2021 AI 编年史：AlphaFold2 与 AI for Science（RosettaFold、TRFold）

发表于 分类于 阅读次数：
2021 年 DeepMind AlphaFold2 在 CASP14 解决蛋白质结构预测，RosettaFold 与 TRFold 紧随其后。本文详解算法架构、Evoformer、优缺点与开源生态。

2021 AI 编年史：AlphaFold2 与 AI for Science | AlphaFold2 and AI for Science in 2021

一、概述与背景知识 | Overview & Background

English

Protein structure prediction — determining the 3D atomic coordinates of a protein from its amino acid sequence — is a 50-year grand challenge in biology. In November 2020, DeepMind’s AlphaFold2 achieved atomic-level accuracy at CASP14 (Critical Assessment of Structure Prediction), effectively solving the problem for single-chain proteins. Throughout 2021, the field exploded:

  • AlphaFold2 paper and open-source code (July 2021)
  • AlphaFold Protein Structure Database — 350k+ structures released
  • RosettaFold (Baker Lab, UW) — competitive accuracy with different architecture
  • TRFold and other transformer-based fold predictors

Key terms:

Term Definition
Primary structure Linear amino acid sequence
Secondary structure Local motifs: α-helices, β-sheets
Tertiary structure Full 3D fold of a single polypeptide chain
MSA (Multiple Sequence Alignment) Evolutionary homolog sequences revealing co-evolution constraints
CASP Biennial blind benchmark for structure prediction
TM-score Metric (0–1) comparing predicted vs. experimental structures; >0.5 ≈ correct fold
Evoformer AlphaFold2’s core module processing MSA and pair representations
pLDDT Per-residue confidence score output by AlphaFold2

中文

蛋白质结构预测 — 从氨基酸序列推断蛋白质三维原子坐标 — 是生物学 50 年重大挑战。2020 年 11 月 DeepMind AlphaFold2 在 CASP14 达到 原子级精度，基本解决单链蛋白预测。 2021 年 领域爆发：

  • AlphaFold2 论文与 开源代码（2021 年 7 月）
  • AlphaFold 蛋白质结构数据库 发布 35 万+ 结构
  • RosettaFold（Baker 实验室）以不同架构达到可比精度
  • TRFold 等 Transformer 折叠预测器涌现

核心术语：

术语 含义
一级结构 氨基酸线性序列
二级结构 局部 motif：α 螺旋、β 折叠
三级结构 单条多肽链的完整三维折叠
MSA（多序列比对） 进化同源序列，揭示共进化约束
CASP 结构预测双年盲评 benchmark
TM-score 预测与实验结构相似度（0–1）；>0.5 通常表示折叠正确
Evoformer AlphaFold2 核心模块，处理 MSA 与 pair 表示
pLDDT AlphaFold2 输出的逐残基置信度

AlphaFold2 被《Science》评为 2021 年度突破，标志着 AI for Science 从辅助工具升级为 科学发现基础设施。

二、技术架构 | Architecture

2.1 AlphaFold2 端到端流水线

flowchart TB
  subgraph Input["Input"]
    SEQ[Amino Acid Sequence]
  end
  subgraph Prep["Feature Preparation"]
    MSA[MSA Search - JackHMMER/MMGalign]
    TPL[Template Search]
    FEAT[Input Features]
  end
  subgraph Evo["Evoformer Stack"]
    MSAR[MSA Representation]
    PAIR[Pair Representation]
    EB[Evoformer Blocks x48]
  end
  subgraph Struct["Structure Module"]
    IPA[Invariant Point Attention]
    FA[Frame Aligned Point Error]
    IT[Iterative Refinement x8]
  end
  subgraph Out["Output"]
    PDB[3D Coordinates]
    PLDDT[pLDDT Confidence]
  end
  SEQ --> MSA
  SEQ --> TPL
  MSA --> FEAT
  TPL --> FEAT
  FEAT --> MSAR
  FEAT --> PAIR
  MSAR --> EB
  PAIR --> EB
  EB --> IPA
  IPA --> IT
  IT --> PDB
  IT --> PLDDT

English

AlphaFold2 replaces the traditional template modeling + physics-based refinement pipeline with a single end-to-end differentiable network:

  1. Input features: MSA (evolutionary co-variation), optional structural templates, residue indices
  2. Evoformer: Alternating updates to MSA row representation (per-sequence) and pair representation (residue i–j relationships) via row/column attention and outer-product mean
  3. Structure Module: Predicts backbone frames using Invariant Point Attention (IPA) — SE(3)-aware attention in 3D — with iterative coordinate refinement
  4. Loss: FAPE (Frame Aligned Point Error) + auxiliary distogram/MLM losses

中文

AlphaFold2 以 端到端可微网络 取代传统 模板建模 + 物理精炼 流水线：

  1. 输入特征：MSA（进化共变）、可选结构模板、残基索引
  2. Evoformer：交替更新 MSA 行表示 与 残基对表示，经行/列注意力与外积均值
  3. 结构模块：不变点注意力（IPA） 在 3D 中 SE(3) 感知地预测骨架框架，迭代精炼坐标
  4. 损失：FAPE + 辅助 distogram/MLM 损失

2.2 RosettaFold 对比架构

1
2
3
4
5
6
7
8
RosettaFold (Three-track Network)
Track 1: MSA (1D per-sequence features)
Track 2: Pair (2D residue-residue)
Track 3: Structure (3D coordinates, updated iteratively)
         ↓
    Triangle Attention + Triangular Multiplicative Update
         ↓
    Structure Prediction + Energy Minimization (optional)
特性 AlphaFold2 RosettaFold
核心模块 Evoformer + IPA Three-track + Triangle updates
开源时间 2021-07 2021-08
复合物预测 AF-Multimer (later) 需 Rosetta 框架组合
典型 TM-score >0.9 (easy targets) 接近 AF2

2.3 TRFold：纯 Transformer 路线

English

TRFold (Transformer-based protein folding) explored whether standard Transformer blocks with geometric constraints could match specialized architectures — contributing to the 2021 debate on inductive bias vs. scale in scientific ML.

中文

TRFold 探索 标准 Transformer 加几何约束能否媲美专用架构，参与 2021 年科学 ML 中 归纳偏置 vs. 规模 的讨论。

English

  1. Structure → Function → Design: Accurate structures enabled de novo enzyme design, drug target identification, and mutation effect prediction.
  2. Open data revolution: AlphaFold DB democratized structural biology — previously X-ray crystallography took months per protein.
  3. Multimer & dynamics: 2021 laid groundwork for protein complex (AF-Multimer) and conformational ensemble prediction.
  4. Integration with cryo-EM: AI predictions + experimental density maps accelerated hybrid modeling.
  5. Competitive open ecosystem: RosettaFold, OpenFold (2022) ensured non-Google reproducibility.
  6. AI for Science template: Success pattern — large curated datasets + strong inductive bias + massive compute — copied in materials, climate, and drug discovery.

中文

  1. 结构 → 功能 → 设计：精准结构推动 从头酶设计、药物靶点发现、突变效应预测。
  2. 开放数据革命：AlphaFold DB 民主化结构生物学——此前 X 射线晶体学 每个蛋白需数月。
  3. 多聚体与动力学：为 蛋白复合物（AF-Multimer）与 构象 ensemble 预测铺路。
  4. 与 cryo-EM 融合：AI 预测 + 实验密度图加速 混合建模。
  5. 开放竞争生态：RosettaFold、OpenFold 保障 非 Google 可复现性。
  6. AI for Science 模板：大数据 + 强归纳偏置 + 大算力 模式被材料、气候、药物等领域复制。

四、优缺点分析 | Pros & Cons

维度 优点 Advantages 缺点 Disadvantages
精度 CASP14 原子级，远超传统方法 内在无序蛋白（IDP）仍困难
速度 分钟级 vs. 实验数月 GPU 内存与 MSA 搜索仍耗时
覆盖 AF DB 覆盖 UniProt 大部分 膜蛋白、超大复合物精度下降
置信度 pLDDT 指示可靠区域 低 pLDDT 区域易被误用
开源 代码+权重公开（学术许可） 商用需单独授权
生物学 静态结构，非动力学 缺少配体、翻译后修饰上下文
依赖 MSA 进化信息提升精度 orphan 序列（无同源）性能降

五、应用场景 | Use Cases

场景 说明
药物发现 靶点结构用于 virtual screening 与先导化合物优化
酶工程 预测突变对稳定性的影响，指导定向进化
农业生物技术 作物抗病蛋白结构解析
基础生物学 未解析 ORF 的功能注释
合成生物学 设计新-to-nature 蛋白骨架
疾病研究 致病突变如何破坏蛋白折叠
教育科研 零实验成本获取教学用高质量结构

六、开源项目与工具 | Open Source & Tools

项目 说明 URL
AlphaFold DeepMind 官方实现 https://github.com/deepmind/alphafold
ColabFold MSA 加速 + Google Colab 一键预测 https://github.com/sokrypton/ColabFold
RosettaFold Baker 实验室官方代码 https://github.com/RosettaCommons/RoseTTAFold
OpenFold AlphaFold2 可训练复现（2022，源于 2021 需求） https://github.com/aqlaboratory/openfold
ESMFold 无 MSA 语言模型折叠（后续演进） https://github.com/facebookresearch/esm
PyMOL / ChimeraX 结构可视化 https://pymol.org/
PDBe 实验结构数据库 https://www.ebi.ac.uk/pdbe/

七、参考文献 | References

  1. Jumper, J., et al. “Highly accurate protein structure prediction with AlphaFold.” Nature 596, 2021. https://www.nature.com/articles/s41586-021-03819-2
  2. Baek, M., et al. “Accurate prediction of protein structures and interactions using a three-track neural network.” Science 373, 2021 (RosettaFold). https://www.science.org/doi/10.1126/science.abj8754
  3. Senior, A.W., et al. “Improved protein structure prediction using potentials from deep learning.” Nature 577, 2020 (AlphaFold1). https://www.nature.com/articles/s41586-019-1923-7
  4. Varadi, M., et al. “AlphaFold Protein Structure Database.” Nucleic Acids Research, 2022. https://alphafold.ebi.ac.uk/
  5. CASP14 Official Results. https://predictioncenter.org/casp14/
  6. Callaway, E. “‘The game has changed.’ AI triumphs at protein folding.” Nature News, 2020. https://www.nature.com/articles/d41586-020-03348-4
  7. Mirdita, M., et al. “ColabFold: making protein folding accessible to all.” Nature Methods, 2022. https://doi.org/10.1038/s41592-022-01488-1

English Summary: AlphaFold2 in 2021 transformed structural biology from a bottleneck into a solved inference problem for most single-chain proteins — launching the modern AI for Science era that extends to materials, climate, and genomics.

中文总结：2021 年 AlphaFold2 将结构生物学从瓶颈转变为对大多数单链蛋白的「已解推理问题」，开启延伸至材料、气候、基因组等领域的 AI for Science 新时代。