图1 理性设计，定向进化和人工智能辅助的蛋白质工程策略示意图（理性设计依赖序列和结构信息，精准设计突变体文库，但难以应用于缺少结构功能信息的蛋白质。定向进化中对目标基因进行多轮突变和筛选实验，不受结构功能信息限制，但是需要进行高通量的筛选方法。人工智能辅助的蛋白质工程则需要大量的序列-功能数据，可以来源于实验、计算和数据库等多方面，通过构建的预测模型，能够更有效地探索蛋白质突变体序列空间）

Fig. 1 Schematic diagram for rational design, directed evolution and artificial intelligence-assisted protein engineering(Rational design relies on sequence and structural information to design mutant libraries accurately. However, it is difficult for being applied to proteins lacking structural and functional information. In the directed evolution strategy, multiple rounds of mutation and screening experiments are performed on target genes, which are not limited by structural and functional information, but high-throughput screening methods are required. Artificial intelligence-assisted protein engineering requires a large amount of sequence-function data, which can be derived from experiments, calculations, and databases. Through the predictive model, the sequence space of protein mutants can be explored more effectively)

图2 GT1家族糖基转移酶预测模型（GT-Predict）的工作流程［19］（基于功能的算法学习方法GT-Predict，使用来源于酶、亲电试剂和亲核试剂的多种训练集来创建基于物理化学和局部序列的分类器，从而预测GT1糖基转移酶的催化活性和功能信息。Nuc表示亲核基团的数量/类型）

Fig. 2 Workflow for predicting the GT1 glycosyltransferase model (GT-Predict)[19](The function-based algorithmic learning approach, GT-Predict, uses a diverse training set of enzymes, electrophiles, and nucleophiles to create a physicochemical and local-sequence-based classifier for predicting the novel transformations and functional annotation of GT group-transfer enzymes.)

图3 人工智能辅助的视紫红质通道蛋白改造的工作流程［26］［在重组文库中表征的102种ChR蛋白和文献中报道的61种变体，共同构成了（1）分类模型的训练集。然后，使用经过训练的分类模型来预测12000个未表征的ChR序列变体是否具有功能。接下来，构建了三个（2）回归模型，分别针对不同的ChR光电流特性：光电流强度，关闭动力学和光电流的波长敏感性］

Fig. 3 Workflow for machine learning-guided channelrhodopsin engineering[26][102 ChR proteins characterized in the recombinant library, together with 61 variants reported in the literature, constitute the training set of the classification model (1). Then the trained classification model was used to predict whether 12000 uncharacterized ChR sequence variants are functional, and three regression models (2) were trained, one for each of the ChR photocurrent properties of interest: photocurrent strength, off-kinetics and wavelength sensitivity of the photocurrents.]

图4 监督学习的流程示意图［27］（a）准备数据：来源于实验，计算或数据库的数据通常会转换成计算机可以识别的格式，并拆分为训练集和测试集；（b）构建预测模型：利用训练集训练不同的算法以找到决策边界，构建预测模型，例如随机森林，神经网络和支持向量机；（c）验证模型：对于分类问题或者回归问题，应选择合适的评估方法

Fig. 4 Schematic diagram of the supervised learning process[27]Step (a): Preparing data. The data from experiments, calculations or databases are usually converted to a format that the computer can recognize and split into the training and test parts. Step(b): Constructing a predictive model. Using the training set to train different algorithms to find decision boundaries, such as random forests, neural networks and support vector machines, so as to build predictive models. Step (c): Validating the model. An appropriate evaluation method should be selected for tasks with classification or regression.

图5 k折交叉验证示意图（将训练数据进一步细分为k个子集，并且将训练工作流程重复k次，同时保留k个子集中的一个用于评估，其余k-1个子集用于训练）

Fig. 5 Schematic diagram for k-fold cross-validation(The training data is further subsplit into k subsets, and the training workflow is repeated k times with each of the k subsets holding for evaluation and the remaining k-1 subsets used for training)

图6 独热编码示意图（N个蛋白质突变体序列中L个氨基酸中某一相同位置包含S种不同的氨基酸，独热编码将这S个氨基酸都表示为包括S-1个0和一个1的S维向量，其中1的位置表示该位置的氨基酸的种类）

Fig. 6 Schematic diagram for one-hot encoding(A certain position of the L amino acids in the N protein mutant sequence contains S different amino acids. The one-hot encoding represents all S amino acids as an S-dimensional vector including S-1 zeros and one 1. The position of 1 indicates the type of amino acid at that position.)

图7 UniRep模型的工作流程［67］［在训练部分，UniRep模型使用了2400万个氨基酸序列作为训练集。然后使用训练好的模型来预测下一个氨基酸（使交叉熵损失最小化），从而学会如何正确表示氨基酸。在应用部分中，训练后的模型通过提取和平均各个氨基酸的数字向量，从而生成输入序列的单个固定长度矢量表示。这些向量可以用于训练顶级模型，从而应用于多种序列-功能预测任务］

Fig. 7 Workflow for the UniRep model[67][In the training part, 24 million amino acid sequences are used to train the UniRep model. Then the trained model is used to predict the next amino acid (minimizing the cross-entropy loss), so as to learn how to correctly represent the amino acid. In the application part, by extracting and assessing the numerical vector associated with the amino acid, the trained model is used to generate a single fixed-length vector representing the input sequence. Next, these vectors can be used to train top models, which can be applied to various sequence-function prediction tasks.]