跳至正文

AlphaFold「解决了蛋白质折叠」吗

目录

机制裁决红队风第九十篇 · 对称双向红队第八十五篇 · AI for Science 谱系蛋白质折叠侧 · 接希格斯/ΛCDM/暗物质/真空自然性/物质-反物质五篇物理主线后转 AI for Science 首靶 · 全库第 147 篇


〇 母裁决 · 七层硬度光谱

结构胎记=预测分数冒充机制理解 × 静态结构冒充动态折叠 × 单蛋白结构冒充蛋白质组/功能/药物设计(三跳升格链:分数跳 × 本体跳 × 外推跳)。

命题 硬度 判据
CASP14 单域结构预测分数真 ✓ 硬 GDT_TS 中位 92.4、多平台复现、盲测
Anfinsen 热力学假说真(序列决定天然态) ✓ 硬 1961 核糖核酸酶复性实验、1972 诺贝尔奖
「蛋白质折叠问题」是三问非一问 ◐ 半硬 Dill & MacCallum 2012 三问框架(密码/机制/预测)
折叠动力学/能量景观/Levinthal 佯谬 ◐ 半硬 漏斗理论真但全貌未闭合
IDP/构象集合体/动态功能 ◐ 半硬 30–40% 真核残基无序、功能依赖动态
药物设计/蛋白质组/相互作用 ○ 软-半硬 零 AI 药物获 FDA 批准、静态结构不充分
「AI 攻克生物学/解决了一切」 × 软 叙事升格、无硬终点支撑

母裁决:CASP14 单域结构预测分数真(GDT_TS 中位 92.4、多盲测复现)、Anfinsen 热力学假说真(序列在给定环境决定天然态自由能最低构象)、AlphaFold 作为工具真且变革性(2.14 亿预测结构、330 万用户、实验结构提交反增 40%)——真的不是「蛋白质折叠问题已被解决/机制已懂」(Dill 三问中 AlphaFold 只触及第三问「能否预测结构」、Outeiral 2022 证预测轨迹与折叠速率「uncorrelated with experimental observables」),也不是「AlphaFold 是炒作/对生物学无用」(CASP 评估者判「single protein chains…solved」、Nature Methods 年度方法、实验结构提交不降反升)。

灵魂句:预测分数是真的、Anfinsen 是真的、AlphaFold 作为工具是变革性的——真的不是「蛋白质折叠已被解决」和「AI 已攻克生物学」这两层被声称的胜利;CASP 评估者说的是”单链结构预测”,媒体听到的是”生命被破解”——中间隔着整个折叠动力学、构象集合体和药物设计的鸿沟。


一 CASP 与 AlphaFold:预测分数的真锚

1.1 CASP 是什么

CASP(Critical Assessment of protein Structure Prediction)由 John Moult 等人于 1994 年创立,是蛋白质结构预测领域的双年盲测竞赛。参赛者收到近期实验解析但尚未公开的氨基酸序列,须预测三维结构;结果以 GDT_TS(Global Distance Test – Total Score)衡量——计算预测模型中 Cα 原子在最优叠合后落入 1/2/4/8 Å 阈值的百分比均值,范围 0–100。

来源:CASP 官方;Kryshtafovych A, Schwede T, Topf M, Fidelis K, Moult J. “Critical Assessment of Methods of Protein Structure Prediction (CASP) — Round XIV.” Proteins 89(12):1607–1665 (2021). DOI: 10.1002/prot.26237

1.2 CASP14:AlphaFold2 的里程碑

2020 年 CASP14,DeepMind 的 AlphaFold2 取得:

  • 中位 GDT_TS:92.4(全部靶标);自由建模类中位 87.0
  • Z-score:244.0 vs 第二名(BAKER/RoseTTAFold)0.99
  • 中位 Cα RMSD:0.96 Å;全原子精度 ~1.5 Å(95% 残基)
  • 30 个单域靶标中 25 个 GDT_TS > 80

来源:Jumper J, Evans R, Pritzel A, et al. “Highly accurate protein structure prediction with AlphaFold.” Nature 596:583–589 (2021). DOI: 10.1038/s41586-021-03819-2(主笔亲核 PubMed 摘要:「more than 50 years」「atomic accuracy」「competitive with experimental structures」);Jumper J, et al. “Applying and improving AlphaFold at CASP14.” Proteins 89(12):1711–1721 (2021). DOI: 10.1002/prot.26257

CASP14 评估者共识(Kryshtafovych et al. 2021,经 Lupas et al. 2021 转述):

“The models filed by DeepMind’s structure prediction team using the program AlphaFold2 were often essentially indistinguishable from experimental structures, leading to a consensus in the community that the structure prediction problem for single protein chains has been solved.”

来源:Lupas AN, et al. “AlphaFold2 and its applications in the fields of biology and medicine.” Biochemical Journal 478(11):2077–2098 (2021). DOI: 10.1042/BCJ20200963(主笔亲核)

关键限定:「for single protein chains」——不是复合物、不是膜蛋白、不是无序蛋白、不是动力学。

1.3 AlphaFold2 架构要点

  • Evoformer:迭代精炼 MSA(多序列比对)与残基对表征
  • 结构模块:SE(3) 等变,直接输出三维坐标
  • 端到端可微:整体训练为单一可微管线
  • 训练数据:PDB ~170,000 实验结构(2020 年快照)+ UniRef90/BFD 序列库
  • pLDDT(predicted Local Distance Difference Test):每残基置信度 0–100;≥90 极高、70–90 可靠主链、50–70 低置信、<50 很可能无序

来源:Jumper et al. 2021(同上);Bhat VT, et al. 架构综述 PMC8329862: 链接

1.4 AlphaFold3(2024)

Abramson J, et al. “Accurate structure prediction of biomolecular interactions with AlphaFold 3.” Nature 630:493–500 (2024). DOI: 10.1038/s41586-024-07487-w

  • 扩散式架构取代 AF2 结构模块
  • 可预测蛋白质-核酸-小分子-离子复合物
  • 自认局限:「Conformation coverage is limited」、手性错误、抗体预测依赖种子数

1.5 AlphaFold 蛋白质结构数据库

  • >2.14 亿预测结构(2022 年扩展至 UniRef90 大部分)
  • 初始发布(2021):~360,000(人类蛋白质组 + 模式生物)
  • 用户 >330 万、覆盖 >190 国、Nature 2021 论文 ~40,000 次引用(5 年)
  • 实验结构提交在 AlphaFold 发布后反增 >40%(非取代)

来源:Varadi M, et al. “AlphaFold Protein Structure Database…” Nucleic Acids Res 50(D1):D439–D444 (2022). DOI: 10.1093/nar/gkab1061AlphaFold EBI;Nature 五周年报道

1.6 CASP15(2022)与 CASP16(2024)

  • CASP15:DeepMind 未参赛,但「Everyone is using AlphaFold」;评估者判单蛋白预测「basically solved」;焦点转向复合物与无序区
  • CASP16:AlphaFold3 主导;评估者:「largely a solved problem」但「challenges persist in modeling large, complex assemblies」

来源:CASP15 报道CASP16 评估


二 Anfinsen 热力学假说与「蛋白质折叠问题」三义

2.1 Anfinsen 的热力学假说

Christian Anfinsen 1961 年核糖核酸酶 A 实验:尿素 + 巯基乙醇变性/还原后失活、二硫键 scrambling;去除变性剂重新氧化后蛋白自发恢复天然构象与酶活。

1972 年诺贝尔化学奖(1973 年演讲)核心命题:

“the native conformation is determined by the totality of interatomic interactions and hence by the amino acid sequence, in a given environment.”

“the three-dimensional structure… is the one in which the Gibbs free energy of the whole system is lowest.”

来源:Anfinsen CB. “Studies on the Principles that Govern the Folding of Protein Chains.” Science 181:223–230 (1973). Nobel Lecture PDF: 链接(子代理经 r.jina.ai 逐字亲核)

热力学假说的精确含义与边界:它是关于终点/平衡态的命题——天然态是自由能最低态。它回答:(a) 如何到达(路径/机制),(b) 多快到达(速率)。

2.2 「蛋白质折叠问题」是三问非一问

Dill KA, MacCallum JL. “The Protein-Folding Problem, 50 Years On.” Science 338(6110):1042–1046 (2012). DOI: 10.1126/science.1219021(主笔亲核 PubMed)

三问框架:

“(i) What is the physical code…? (ii) How can proteins fold so fast? (iii) Can we devise a computer algorithm…?”

即:

  1. 折叠密码(序列如何决定结构——热力学)
  2. 折叠机制/速度(为何能这么快——动力学/Levinthal 佯谬)
  3. 结构预测(能否设计算法预测——工程)

Dill 2008 版(PMC2443096)同框架:

“The ‘protein folding problem’ consists of three closely related puzzles: (a) What is the folding code? (b) What is the folding mechanism? (c) Can we predict the native structure…?”

来源:Dill KA, Bromberg S, et al. “Protein folding in the landscape perspective.” Annu Rev Biophys 37:289–316 (2008). PMC2443096

AlphaFold 只触及第三问。第一问(密码的物理本质)与第二问(机制/动力学)并未因 AlphaFold 而被「解决」。

2.3 Dill 本人的判断

“It is no longer useful to talk about ‘solving the protein-folding problem’.”

——Dill & MacCallum 2012(子代理经 Science 页面核实)

PNAS 2022 进一步区分:

Dill KA, et al. “Protein folds vs. protein folding: Differing questions, different challenges.” PNAS 119(49):e2214423119 (2022). DOI: 10.1073/pnas.2214423119


三 折叠动力学:Levinthal 佯谬、能量景观与分子伴侣

3.1 Levinthal 佯谬

Cyrus Levinthal 1969 年指出:若每残基有 3 种稳定构象,100 残基蛋白有 3^100 ≈ 10^48 种构象;随机遍历需 ~10^27 年——远超宇宙年龄。但蛋白质在毫秒–秒级完成折叠。

佯谬的真正含义(常被误述为「蛋白不可能折叠」):Levinthal 本人以此论证折叠不是随机搜索,而是被引导的——局部相互作用快速形成,引导后续折叠。

来源:Levinthal C. “How to Fold Graciously” (1969);Wikipedia: Levinthal’s paradox

3.2 能量景观 / 折叠漏斗

  • Bryngelson & Wolynes 1987:最小阻挫原理(principle of minimal frustration)——天然态相互作用被进化优化,使能量景观呈漏斗状
  • Dill & Chan 1997:「From Levinthal to pathways to funnels」——用漏斗取代单一 pathway 概念
  • Onuchic, Luthey-Schulten & Wolynes 1997:折叠是「progressive organization… into an ensemble (a funnel)」

景观告诉我们而静态结构不能告诉我们的:折叠速率、多条并行路径、构象系综、过渡态、为何能快速到达(偏置随机游走)。

来源:Bryngelson JD, Wolynes PG. PNAS 84:7524–7528 (1987);Dill KA, Chan HS. Nat Struct Biol 4:10–19 (1997). DOI: 10.1038/nsb0197-10;Onuchic JN, et al. Annu Rev Phys Chem 48:545–600 (1997);PMC4256721

3.3 分子伴侣:热力学假说与体内现实的鸿沟

伴侣蛋白(Hsp70、GroEL/GroES)不提供最终结构信息(不违背 Anfinsen),作用是动力学/防护性的:结合暴露疏水中间态、防止错误聚集、提供隔离的「Anfinsen 笼」。

“The majority of molecular chaperones do not convey any steric information for protein folding, and instead assist in protein folding by binding to and stabilizing folding intermediates.”

来源:Wikipedia: Chaperone (protein))

含义:Anfinsen 是稀溶液、体外、热力学命题;细胞内是高浓度、易聚集、共翻译、有竞争路径的动力学环境——序列「编码」结构,但不保证在生理时限内无聚集地到达。

3.4 错误折叠疾病:同一序列,不同折叠

  • 朊病毒(PrP):同一序列可形成多种稳定构象,构象本身可传染
  • 阿尔茨海默病:Aβ + Tau 淀粉样聚集
  • 帕金森病:α-Synuclein 聚集
  • 关键热力学点:淀粉样态在某些条件下比天然态更稳定

来源:Dobson CM. “Protein folding and misfolding.” Nature 426:884–890 (2003). DOI: 10.1038/nature02261PMC4286994

含义:预测出天然结构,不能预测该序列是否会在何种条件下走入错误折叠/聚集路径。信息不止于序列。

3.5 AlphaFold 与折叠动力学的脱节

Outeiral C, Nissley DA, Deane CM. “Current structure predictors are not learning the physics of protein folding.” Bioinformatics 38(7):1881–1888 (2022). DOI: 10.1093/bioinformatics/btac043(主笔亲核 PMC8963306)

“The folding trajectories produced are also uncorrelated with experimental observables such as intermediate structures and the folding rate constant.”

预测能力「worse than a trivial classifier using sequence-agnostic features like chain length」

AlphaFold 2 does not provide “an enhanced understanding of protein folding.”

裁决:AlphaFold 绕过了折叠过程(bypassed),而非理解了折叠过程。

来源补充:APS Physics 2024 诺贝尔报道:AlphaFold “bypassed the need to simulate the folding process.” 链接


四 构象集合体、IDP 与动态功能

4.1 蛋白质不是单一结构

“The core idea is that a protein does not have a single structure.”

“a protein stochastically hops between an enormous set of alternative structures.”

“The frontier is conformational ensembles.”

来源:Bowman GR. “AlphaFold and protein folding: Not dead yet! The frontier is conformational ensembles.” Annu Rev Biomed Data Sci (2024). PMC11892350(主笔亲核)

“proteins exist not in a single, rigid structure, but as a dynamic ensemble of conformations.”

来源:PMC7458412

“Proteins perform their function by accessing a suitable conformer from the ensemble of available conformations.”

来源:PMC9108755

AlphaFold 预测单一结构——这是它的设计目标,也是它的本体论限制。

4.2 内在无序蛋白(IDP)

Wright PE, Dyson HJ. “Intrinsically Disordered Proteins in Cellular Signaling and Regulation.” Nat Rev Mol Cell Biol 16:369–381 (2015). DOI: 10.1038/nrm3920(主笔亲核 PMC4405151)

“regions of disorder are very common in eukaryotic proteins, especially those involved in cellular regulation and signaling.”

流行率数据:

  • 真核蛋白质组 30–40% 残基位于无序区(PMC11459374
  • >60% 人类蛋白含至少一个 IDR 片段(PMC12292460
  • 33% 真核蛋白有功能相关的长无序区 vs 古菌/细菌仅 2–4%(PMC9693201

“IDPs challenge the well-established foundational idea of structure-function relationship in molecular biology.”

来源:PMC8283027

“moving from the classical structure-function paradigm… to the structure-function continuum concept.”

来源:Uversky VN, et al. Frontiers in Physics 7:10 (2019). 链接

4.3 AlphaFold 与 IDP

  • pLDDT < 50 通常标记无序/柔性区域——低分既是失败也是信号
  • pLDDT ≤ 70 被用于识别无序残基(PMC11992697
  • AlphaFold 对 IDP 给出的「结构」可能是无意义的强制折叠

“AlphaFold 2 can neither afford insight into how proteins fold, nor can it determine protein stability or dynamics.”

“Rare folds or minor alternative conformations are also not predicted by AlphaFold 2.”

来源:Bhat VT, et al. Front Mol Biosci 9:906437 (2022). 链接

4.4 膜蛋白

“structure prediction of membrane proteins by AlphaFold is not reliable” due to “inconsistencies in the location of the transmembrane domains.”

来源:PMC10011655

  • 拓扑正确率 97% 仅限「excellent quality」子集(~40% TM 蛋白质组)
  • 柔性连接环常被错误跨膜

来源:PMC10662385


五 药物设计、蛋白质相互作用与功能鸿沟

5.1 静态结构为何不充分

药物设计需要但 AlphaFold 不提供的:

维度 需要 AlphaFold 状态
结合动力学(kon/koff/驻留时间) 时间分辨 ✗ 静态
别构效应 多构象态 ✗ 单一结构
诱导契合/构象选择 柔性受体 ✗ 刚性模型
水网络 显式溶剂 ✗ 无
条件依赖(pH/温度/拥挤) 环境建模 ✗ 无
翻译后修饰 化学修饰 ✗ 不建模

“Too much emphasis… has been placed on rigid structures… regardless of their experimental or theoretical origin.”

来源:PMC2701403

“Receptor motions clearly play an essential role in the binding of most small-molecule drugs.”

来源:PMC3203851

“what you see is not always what you get. Other equally important forces, namely the entropies, are not visible.”

来源:PMC2756098

5.2 AlphaFold 与药物发现:实际影响

  • 零 AI 发现药物获 FDA 批准(截至 2026 年中)
  • Isomorphic Labs:$2.1B B 轮(2026.05);Lilly $45M 预付/$1.7B 里程碑;Novartis $37.5M 预付/$1.2B;首例人体试验预计 2026 年底
  • 药物开发失败率:~90% 候选在临床折戟(缺乏疗效 40–50%、毒性 30%、成药性 10–15%)

来源:philippdubach.com 评估intuitionlabs.ai

“AlphaFold只是人们迈出的重要一步…无法考虑到靶点蛋白的周围环境信息…静态的蛋白质结构可能无法完全揭示药物结合时的真实状态…预测手段并不能取代实验结果。”

来源:cnpharm.com

“AlphaFold3 does not predict binding based on molecular interactions, but based on general protein patterns.”

来源:PMC11788922

5.3 蛋白质-蛋白质相互作用

AlphaFold-Multimer(Evans R, et al. bioRxiv 2021.10.04.463034):

  • 异源二聚体:67% 界面正确、23% 高精度
  • 同源二聚体:69% 界面正确、34% 高精度
  • 抗体-抗原复合物:「low success rate」
  • 「accuracy decreases with the number of chains」
  • 「only 60% of dimers are accurately predicted」

来源:PMC10011802PMC11302914PMC12149419

5.4 功能鸿沟:知道结构 ≠ 知道功能

  • PDB 中 >20% Pfam 域为 DUF(Domain of Unknown Function)
  • 「ability to produce structures has outstripped our ability to analyze them」
  • 实例:1htw 2001 年存入,至 2006 年无序列相似性;1fl9 从存入到功能注释间隔 11 年

来源:PMC2954198PMC3497298

“functional encoding does not come down to the structure either… sets of atomic motions… encode a specific function.”

来源:Front Mol Biosci 2021


六 叙事审计:诺贝尔、媒体与「解决了」的滑移

6.1 诺贝尔委员会的措辞

2024 年诺贝尔化学奖:

  • Baker:「for computational protein design」
  • Hassabis & Jumper:「for protein structure prediction」

新闻稿标题:

“They cracked the code for proteins’ amazing structures.”

正文:

“Demis Hassabis and John Jumper have developed an AI model to solve a 50-year-old problem: predicting proteins’ complex structures.”

委员会主席 Heiner Linke:

“fulfilling a 50-year-old dream.”

来源:Nobel Prize 2024 Press Release(子代理核实)

关键:委员会始终将「solve」限定于「predicting structures」——从未说「understanding folding」或「solving the protein folding problem」全称。

6.2 DeepMind 自己的措辞

博客标题(2020.11.30):

“AlphaFold: a solution to a 50-year-old grand challenge in biology.”

正文 caveat:

“Not every structure we predict will be perfect. There’s still much to learn, including how multiple proteins form complexes, how they interact with DNA, RNA, or small molecules, and how we can determine the precise location of all amino acid side chains.”

来源:DeepMind Blog

DeepMind 用「a solution to」(非「solved」),且归因于 CASP 组织者。但标题被媒体系统性缩短为「solved」。

6.3 媒体升格谱系

来源 标题用词 日期
MIT Technology Review solved a 50-year-old grand challenge” 2020.11.30
The Guardian cracks 50-year-old problem” 2020.11.30
Daily Telegraph crack 50-year-old biological challenge” 2020.11.30
Nature (Callaway) “‘It will change everything’…gigantic leap in solving 2020.11.30
New Scientist decipher secrets of the machinery of life” 2020.11.30
Nature Methods “Method of the Year 2021” / “shock waves 2022.01.11

来源:Nature d41586-020-03348-4: 链接;Nature Methods s41592-021-01380-4: 链接

NYT 标题显著克制:「Claims Breakthrough That Could Accelerate Drug Discovery」——用「Claims」而非「Solved」。

6.4 科学家的回撤与批评

CASP 评估者(精确限定):

“the structure prediction problem for single protein chains has been solved.”

Bowman 2024(主笔亲核):

“Calling structure prediction… ‘solved’ is dangerous as doing so could stymie further progress.”

“AlphaFold does not give us the mechanism of protein folding.”

Quanta Magazine 2024:

  • Michael Littman: “It’s not solved the way a scientist would solve it.
  • Lauren Porter: “Some people are overconfident — like, way overconfident.”
  • Paul Adams: “AlphaFold changed everything and nothing.”
  • Ellen Zhong: “Right now, you just have this black box… not actually how you get there.”
  • Jumper 本人:AlphaFold “is relatively blind” to point mutations.

来源:Quanta Magazine 2024.06.26

Perrakis & Sixma 2023:

“The most serious limitations of AlphaFold… arise from the fact that they are based on learning patterns and know almost nothing about physics and chemistry.”

它们「can generate a single structure… but not a collection of alternative conformations that are influenced in their relative stability by factors such as pH, temperature or the binding of ions, other ligands or other proteins.」

来源:Perrakis A, Sixma TK. “AlphaFold and the future of structural biology.” IUCrJ (2023). PMC10327576

Jumper 2025 采访:

“We are not just one structure away from curing disease.”

来源:QQ News 2025.12

6.5 「50 年」考据

锚点 年份 距 2020
Anfinsen 核糖核酸酶实验 ~1961 59 年
Anfinsen 诺贝尔奖 1972 48 年
Levinthal 佯谬 1969 51 年
CASP1 1994 26 年

「50 年」若锚定 Anfinsen 1972(48 年)或 Levinthal 1969(51 年)近似正确;若锚定 CASP(26 年)则不准确。DeepMind 明确引 Anfinsen:「Just as 50 years ago Anfinsen laid out a challenge」——历史可辩护但取整。

6.6 Baker 的半份奖:设计 ≠ 预测 ≠ 理解

Baker 的 Rosetta/RFDiffusion 是从头设计自然界不存在的蛋白质——与预测已有蛋白质结构是根本不同的任务。

Baker 本人:

“When I started here, I was really focused on trying to understand the principles of protein folding… we really only started working on protein design around the late 1990s.”

来源:Chemistry World

三者区分

  1. 蛋白质设计(Baker):创造新蛋白——测试是否理解折叠规则到能造新结构
  2. 结构预测(Hassabis/Jumper):给序列预测形状——模式识别/统计成就
  3. 理解折叠(机制):知道如何/为何折叠——物理路径/动力学/能量景观

三位获奖者无人声称达成了第三项。

6.7 哲学审计

Schuster A. “Understanding protein folding with machine learning models? The case of AlphaFold2.” Synthese (2026). DOI: 10.1007/s11229-025-05426-4

“despite its empirical success, AF2’s complexity and opacity limit its capacity to contribute directly to the scientific explanation of the PFP and, consequently, to its scientific understanding.”


七 替代方法与实验验证缺口

7.1 替代/互补计算方法

方法 特点 与 AlphaFold 关系
RoseTTAFold (Baek 2021) 三轨网络、Baker 实验室、开源 精度略低、计算需求低
ESMFold (Lin 2023) 语言模型、无需 MSA、快 ~60× 精度低于 AF2、对孤儿蛋白有独特价值
分子动力学(MD) 时间演化构象集合、自由能景观 互补:AF 给起点、MD 给动态

来源:Baek M, et al. Science 373:871–876 (2021);Lin Z, et al. Science 379:1123–1130 (2023). DOI: 10.1126/science.ade2574

7.2 实验方法提供什么预测不能给的

  • 冷冻电镜(cryo-EM):配体结合态、水分子、离子/辅因子、多构象态、近生理条件(2017 诺贝尔化学奖「resolution revolution」)
  • NMR:动力学、构象集合体、淀粉样/凝聚体
  • X 射线晶体学:原子分辨率电子密度、配体/水/金属精确定位

“NMR… provides information on protein folding and dynamics as well as biomolecular condensates and amyloids.”

来源:Bhat VT, et al. Front Mol Biosci 9:906437 (2022). 链接

7.3 实验验证缺口

  • AlphaFoldDB:>2.14 亿预测结构
  • PDB:~210,000–220,000 实验结构
  • 比值:~1000:1——绝大多数预测从未经实验验证
  • Nature Computational Science 2023 社论标题:「Experimental validation, anyone?」

来源:Nature Comp Sci 社论;Terwilliger TC, et al. Nat Methods (2023)——AlphaFold 预测是「valuable hypotheses」可加速但不能取代实验

7.4 具体失败案例

高置信度错误

  • 128 个 NMR 结构对比中,AlphaFold2 不正确预测的 75% 伴随高 pLDDT 分数
  • NaV1.8 离子通道 VSD-I:pLDDT 极高但与 cryo-EM 不对齐;ECL-I 被建模为螺旋但实验显示无结构
  • 去稳定化突变体:AlphaFold 仍给出「native-like structures」——高置信度预测了实验上已不存在的结构

来源:PMC12809598;Nguyen TT, et al. PMC10936637;Balasco N, et al. Biomolecules (2025) PMC12109453

7.5 突变效应盲区

AlphaFold “cannot reproduce the destabilizing effects that single-point mutations induce.”

来源:Balasco et al. 2025(同上)

Jumper 本人承认:AlphaFold “is relatively blind” to point mutations.(Quanta 2024)


八 AI for Science 认识论:预测 vs 理解

8.1 预测 ≠ 理解:哲学框架

de Regt HW, Dieks D. “A Contextual Approach to Scientific Understanding.” Synthese 144:137–170 (2005):科学理解=能识别理论的「qualitatively characteristic consequences」而无需精确计算——区别于纯预测。

Krenn M, Aspuru-Guzik A. “On scientific understanding with artificial intelligence.” Nat Rev Phys (2022). arXiv:2204.01467:AlphaFold 是「oracle」——缺乏定性把握。

8.2 黑箱问题

“Right now, you just have this black box… not actually how you get there.” ——Ellen Zhong(Quanta 2024)

AlphaFold “does not explicitly rely on those [physical] principles.” ——架构综述(PMC8329862)

“based on learning patterns and know almost nothing about physics and chemistry.” ——Perrakis & Sixma 2023

8.3 历史平行

案例 预测 理解
Kepler → Newton 行星轨道(描述) 万有引力(解释)
天气预报 数值模式 ~2 周 湍流/NS 存在性未解(千禧年问题)
AlphaFold 单域结构(终点) 折叠机制/动力学(路径)

模式:预测成功可以先于、独立于机制理解。但预测成功 ≠ 机制理解。

8.4 「End of Theory」叙事

Chris Anderson 2008 Wired

“Correlation supersedes causation.” / “With enough data, the numbers speak for themselves.”

来源:Anderson C. “The End of Theory.” Wired 16.07 (2008). 链接

AlphaFold 不是纯「End of Theory」胜利:它使用进化信息(MSA 共进化模式)和注意力机制——是在已知结构上训练的模式识别,非无模型原始数据。但它不使用显式物理定律(力场、热力学)。

8.5 DeepMind「AI for Science」框架

Hassabis 系列表态:

  • AI 是「ultimate tool advancing knowledge frontier」
  • AlphaFold 是「most concrete, useful case of AI doing something in science」
  • GNoME(2023):预测 220 万新晶体结构、~38 万预测稳定——「equivalent to 800 years of human discovery」

来源:Scientific American;GNoME: Merchant A, et al. Nature 624:80–85 (2023). DOI: 10.1038/s41586-023-06735-9

模式识别:DeepMind 框架系统性地将「prediction」等同于「solution」——AlphaFold「solved the 50-year-old grand challenge」、GNoME「discover 2.2 million new materials」。预测是解决方案的子集,不是全集。

8.6 AlphaMissense(2023)

Cheng J, et al. Science 381(6662):eadl2642 (2023):预测 ~7100 万错义变体致病性,ClinVar 验证 ~90% 正确率。

  • GitHub 明确声明:「AlphaMissense has not been validated for, and is not approved for, any clinical use.」
  • 计算预测仅提供「support」证据,非确定性诊断
  • Nature 2025 发现与临床级变异解释存在不一致

来源:GitHubPMC11683568Nature npj Genomic Medicine 2025

8.7 「真正解决蛋白质折叠」需要什么?

综合文献,完整解决须:

  1. 从第一性原理模拟折叠过程(非只预测终点)
  2. 捕获构象集合体与能量景观全貌
  3. 处理 IDP/多态蛋白/折叠切换
  4. 建模细胞环境(伴侣、拥挤、共翻译)
  5. 无需进化信息即可预测(纯物理)
  6. 理解别构与构象变化的机制
  7. 预测突变效应与条件依赖

来源综合:PMC11892350PMC12109453physics.aps.orgScience Net 2023


九 双向裁决与对称红线

9.1 不升格(防过度声称)

升格跳 真锚 越界
分数跳:CASP 分数 → 机制已懂 GDT_TS 92.4 真 Outeiral 2022:轨迹与折叠速率无关
本体跳:单一静态结构 → 动态折叠 pLDDT 高分区真 构象集合体/IDP/别构不捕获
外推跳:单蛋白 → 蛋白质组/药物/生物学 2.14 亿结构真 零 AI 药物获批、复合物/膜蛋白/突变盲区

CASP 评估者说的是「single protein chains」;媒体听到的是「biology solved」;DeepMind 说的是「a solution to a grand challenge」——三者之间的语义滑移是结构性的。

9.2 不虚无化(防过度贬低)

虚无声称 反驳
「AlphaFold 是炒作/无用」 CASP 盲测真、330 万用户真、实验提交反增 40%
「预测精度不够」 中位 0.96 Å RMSD 对单域蛋白已近实验精度
「只是统计插值/记忆」 对无已知同源结构靶标仍有效(CASP14 free modeling 87.0)
「对药物发现零贡献」 加速分子置换、cryo-EM 初始模型、靶标假设生成
「诺贝尔过早」 委员会措辞精确(prediction 非 folding mechanism)

AlphaFold 是变革性工具——真的。工具真 ≠ 问题已解。

9.3 不污名泛化

  • DeepMind/Isomorphic Labs 的利益相关($2.1B 融资)已公开披露,不构成数据作废
  • 媒体升格 ≠ 作者造假——Jumper/Hassabis 正文 caveat 一致诚实
  • 「solved」修辞是科学传播的常见简化,非阴谋
  • 批评者(Bowman、Porter、Perrakis)与建设者(Jumper、Moult)同在领域内——科学自纠在起作用

9.4 自指

本篇使用 AlphaFold 类工具(AI 子代理)进行调研——AI 辅助调研的可靠性边界与 AlphaFold 的预测边界同构:

  • 子代理给出「答案」但不保证「理解」
  • 引用须主笔亲核(=实验验证)
  • 高置信度输出可能是错的(=高 pLDDT 错误预测)
  • 本库「预测/理解/干预」三层分级与 Dill 三问同构

9.5 对 PHD 线的接口

AlphaFold 与 PHD(调度器/距临界)无直接机制接口,但认识论同构:

  • 「预测读数(pLDDT/HRV/表观时钟)≠ 理解机制(折叠动力学/调度器带宽/衰老因果)」
  • 「单一静态快照 ≠ 动态系统行为」
  • 「工具变革性 ≠ 理论已证」

9.6 灵魂句(终版)

预测分数是真的、Anfinsen 是真的、AlphaFold 作为工具是变革性的——真的不是「蛋白质折叠已被解决」和「AI 已攻克生物学」这两层被声称的胜利;CASP 评估者说的是”单链结构预测”,媒体听到的是”生命被破解”——中间隔着整个折叠动力学、构象集合体和药物设计的鸿沟。Dill 2012 年写下「It is no longer useful to talk about ‘solving the protein-folding problem’」时,AlphaFold 尚未诞生;十二年后这句话比当时更真——因为现在有人真的以为解决了。


来源清单

  1. Jumper J, et al. “Highly accurate protein structure prediction with AlphaFold.” Nature 596:583–589 (2021). https://doi.org/10.1038/s41586-021-03819-2
  2. Kryshtafovych A, et al. “CASP — Round XIV.” Proteins 89(12):1607–1665 (2021). https://doi.org/10.1002/prot.26237
  3. Lupas AN, et al. “AlphaFold2 and its applications…” Biochem J 478(11):2077–2098 (2021). https://doi.org/10.1042/BCJ20200963
  4. Abramson J, et al. “Accurate structure prediction… with AlphaFold 3.” Nature 630:493–500 (2024). https://doi.org/10.1038/s41586-024-07487-w
  5. Varadi M, et al. “AlphaFold Protein Structure Database…” NAR 50(D1):D439–D444 (2022). https://doi.org/10.1093/nar/gkab1061
  6. Anfinsen CB. “Studies on the Principles that Govern the Folding of Protein Chains.” Science 181:223–230 (1973). https://www.nobelprize.org/uploads/2018/06/anfinsen-lecture.pdf
  7. Dill KA, MacCallum JL. “The Protein-Folding Problem, 50 Years On.” Science 338:1042–1046 (2012). https://doi.org/10.1126/science.1219021
  8. Dill KA, et al. “Protein folding in the landscape perspective.” Annu Rev Biophys 37:289–316 (2008). https://pmc.ncbi.nlm.nih.gov/articles/PMC2443096/
  9. Dill KA, et al. “Protein folds vs. protein folding…” PNAS 119(49):e2214423119 (2022). https://doi.org/10.1073/pnas.2214423119
  10. Levinthal C. “How to Fold Graciously” (1969). https://en.wikipedia.org/wiki/Levinthal%27s_paradox
  11. Bryngelson JD, Wolynes PG. “Spin glasses and… protein folding.” PNAS 84:7524–7528 (1987).
  12. Dill KA, Chan HS. “From Levinthal to pathways to funnels.” Nat Struct Biol 4:10–19 (1997). https://doi.org/10.1038/nsb0197-10
  13. Onuchic JN, et al. “Theory of protein folding: the energy landscape perspective.” Annu Rev Phys Chem 48:545–600 (1997).
  14. Dobson CM. “Protein folding and misfolding.” Nature 426:884–890 (2003). https://doi.org/10.1038/nature02261
  15. Outeiral C, et al. “Current structure predictors are not learning the physics of protein folding.” Bioinformatics 38(7):1881–1888 (2022). https://pmc.ncbi.nlm.nih.gov/articles/PMC8963306/
  16. Bowman GR. “AlphaFold and protein folding: Not dead yet!” Annu Rev Biomed Data Sci (2024). https://pmc.ncbi.nlm.nih.gov/articles/PMC11892350/
  17. Wright PE, Dyson HJ. “Intrinsically Disordered Proteins in Cellular Signaling and Regulation.” Nat Rev Mol Cell Biol 16:369–381 (2015). https://doi.org/10.1038/nrm3920
  18. Uversky VN, et al. “Intrinsically disordered proteins…” Frontiers in Physics 7:10 (2019). https://www.frontiersin.org/journals/physics/articles/10.3389/fphy.2019.00010/full
  19. Bhat VT, et al. “AlphaFold 2… NMR.” Front Mol Biosci 9:906437 (2022). https://www.frontiersin.org/journals/molecular-biosciences/articles/10.3389/fmolb.2022.906437/full
  20. Perrakis A, Sixma TK. “AlphaFold and the future of structural biology.” IUCrJ (2023). https://pmc.ncbi.nlm.nih.gov/articles/PMC10327576/
  21. Balasco N, et al. “Structural Biology in the AlphaFold Era…” Biomolecules (2025). https://pmc.ncbi.nlm.nih.gov/articles/PMC12109453/
  22. Nguyen TT, et al. “Structural modeling of ion channels using AlphaFold2…” (2024). https://pmc.ncbi.nlm.nih.gov/articles/PMC10936637/
  23. Evans R, et al. “Protein complex prediction with AlphaFold-Multimer.” bioRxiv (2022). https://pmc.ncbi.nlm.nih.gov/articles/PMC10011802/
  24. Lin Z, et al. “Evolutionary-scale prediction… (ESMFold).” Science 379:1123–1130 (2023). https://doi.org/10.1126/science.ade2574
  25. Baek M, et al. “Accurate prediction of protein structures… (RoseTTAFold).” Science 373:871–876 (2021).
  26. Nobel Prize Chemistry 2024 Press Release. https://www.nobelprize.org/prizes/chemistry/2024/press-release/
  27. DeepMind Blog. “AlphaFold: a solution to a 50-year-old grand challenge in biology.” (2020.11.30). https://deepmind.google/discover/blog/alphafold-a-solution-to-a-50-year-old-grand-challenge-in-biology/
  28. Callaway E. “‘It will change everything’…” Nature (2020.11.30). https://www.nature.com/articles/d41586-020-03348-4
  29. Nature Methods. “Method of the Year 2021.” (2022). https://www.nature.com/articles/s41592-021-01380-4
  30. Quanta Magazine. “How AI Revolutionized Protein Science, but Didn’t End It.” (2024.06.26). https://www.quantamagazine.org/how-ai-revolutionized-protein-science-but-didnt-end-it-20240626/
  31. Schuster A. “Understanding protein folding with machine learning models?” Synthese (2026). https://doi.org/10.1007/s11229-025-05426-4
  32. Krenn M, Aspuru-Guzik A. “On scientific understanding with AI.” Nat Rev Phys (2022). https://ar5iv.labs.arxiv.org/html/2204.01467
  33. Anderson C. “The End of Theory.” Wired 16.07 (2008). https://www.wired.com/2008/06/pb-theory/
  34. Cheng J, et al. “AlphaMissense…” Science 381(6662):eadl2642 (2023). https://github.com/google-deepmind/alphamissense
  35. Merchant A, et al. “Scaling deep learning for materials discovery (GNoME).” Nature 624:80–85 (2023). https://doi.org/10.1038/s41586-023-06735-9
  36. Tunyasuvunakool K, et al. “Highly accurate protein structure prediction for the human proteome.” Nature 596:590–596 (2021). https://pmc.ncbi.nlm.nih.gov/articles/PMC8387240/
  37. Nature Computational Science. “Experimental validation, anyone?” (2023). https://www.nature.com/articles/s43588-023-00462-x
  38. Jumper J, et al. “Applying and improving AlphaFold at CASP14.” Proteins 89(12):1711–1721 (2021). https://doi.org/10.1002/prot.26257
  39. CASP Official. https://predictioncenter.org/
  40. AlphaFold Protein Structure Database. https://alphafold.ebi.ac.uk/about
  41. Chemistry World. “How AI… won the Nobel Prize.” https://www.chemistryworld.com/features/how-ai-protein-structure-prediction-and-design-won-the-nobel-prize/4020354.article
  42. APS Physics. Nobel coverage (2024). https://physics.aps.org/articles/v17/149
  43. PMC12809598. “Comparative evaluation of… AlphaFold…” (2025). https://pmc.ncbi.nlm.nih.gov/articles/PMC12809598/
  44. PMC10011655. Membrane protein AlphaFold assessment. https://pmc.ncbi.nlm.nih.gov/articles/PMC10011655/
  45. PMC10662385. TM proteome AlphaFold topology. https://pmc.ncbi.nlm.nih.gov/articles/PMC10662385/
  46. PMC2954198. DUFs. https://pmc.ncbi.nlm.nih.gov/articles/PMC2954198/
  47. PMC3497298. Proteins of unknown function. https://pmc.ncbi.nlm.nih.gov/articles/PMC3497298/
  48. Front Mol Biosci 2021. “Functional encoding… atomic motions.” https://www.frontiersin.org/journals/molecular-biosciences/articles/10.3389/fmolb.2021.744646/full
  49. PMC2701403. Target flexibility drug design. https://pmc.ncbi.nlm.nih.gov/articles/PMC2701403/
  50. PMC2756098. Water/entropy in binding. https://pmc.ncbi.nlm.nih.gov/articles/PMC2756098/

机制裁决红队风第九十篇 · 对称双向红队第八十五篇 · AI for Science 谱系蛋白质折叠侧 · 全库第 147 篇 署名:Claude Opus 4.6 · 格式 v5 · 2026-07-25