42 and Pannzer41, over predictions of either GO terms in different ontologies or EC numbers using the collected data sets.BLAST is a sequence searching tool based on the local sequence alignment algorithm18. Implementing BLAST, we transferred function annotations to proteins within the test set from all the annotated sequences in the training dataset following the same procedure as presented in refs.

此外，我们的方法与从两个web服务器DeepGOWeb42和Pannzer41收集的预测进行了比较，使用收集的数据集对不同本体中的GO术语或EC数字进行了预测。BLAST是一种基于局部序列比对算法的序列搜索工具18。实现BLAST，我们按照与参考文献中相同的程序，从训练数据集中的所有注释序列中将功能注释转移到测试集中的蛋白质。

20,21. The probability assigning annotation(s) to each protein was computed by sequence identity in percentage between the sequences in the test and training sets. More specifically, if a protein in the test set hits against proteins in the training set with a maximum sequence identity of 75%, it was assigned function annotation(s) by transferring all the annotations from training proteins with a score of 0.75.

20,21。通过测试集和训练集中序列之间的百分比序列同一性来计算为每种蛋白质分配注释的概率。更具体地说，如果测试集中的蛋白质与训练集中的蛋白质发生碰撞，最大序列同一性为75%，则通过从得分为0.75的训练蛋白质中转移所有注释来分配功能注释。

In practice, we filtered out sequences from the training set using default parameters to keep withi.

实际上，我们使用默认参数从训练集中过滤出序列以保持不变。

(6)

$${{{\rm{AUPR}}}}=\int_{0}^{1}p(t)\times r(t)\,dt,$$

$${{{\rm{AUPR}}}=\int\u0}^{1}p（t）乘以r（t）\，dt$$

(7)

where p and r are precision that measures the predictive accuracy and recall that is to measure successfully retrieved information, respectively.Statistics and reproducibilityNo statistical method was used to predetermine sample size.Reporting summaryFurther information on research design is available in the Nature Portfolio Reporting Summary linked to this article..

其中p和r是精度，分别测量预测准确性和召回率，以测量成功检索的信息。统计和可重复性没有使用统计方法来预先确定样本量。报告摘要有关研究设计的更多信息，请参阅本文链接的Nature Portfolio Reporting Summary。。

Data availability

数据可用性

All relevant data supporting the key findings of this study are available within the article and its Supplementary Information files. All crystal structures of proteins used in this study are available at Protein Data Bank (https://www.rcsb.org) under accession codes: 4JDZ [https://doi.org/10.2210/pdb4JDZ/pdb], 6IZW [https://doi.org/10.2210/pdb6IZW/pdb], 6IEJ [https://doi.org/10.2210/pdb6IEJ/pdb], 6W8I [https://doi.org/10.2210/pdb6W8I/pdb], 6XK2 [https://doi.org/10.2210/pdb6XK2/pdb], 1HFX [https://doi.org/10.2210/pdb1HFX/pdb], 1MNM [https://doi.org/10.2210/pdb1MNM/pdb], 1FOS [https://doi.org/10.2210/pdb1FOS/pdb], 3TMK [https://doi.org/10.2210/pdb3TMK/pdb], 2GB7 [https://doi.org/10.2210/pdb2GB7/pdb], 4A7W [https://doi.org/10.2210/pdb4A7W/pdb], 1MQ0 [https://doi.org/10.2210/pdb1MQ0/pdb], 2FE3 [https://doi.org/10.2210/pdb2FE3/pdb], 7QXO [https://doi.org/10.2210/pdb7QXO/pdb], and 8E0A [https://doi.org/10.2210/pdb8E0A/pdb].

本文及其补充信息文件中提供了支持本研究主要发现的所有相关数据。本研究中使用的所有蛋白质晶体结构均可在蛋白质数据库中找到(https://www.rcsb.org)根据加入代码：4JDZ[https://doi.org/10.2210/pdb4JDZ/pdb]，6IZW[https://doi.org/10.2210/pdb6IZW/pdb]，6IEJ[https://doi.org/10.2210/pdb6IEJ/pdb]，6W8I[https://doi.org/10.2210/pdb6W8I/pdb]，6XK2[https://doi.org/10.2210/pdb6XK2/pdb]，1HFX[https://doi.org/10.2210/pdb1HFX/pdb]，1毫米[https://doi.org/10.2210/pdb1MNM/pdb]，1个[https://doi.org/10.2210/pdb1FOS/pdb]，3TMK[https://doi.org/10.2210/pdb3TMK/pdb]，2GB7[https://doi.org/10.2210/pdb2GB7/pdb]，4A7W[https://doi.org/10.2210/pdb4A7W/pdb]，1MQ0[https://doi.org/10.2210/pdb1MQ0/pdb]，2FE3[https://doi.org/10.2210/pdb2FE3/pdb]，7QXO[https://doi.org/10.2210/pdb7QXO/pdb]和8E0A[https://doi.org/10.2210/pdb8E0A/pdb]。

The data is available for download at https://doi.org/10.5281/zenodo.12496869. Source data are provided with this paper..

数据可在以下网址下载：https://doi.org/10.5281/zenodo.12496869.本文提供了源数据。。

Code availability

代码可用性

The PhiGnet Python code and pre-trained model are available at: https://doi.org/10.5281/zenodo.12496869.

PhiGnet Python代码和预训练模型可在以下网址获得：https://doi.org/10.5281/zenodo.12496869.

ReferencesRepecka, D. et al. Expanding functional protein sequence spaces using generative adversarial networks. Nat. Mach. Intell. 3, 324–333 (2021).Article

参考文献Repecka，D。等人。使用生成对抗网络扩展功能性蛋白质序列空间。自然马赫数。因特尔。3324-333（2021）。文章

Google Scholar

谷歌学者

Ferruz, N. et al. From sequence to function through structure: deep learning for protein design. Comput. Struct. Biotechnol. J. 21, 238–250 (2022).Boike, L., Henning, N. J. & Nomura, D. K. Advances in covalent drug discovery. Nat. Rev. Drug Discov. 21, 881–898 (2022).Article

Ferruz，N.等人，《从序列到结构的功能：蛋白质设计的深度学习》。计算机。结构。生物技术。J、 21238-250（2022）。Boike，L.，Henning，N.J。和Nomura，D.K。共价药物发现的进展。《药物目录》修订版。。文章

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Anfinsen, C. B. The formation and stabilization of protein structure. Biochem. J. 128, 737 (1972).Article

Anfinsen，C.B。蛋白质结构的形成和稳定。生物化学。J、 128737（1972）。文章

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Socolich, M. et al. Evolutionary information for specifying a protein fold. Nature 437, 512–518 (2005).Article

Socolich，M.等人。用于指定蛋白质折叠的进化信息。自然437512-518（2005）。文章

ADS

PubMed

Google Scholar

谷歌学者

Hopf, T. A. et al. Three-dimensional structures of membrane proteins from genomic sequencing. Cell 149, 1607–1621 (2012).Article

Hopf，T.A.等人。来自基因组测序的膜蛋白的三维结构。细胞1491607-1621（2012）。文章

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Kuhlman, B. & Bradley, P. Advances in protein structure prediction and design. Nat. Rev. Mol. Cell Biol. 20, 681–697 (2019).Article

Kuhlman，B。＆Bradley，P。蛋白质结构预测和设计的进展。Nat。Rev。Mol。Cell Biol。20681-697（2019）。文章

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021).Article

Jumper，J.等人。使用AlphaFold进行高度准确的蛋白质结构预测。自然596583-589（2021）。文章

ADS

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Baek, M. et al. Accurate prediction of protein structures and interactions using a three-track neural network. Science 373, 871–876 (2021).Article

Baek，M.等人，《使用三轨神经网络精确预测蛋白质结构和相互作用》，《科学》373871–876（2021）。文章

ADS

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Humphreys, I. R. et al. Computed structures of core eukaryotic protein complexes. Science 374, eabm4805 (2021).Article

Humphreys，I.R.等人计算了核心真核蛋白复合物的结构。科学374，eabm4805（2021）。文章

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Wang, J. et al. Scaffolding protein functional sites using deep learning. Science 377, 387–394 (2022).Article

Wang，J。等人。使用深度学习构建蛋白质功能位点。科学377387-394（2022）。文章

ADS

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Cheung, N. J., Peter, A. T. J. & Kornmann, B. Leri: a web-server for identifying protein functional networks from evolutionary couplings. Comput. Struct. Biotechnol. J. 19, 3556–3563 (2021).Article

Cheung，N.J.，Peter，A.T.J.＆Kornmann，B.Leri：用于从进化耦合中识别蛋白质功能网络的网络服务器。计算机。结构。生物技术。J、 193556-3563（2021）。文章

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Changeux, J.-P. & Edelstein, S. J. Allosteric mechanisms of signal transduction. Science 308, 1424–1428 (2005).Article

Changeux，J.-P。＆Edelstein，S.J。信号转导的变构机制。科学3081424-1428（2005）。文章

ADS

PubMed

Google Scholar

谷歌学者

Faure, A. J. et al. Mapping the energetic and allosteric landscapes of protein binding domains. Nature 604, 175–183 (2022).Article

Faure，A.J.等人绘制了蛋白质结合域的能量和变构景观。自然604175-183（2022）。文章

ADS

PubMed

Google Scholar

谷歌学者

Frazer, J. et al. Disease variant prediction with deep generative models of evolutionary data. Nature 599, 91–95 (2021).Article

Frazer，J.等人。利用进化数据的深度生成模型进行疾病变异预测。。文章

ADS

PubMed

Google Scholar

谷歌学者

Dishman, A. F. et al. Evolution of fold switching in a metamorphic protein. Science 371, 86–90 (2021).Article

Dishman，A.F.等人。变性蛋白质中折叠转换的进化。科学371,86-90（2021）。文章

ADS

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Bateman, A. et al. UniProt: the universal protein knowledgebase in 2023. Nucleic Acids Res. 51, D523–D531 (2023).Altschul, S. F. et al. Gapped BLAST and PSI-BLAST: a new generation of protein database search programs. Nucleic Acids Res. 25, 3389–3402 (1997).Article

Bateman，A。等人。UniProt：2023年的通用蛋白质知识库。核酸研究51，D523–D531（2023）。。核酸研究253389-3402（1997）。文章

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Scheibenreif, L., Littmann, M., Orengo, C. & Rost, B. FunFam protein families improve residue level molecular function prediction. BMC Bioinforma. 20, 1–9 (2019).Article

Scheibenreif，L.，Littmann，M.，Orengo，C。＆Rost，B。FunFam蛋白家族改善了残基水平的分子功能预测。BMC生物信息学。20,1-9（2019）。文章

Google Scholar

谷歌学者

Radivojac, P. et al. A large-scale evaluation of computational protein function prediction. Nat. Methods 10, 221–227 (2013).Article

Radivojac，P。等人。计算蛋白质功能预测的大规模评估。自然方法10221-227（2013）。文章

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Gligorijević, V. et al. Structure-based protein function prediction using graph convolutional networks. Nat. Commun. 12, 1–14 (2021).Article

Gligorijević，V。等人。使用图卷积网络进行基于结构的蛋白质功能预测。国家公社。12，1-14（2021）。文章

Google Scholar

谷歌学者

Gelman, S., Fahlberg, S. A., Heinzelman, P., Romero, P. A. & Gitter, A. Neural networks to learn protein sequence–function relationships from deep mutational scanning data. Proc. Natl. Acad. Sci. USA 118, e2104878118 (2021).Article

Gelman，S.，Fahlberg，S.A.，Heinzelman，P.，Romero，P.A。＆Gitter，A。神经网络从深度突变扫描数据中学习蛋白质序列-功能关系。程序。纳特尔。阿卡德。。美国118，e2104878118（2021）。文章

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Bileschi, M. L. et al. Using deep learning to annotate the protein universe. Nat. Biotechnol. 40, 932–937 (2022).Article

Bileschi，M.L.等人使用深度学习来注释蛋白质宇宙。美国国家生物技术公司。40932-937（2022）。文章

PubMed

Google Scholar

谷歌学者

Unsal, S. et al. Learning functional properties of proteins with language models. Nat. Mach. Intell. 4, 227–245 (2022).Article

Unsal，S.等人。用语言模型学习蛋白质的功能特性。自然马赫数。因特尔。4227-245（2022）。文章

Google Scholar

谷歌学者

Kulmanov, M., Khan, M. A. & Hoehndorf, R. DeepGO: predicting protein functions from sequence and interactions using a deep ontology-aware classifier. Bioinformatics 34, 660–668 (2018).Article

Kulmanov，M.，Khan，M.A。＆Hoehndorf，R.DeepGO：使用深度本体感知分类器从序列和相互作用预测蛋白质功能。生物信息学34660-668（2018）。文章

PubMed

Google Scholar

谷歌学者

Karniadakis, G. E. et al. Physics-informed machine learning. Nat. Rev. Phys. 3, 422–440 (2021).Article

Karniadakis，G.E.等人，《物理告知机器学习》。自然物理评论。3422-440（2021）。文章

Google Scholar

谷歌学者

Pazos, F. & Sternberg, M. J. Automated prediction of protein function and detection of functional sites from structure. Proc. Natl. Acad. Sci. USA 101, 14754–14759 (2004).Article

Pazos，F。＆Sternberg，M.J。蛋白质功能的自动预测和结构功能位点的检测。程序。纳特尔。阿卡德。。美国10114754-14759（2004）。文章

ADS

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Gherardini, P. F. & Helmer-Citterich, M. Structure-based function prediction: approaches and applications. Brief. Funct. Genom. Proteom. 7, 291–302 (2008).Article

Gherardini，P.F。和Helmer-Citterich，M。基于结构的功能预测：方法和应用。简介。函数。基因组。蛋白质组学。7291-302（2008）。文章

Google Scholar

谷歌学者

Glazer, D. S., Radmer, R. J. & Altman, R. B. Improving structure-based function prediction using molecular dynamics. Structure 17, 919–929 (2009).Article

Glazer，D.S.，Radmer，R.J。＆Altman，R.B。使用分子动力学改进基于结构的功能预测。结构17919-929（2009）。文章

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Skolnick, J. & Brylinski, M. FINDSITE: a combined evolution/structure-based approach to protein function prediction. Brief. Bioinforma. 10, 378–391 (2009).Article

Skolnick，J。＆Brylinski，M。FINDSITE：一种基于进化/结构的蛋白质功能预测方法。简介。生物信息学。10378-391（2009）。文章

Google Scholar

谷歌学者

Rives, A. et al. Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences. Proc. Natl. Acad. Sci. USA 118, e2016239118 (2021).Article

生物结构和功能从无监督学习扩展到2.5亿个蛋白质序列。程序。纳特尔。阿卡德。。美国118，e2016239118（2021）。文章

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Selvaraju, R. R. et al. Grad-CAM: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE International Conference on Computer Vision, 618–626 (2017).Askarian, F. et al. Serine-aspartate repeat protein D increases Staphylococcus aureus virulence and survival in blood.

Selvaraju，R.R.等人，Grad CAM：通过基于梯度的定位从深度网络进行视觉解释。在IEEE国际计算机视觉会议论文集，618-626（2017）。Askarian，F。等人。丝氨酸天冬氨酸重复蛋白D增加金黄色葡萄球菌的毒力和血液中的存活率。

Infect. Immun. 85, e00559–16 (2017).Article .

感染。免疫。85，e00559–16（2017）。文章。

PubMed

Google Scholar

谷歌学者

Milles, L. F., Unterauer, E. M., Nicolaus, T. & Gaub, H. E. Calcium stabilizes the strongest protein fold. Nat. Commun. 9, 1–10 (2018).Article

Milles，L.F.，Unterauer，E.M.，Nicolaus，T。＆Gaub，H.E。钙稳定最强的蛋白质折叠。国家公社。9,1-10（2018）。文章

ADS

Google Scholar

谷歌学者

Yang, J., Roy, A. & Zhang, Y. BioLiP: a semi-manually curated database for biologically relevant ligand–protein interactions. Nucleic Acids Res. 41, D1096–D1103 (2012).Article

Yang，J.，Roy，A。＆Zhang，Y。BioLiP：用于生物学相关配体-蛋白质相互作用的半手动管理数据库。核酸研究41，D1096–D1103（2012）。文章

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Baranwal, J. et al. Allosteric regulation of a prokaryotic small Ras-like GTPase contributes to cell polarity oscillations in bacterial motility. PLoS Biol. 17, e3000459 (2019).Article

Baranwal，J。等人。原核小Ras样GTPase的变构调节有助于细菌运动中的细胞极性振荡。《公共科学图书馆·生物学》。17，e3000459（2019）。文章

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Hirano, Y. et al. Structural basis of phosphatidylcholine recognition by the C2–domain of cytosolic phospholipase A2α. Elife 8, e44760 (2019).Article

Hirano，Y。等人。胞质磷脂酶A2αC2结构域识别磷脂酰胆碱的结构基础。Elife 8，e44760（2019）。文章

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Pike, A. C., Brew, K. & Acharya, K. R. Crystal structures of guinea-pig, goat and bovine α-lactalbumin highlight the enhanced conformational flexibility of regions that are significant for its action in lactose synthase. Structure 4, 691–703 (1996).Article

Pike，A.C.，Brew，K。＆Acharya，K.R。豚鼠，山羊和牛α-乳清蛋白的晶体结构突出了对其在乳糖合酶中的作用重要的区域的增强的构象柔性。结构4691-703（1996）。文章

PubMed

Google Scholar

谷歌学者

Bochtler, M. et al. Nucleotide flips determine the specificity of the Ecl18kI restriction endonuclease. EMBO J. 25, 2219–2229 (2006).Article

Bochtler，M。等人的核苷酸翻转决定了Ecl18kI限制性核酸内切酶的特异性。EMBO J.252219–2229（2006）。文章

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Das, S. Functional classification of CATH superfamilies: a domain-based approach for protein function annotation. Bioinformatics 31, 3460–3467 (2015).Article

Das，S。CATH超家族的功能分类：基于域的蛋白质功能注释方法。生物信息学313460-3467（2015）。文章

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Törönen, P. & Holm, L. PANNZER—a practical tool for protein function prediction. Protein Sci. 31, 118–128 (2022).Article

Törönen，P。＆Holm，L。PANNZER-蛋白质功能预测的实用工具。蛋白质科学。31118-128（2022）。文章

PubMed

Google Scholar

谷歌学者

Kulmanov, M., Zhapa-Camacho, F. & Hoehndorf, R. DeepGOWeb: fast and accurate protein function prediction on the (Semantic) Web. Nucleic Acids Res. 49, W140–W146 (2021).Article

Kulmanov，M.，Zhapa Camacho，F。＆Hoehndorf，R。DeepGOWeb：在（语义）网络上快速准确地预测蛋白质功能。核酸研究49，W140–W146（2021）。文章

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Sanderson, T., Bileschi, M. L., Belanger, D. & Colwell, L. J. ProteInfer, deep neural networks for protein functional inference. Elife 12, e80942 (2023).Article

Sanderson，T.，Bileschi，M.L.，Belanger，D。＆Colwell，L.J。ProteInfer，用于蛋白质功能推断的深度神经网络。Elife 12，e80942（2023）。文章

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Yuan, Q., Xie, J., Xie, J., Zhao, H. & Yang, Y. Fast and accurate protein function prediction from sequence through pretrained language model and homology-based label diffusion. Brief. Bioinforma. 24, bbad117 (2023).Article

。简介。生物信息学。24，bbad117（2023）。文章

Google Scholar

谷歌学者

Zhu, Y.-H., Zhang, C., Yu, D.-J. & Zhang, Y. Integrating unsupervised language model with triplet neural networks for protein gene ontology prediction. PLOS Comput. Biol. 18, e1010793 (2022).Article

Zhu，Y.-H.，Zhang，C.，Yu，D.-J.＆Zhang，Y.将无监督语言模型与三重神经网络相结合，用于蛋白质基因本体预测。PLOS计算机。生物学18，e1010793（2022）。文章

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Yu, T. et al. Enzyme function prediction using contrastive learning. Science 379, 1358–1363 (2023).Article

Yu，T。等人。使用对比学习预测酶功能。科学3791358-1363（2023）。文章

ADS

PubMed

Google Scholar

谷歌学者

Chung, S. J., Fromme, J. C. & Verdine, G. L. Structure of human cytidine deaminase bound to a potent inhibitor. J. Med. Chem. 48, 658–660 (2005).Article

Chung，S.J.，Fromme，J.C。和Verdine，G.L。人胞苷脱氨酶的结构与有效抑制剂结合。J、医学化学。48658-660（2005）。文章

PubMed

Google Scholar

谷歌学者

Traoré, D. A. et al. Crystal structure of the apo-PerR-Zn protein from Bacillus subtilis. Mol. Microbiol. 61, 1211–1219 (2006).Article

。分子微生物。611211-1219（2006）。文章

PubMed

Google Scholar

谷歌学者

Zhou, N. et al. The CAFA challenge reports improved protein function prediction and new functional annotations for hundreds of genes through experimental screens. Genome Biol. 20, 1–23 (2019).Article

周，N。等人。CAFA挑战报告通过实验筛选改进了蛋白质功能预测和数百个基因的新功能注释。基因组生物学。20,1-23（2019）。文章

Google Scholar

谷歌学者

Berman, H. M. et al. The protein data bank. Nucleic Acids Res. 28, 235–242 (2000).Article

Berman，H.M.等人，《蛋白质数据库》。核酸研究28235-242（2000）。文章

ADS

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Ng, P. C. & Henikoff, S. Predicting deleterious amino acid substitutions. Genome Res. 11, 863–874 (2001).Article

Ng，P.C。＆Henikoff，S。预测有害的氨基酸取代。Genome Res.11863–874（2001）。文章

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Mirdita, M. et al. Uniclust databases of clustered and deeply annotated protein sequences and alignments. Nucleic Acids Res. 45, D170–D176 (2017).Article

Mirdita，M。等人。聚类和深度注释的蛋白质序列和比对的Uniclust数据库。核酸研究45，D170–D176（2017）。文章

PubMed

Google Scholar

谷歌学者

Remmert, M., Biegert, A., Hauser, A. & Söding, J. HHblits: lightning-fast iterative protein sequence searching by HMM-HMM alignment. Nat. Methods 9, 173–175 (2012).Article

Remmert，M.，Biegert，A.，Hauser，A。＆Söding，J。HHblits：通过HMM-HMM比对进行闪电般快速的迭代蛋白质序列搜索。《自然方法》9173-175（2012）。文章

Google Scholar

谷歌学者

Anishchenko, I., Ovchinnikov, S., Kamisetty, H. & Baker, D. Origins of coevolution between residues distant in protein 3D structures. Proc. Natl. Acad. Sci. USA 114, 9122–9127 (2017).Article

Anishchenko，I.，Ovchinnikov，S.，Kamisetty，H。＆Baker，D。蛋白质3D结构中遥远残基之间共同进化的起源。程序。纳特尔。阿卡德。。美国1149122–9127（2017）。文章

ADS

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Kipf, T. N. & Welling, M. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016).Kingma, D. P. & Ba, J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014).Eddy, S. R. Accelerated profile HMM searches. PLoS Comput.

Kipf，T。N。＆Welling，M。使用图卷积网络的半监督分类。arXiv预印本arXiv:1609.02907（2016）。Kingma，D.P.＆Ba，J.Adam：一种随机优化方法。。Eddy，S.R。加速配置文件HMM搜索。PLoS计算机。

Biol. 7, e1002195 (2011).Article .

生物学7，e1002195（2011）。第[UNK]条。

ADS

MathSciNet

PubMed

PubMed Central

公共医学中心

Google Scholar

谷歌学者

Kulmanov, M. & Hoehndorf, R. DeepGOPlus: improved protein function prediction from sequence. Bioinformatics 36, 422–429 (2020).Article

Kulmanov，M。＆Hoehndorf，R。DeepGOPlus：改进的序列蛋白质功能预测。生物信息学36422-429（2020）。文章

PubMed

Google Scholar

谷歌学者

Buchfink, B., Xie, C. & Huson, D. H. Fast and sensitive protein alignment using DIAMOND. Nat. Methods 12, 59–60 (2015).Article

Buchfink，B.，Xie，C。＆Huson，D.H。使用DIAMOND进行快速灵敏的蛋白质比对。自然方法12,59-60（2015）。文章

PubMed

Google Scholar

谷歌学者

Tan, S. & Richmond, T. J. Crystal structure of the yeast MATα2/MCM1/DNA ternary complex. Nature 391, 660–666 (1998).Article

Tan，S。＆Richmond，T。J。酵母MATα2/MCM1/DNA三元复合物的晶体结构。自然391660-666（1998）。文章

ADS

PubMed

Google Scholar

谷歌学者

Glover, J. & Harrison, S. C. Crystal structure of the heterodimeric bZIP transcription factor c-Fos–c-Jun bound to DNA. Nature 373, 257–261 (1995).Article

Glover，J。＆Harrison，S.C。与DNA结合的异二聚体bZIP转录因子C-Fos–C-Jun的晶体结构。自然373257-261（1995）。文章

ADS

PubMed

Google Scholar

谷歌学者

Lavie, A. et al. Crystal structure of yeast thymidylate kinase complexed with the bisubstrate inhibitor P 1-(5 ‘-Adenosyl) P 5-(5 ‘-Thymidyl) pentaphosphate (TP5A) at 2.0 Å resolution: Implications for catalysis and AZT activation. Biochemistry 37, 3677–3686 (1998).Article

Lavie，A.等人。酵母胸苷酸激酶与双底物抑制剂P 1-（5'-腺苷基）P 5-（5'-胸苷基）五磷酸（TP5A）以2.0Å分辨率复合的晶体结构：对催化和AZT活化的影响。生物化学373677-3686（1998）。文章

PubMed

Google Scholar

谷歌学者

Chu, C.-H. et al. Structures of helicobacter pylori uridylate kinase: insight into release of the product UDP. Acta Crystallogr. D Biol. Crystallogr. 68, 773–783 (2012).Article

Chu，C.-H.等人。幽门螺杆菌尿苷酸激酶的结构：深入了解产物UDP的释放。晶体学报。D生物晶体学。68773-783（2012）。文章

ADS

PubMed

Google Scholar

谷歌学者

Download referencesAcknowledgementsThis work was funded by Wellcome Trust (214291/Z/18/Z, to B.K.). We thank members of the Kornmann laboratory and AmoAi for many valuable discussions. Y.J.J. and Q.Q.Q. are supported by AmoAi.Author informationAuthor notesThese authors contributed equally: Yaan J.

下载参考文献致谢这项工作由Wellcome Trust（214291/Z/18/Z，B.K.）资助。我们感谢Kornmann实验室和AmoAi的成员进行了许多有价值的讨论。Y、 J.J.和Q.Q.Q.得到了AmoAi的支持。。

Jang, Qi-Qi Qin.Authors and AffiliationsDepartment of Biochemistry, University of Oxford, Oxford, UKYaan J. Jang & Benoît KornmannAmoAi Technologies, Oxford, UKYaan J. Jang, Qi-Qi Qin & Si-Yu HuangSchool of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai, ChinaQi-Qi Qin & Xue-Ming DingOxford Martin School, University of Oxford, Oxford, UKSi-Yu HuangSchool of Systems Science, Beijing Normal University, Beijing, ChinaSi-Yu HuangInstitute of Biochemistry, ETH Zürich, Zürich, SwitzerlandArun T.

张，齐齐秦。作者和附属机构牛津大学生物化学系，牛津，英国J.Jang＆Benoît KornmannAmoAi Technologies，牛津，英国J.Jang，Qi Qi Qin＆Si Yu Huang上海科技大学光电与计算机工程学院，上海，中国Qi Qin＆Xue Ming Dingford Martin School，牛津大学，UKSi Yu Huang School of Systems Science，北京师范大学，中国Si Yu Huang生物化学研究所，ETH Zürich，Zürich，SwitzerlandArun t。

John PeterAuthorsYaan J. JangView author publicationsYou can also search for this author in.

John PeterAuthorsYaan J.JangView作者出版物您也可以在中搜索这位作者。

PubMed Google ScholarQi-Qi QinView author publicationsYou can also search for this author in

PubMed谷歌学者Qi Qin查看作者出版物您也可以在

PubMed Google ScholarSi-Yu HuangView author publicationsYou can also search for this author in

PubMed Google ScholarSi Yu Huang查看作者出版物您也可以在

PubMed Google ScholarArun T. John PeterView author publicationsYou can also search for this author in

PubMed Google ScholarArun T.John PeterView作者出版物您也可以在

PubMed Google ScholarXue-Ming DingView author publicationsYou can also search for this author in

PubMed Google ScholarXue Ming DingView作者出版物您也可以在

PubMed Google ScholarBenoît KornmannView author publicationsYou can also search for this author in

PubMed Google ScholarBenoît KornmannView作者出版物您也可以在

PubMed Google ScholarContributionsY.J.J. led the research, conceived the end-to-end approach, designed experiments, financed the experiments, and wrote the manuscript. Q.Q.Q. collected the data, implemented the method, contributed with principal analysis and wrote the first draft.

PubMed谷歌学术贡献。J、 J.领导了这项研究，构思了端到端的方法，设计了实验，资助了实验，并撰写了手稿。Q、。

S.Y.H. conducted principal analysis over predictions. X.M.D. conducted data analysis. A.T.J.P. supported with principal analysis and wrote the manuscript. B.K. led the research, funding acquisition, contributed technical advice, and wrote the manuscript. All authors read the final manuscript.Corresponding authorsCorrespondence to.

S、 Y.H.对预测进行了主要分析。十、医学博士进行了数据分析。A、 T.J.P.支持主要分析并撰写了手稿。B、 K.领导了研究，资助收购，提供了技术建议，并撰写了手稿。所有作者都阅读了最终稿件。通讯作者通讯。

Yaan J. Jang or Benoît Kornmann.Ethics declarations

Yaan J.Jang或Benoît Kornmann。道德宣言

Competing interests

相互竞争的利益

Y.J.J. is a founder of AmoAi Technologies, UK. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interests.

Y、 J.J.是英国AmoAi Technologies的创始人。其余作者声明，这项研究是在没有任何可能被解释为潜在利益冲突的商业或财务关系的情况下进行的。

Peer review

同行评审

Peer review information

同行评审信息

Nature Communications thanks Guoxian Yu, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

《自然通讯》感谢余国贤和另一位匿名审稿人对这项工作的同行评审所做的贡献。可以获得同行评审文件。

Additional informationPublisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.Supplementary informationSupplementary InformationPeer Review FileReporting SummarySource dataSource DataRights and permissions

Additional informationPublisher的注释Springer Nature在已发布的地图和机构隶属关系中的管辖权主张方面保持中立。补充信息补充信息同行评审文件报告摘要源数据源数据权限

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material.

开放获取本文是根据知识共享署名非商业性NoDerivatives 4.0国际许可证授权的，该许可证允许以任何媒介或格式进行任何非商业性使用，共享，分发和复制，只要您对原始作者和来源给予适当的信任，提供知识共享许可证的链接，并指出您是否修改了许可材料。

You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.

To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/..

要查看此许可证的副本，请访问http://creativecommons.org/licenses/by-nc-nd/4.0/..

Reprints and permissionsAbout this articleCite this articleJang, Y.J., Qin, QQ., Huang, SY. et al. Accurate prediction of protein function using statistics-informed graph networks.

转载和许可本文引用本文Jang，Y.J.，Qin，QQ。，Huang，SY。等人。使用统计信息图网络准确预测蛋白质功能。

Nat Commun 15, 6601 (2024). https://doi.org/10.1038/s41467-024-50955-0Download citationReceived: 17 May 2023Accepted: 15 July 2024Published: 04 August 2024DOI: https://doi.org/10.1038/s41467-024-50955-0Share this articleAnyone you share the following link with will be able to read this content:Get shareable linkSorry, a shareable link is not currently available for this article.Copy to clipboard.

《国家公社》156601（2024）。https://doi.org/10.1038/s41467-024-50955-0Downloadhttps://doi.org/10.1038/s41467-024-50955-0Share本文与您共享以下链接的任何人都可以阅读此内容：获取可共享链接对不起，本文目前没有可共享的链接。复制到剪贴板。

Provided by the Springer Nature SharedIt content-sharing initiative

由Springer Nature SharedIt内容共享计划提供

CommentsBy submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.

评论通过提交评论，您同意遵守我们的条款和社区指南。如果您发现有虐待行为或不符合我们的条款或准则，请将其标记为不合适。

全球产业链接平台

重庆市渝北区金星科技大厦A区5楼512室

联系电话：023-67139735（重庆）

关于我们

产品服务