图1 基于序列相似性的蛋白质挖掘的对比a：通过从数据库中选取具有代表性的种子序列，结合 HMMER、Hhpred等工具进行同源性搜索与特征提取，构建初步候选空间。随后结合系统发育分析与功能注释信息进行筛选与过滤，并通过实验方法验证其潜在功能。b：以原始氨基酸序列为输入，利用大型语言模型（如Transformer架构）学习其潜在语义表示，进而执行结构预测与功能预测任务，输出功能相关的候选序列、结构与标签信息

Fig. 1 Comparison of sequence similarity-based protein mining approachesa: Representative seed sequences are selected from databases and used for homology search and feature extraction via tools such as HMMER and HHpred, forming an initial candidate space. Phylogenetic analysis and functional annotations are then applied for further filtering, followed by experimental validation of potential functions. b: Raw amino acid sequences are used as input for large language models (e.g., Transformer-based architectures) to learn latent semantic representations, enabling structure and function prediction tasks and yielding candidate sequences, structures, and functional labels

图2 蛋白质的设计目标

Fig. 2 Design objectives of proteins

图3 基于序列的蛋白质设计a：输入层包括多序列比对信息与功能标签，模型层分为“仅序列模型”与“序列-标签模型”，输出层为预测的目标序列或功能属性，如热稳定性、水溶性或催化活性变化等。b：上图展示了掩码语言模型与自回归语言模型的区别；下图显示Transformer的典型结构，包括输入嵌入、注意力机制与残差连接。掩码模型支持双向信息流，适用于上下文建模；自回归模型支持单向信息流，用于序列生成。c：输入蛋白质序列和功能标签后，分别通过局部卷积表示和全局注意力机制编码，融合后用于生成目标序列或标签预测。输出端可采用softmax或sigmoid策略

Fig. 3 Sequence-based protein designa: The input layer includes multiple sequence alignment (MSA) data and functional labels. The model layer is divided into sequence-only models and sequence-to-label models. The output layer yields predicted target sequences or functional properties, such as thermal stability, solubility, or changes in catalytic activity. b: The upper panel illustrates the difference between masked language models and autoregressive language models. The lower panel shows the typical transformer architecture, including input embedding, attention mechanisms, and residual connections. Masked models support bidirectional information flow for contextual modeling, while autoregressive models operate with unidirectional flow for sequence generation. c: Given protein sequences and functional labels as inputs, the model encodes local features through convolutional layers and global context via attention mechanisms. These representations are then fused to generate target sequences or predict functional labels. The output can be modeled using a softmax layer for sequence generation or a sigmoid layer for multi-label classification

图4 基于结构的蛋白质设计a：结构预测模型致力于从序列预测结构P(xz)；结构生成模型直接在结构空间中建模P(z)；结构-序列联合建模方法联合学习P(x,z)，实现结构与序列的协同设计。b：基于模板的方法包括同源建模和穿线方法，依赖已知蛋白质结构模板；自由建模方法在无模板条件下进行结构构建。图中列出各类典型预测工具及其方法归属。c：通过多序列比对生成模板信息，经预训练模型提取残基-残基特征，结合结构推理模块预测原子坐标，实现高精度结构建模。该流程体现了演化信息与几何学习的融合。d：按照功能目标分为4个阶段：（1）生成具备合理折叠结构的蛋白质；（2）构建具有指定几何功能位点的蛋白；（3）增强与小分子的特异性结合能力；（4）构建具备催化功能的蛋白，支持将底物 S 转化为产物P的反应功能

Fig. 4 Structure-based protein designa: Structure prediction models aim to infer protein structures from sequences by learning P(xz); structure generation models directly model the structural distribution P(z); and joint structure-sequence models learn the joint distributionP(x, z), enabling coordinated sequence-structure design. b: Template-based methods, including homology modeling and threading, rely on known protein structure templates, while free modeling methods generate structures without using templates. The figure categorizes representative structure prediction tools according to their methodological classes. c: Template information is generated via multiple sequence alignment and fed into pretrained models to extract residue-residue features. These are combined with structural inference modules to predict atomic coordinates, enabling high-accuracy structure modeling. This workflow integrates evolutionary information with geometric learning.d: Structure generation can be categorized into four stages based on functional objectives: (1) generating proteins with properly folded structures; (2) constructing proteins with specified geometric functional sites; (3) enhancing specific binding affinity to small molecules; and (4) designing proteins with catalytic functions that enable the transformation of substrate S into product P