科学数据视角下AlphaFold的迭代突破与数据策略启示

doi:10.19788/j.issn.2096-6369.000136

Abstract

Abstract:

The transformative breakthroughs of the AlphaFold series in structural biology are often attributed to algorithmic advances, yet the critical role of its evolving data strategy remains underexplored. Adopting a data-centric perspective, this paper deconstructs the iterative mechanisms driving AlphaFold’s progress from versions 1 to 3, emphasizing the optimization of data quality attributes, innovations in representation paradigms, and data-model synergy. The analysis reveals that each performance leap stems from the co-evolution of data and model architectures. AlphaFold’s data strategy follows a clear trajectory: from passive data adoption, to proactive data construction, and finally to generative data augmentation. From this, three core principles emerge: paradigm shifts in data representation are the primary drivers of breakthroughs; data-model co-evolution is a hallmark of system maturity; and the richness of data quality attributes sets the ceiling for an AI’s learning potential. These principles yield four implications for the AI for Science (AI4S) field: data practices should shift from passive preparation to active design; research should prioritize data-model alignment over model- or data-centric approaches; data ecosystems should focus on enhancing key attributes, such as diversity and quality, rather than broad multimodal integration; and a new theoretical and evaluation framework is needed to assess the "scientific efficacy" of data. This study provides a theoretical foundation and practical roadmap for advancing AI-driven scientific discovery.

Key words: AlphaFold, scientific data, data-model synergy, protein structure prediction, AI for science

OUYANG ZhengZheng, MA YuCong, KOU YuanTao, XIAN GuoJian, WANG Hui, ZHAO Qun. Unveiling AlphaFold’s Iterative Breakthroughs: Data Strategy Insights from a Scientific Perspective[J].Journal of Agricultural Big Data, 2025, 7(4): 485-495.

Figures/Tables 4

Table 1

Comparison of core features of AlphaFold Generations 1-3"

维度	AlphaFold 1（2018）^[22]	AlphaFold 2（2020）^[23]	AlphaFold 3（2024）^[24]
研发背景与核心目标	针对传统自由建模（Free Modeling，FM）方法精度有限的问题，验证深度学习在蛋白质结构预测领域的可行性^[25]，重点突破无模板蛋白预测的技术空白。	聚焦AlphaFold 1对长链蛋白、稀有折叠预测的局限，旨在从根本上解决单体蛋白质的高精度结构预测问题，推动AI技术向科研应用落地。	突破AlphaFold 2仅能预测单链蛋白的边界，将目标拓展至生物分子相互作用体系，实现蛋白质-核酸、蛋白质-小分子等复合物的精准预测。
关键成果	1. 在CASP13竞赛中，在‌模板基建模（Template-based Modeling，TBM）的部分目标上实现90 GDT以上的原子级精度，且FM域平均性能显著超越传统方法，实现该领域首次突破； 2. 同时验证MSA与深度学习结合的核心技术路线可行性，但其FM域无模板预测成功率仅56%。	1. CASP14竞赛中90%目标蛋白主链预测精度媲美实验解析（平均GDT 92.4分）^[26]； 2. 支持2180残基长链蛋白（如T1044）完整预测，无需分块； 3. 催生AlphaFold DB，截至2022年收录超2亿条蛋白质结构^[27]。	1. CASP15蛋白质-核酸复合物盲测中95%目标达到实验精度，蛋白质-配体结合预测精度相比AlphaFold 2有大幅提升，在PoseBusters等关键基准测试上表现尤为突出； 2. 实现修饰残基（糖基化、磷酸化）与预测不同功能状态下的构象。
核心技术局限	1. 严重依赖MSA数据质量，对无同源序列蛋白预测精度显著下降； 2. 采用静态距离矩阵表征，长程相互作用（>22Å）预测误差大； 3. 无法端到端输出完整结构，长链蛋白需分块处理。	1. 仅支持单链蛋白质预测，无法处理蛋白质复合物、核酸-蛋白质互作体系^[28]； 2. 无法解析蛋白质动态构象变化（如变构效应）； 3. 元数据未完全遵循FAIR原则，跨平台复现效率受限。	1. 对MSA缺失率>50%的极端场景（如全新化学空间小分子）预测鲁棒性不足； 2. 仅能预测单一静态结构，无法捕捉分子动态过程，且在预测小分子时无法保证正确的手性^[29]； 3. 跨数据库（PDB、UniProt）数据标准化不统一，多模态数据融合效率低。
代表性应用场景	基础研究领域的蛋白质结构初步预测，为后续AI模型提供技术验证样本^[30]。	结构生物学常规研究（如未知单体蛋白结构解析）、AlphaFold DB支撑的跨物种蛋白功能注释。	药物研发（靶点-配体结合模式解析）、生物分子互作机制研究（如病毒-宿主蛋白复合物解析）。

Table 1

Table 2

Core data analysis dimensions of AlphaFold’s iterative breakthroughs"

核心维度	核心内涵	关键子维度	AlphaFold中的典型例证
数据内在属性	聚焦数据作为人工智能系统输入的基础特性，直接构成模型性能的理论潜力约束，即模型性能无法超越数据基础属性所支撑的上限。	质量与可靠性	AlphaFold 2的自蒸馏技术^[35]：通过模型自我预测，生成海量高质量的伪标签结构数据，提升标注可靠性。
		规模与覆盖度	AlphaFold 2整合元基因组数据库^[36]：将数据源从PDB拓展至BFD/MGnify，极大增加了序列的多样性与覆盖广度。
		信息完备性	AlphaFold 3整合多模态数据：将数据从单一蛋白质拓展至包含核酸、配体、离子等的复合物，信息更完整。
数据表征范式	聚焦数据向人工智能模型可解释语义的映射与编码逻辑，其决定了数据蕴含的结构与功能信息向模型的传递效率及价值激活的有效性。	信息维度与抽象层次	AlphaFold 1→AlphaFold 2的革新：从静态二维距离矩阵（信息降维）跃迁至动态三维交互图（高维抽象）。
数据表征范式	聚焦数据向人工智能模型可解释语义的映射与编码逻辑，其决定了数据蕴含的结构与功能信息向模型的传递效率及价值激活的有效性。	物理本质贴合度	AlphaFold 2/AlphaFold 3的图表征：直接建模残基间的空间关系与相互作用，比AlphaFold 1的矩阵表征更贴近蛋白质折叠的物理本质。
数据端主动调控	聚焦为适配特定模型架构与任务需求，通过突破原始数据的固有局限，决定了数据价值的可释放上限与模型性能的实际达成度。	数据预处理与标准化	AlphaFold 2的MSA去冗余策略：通过Neff等指标过滤，提升输入信噪比。AlphaFold 3的跨库元数据标准化：尝试遵循FAIR原则，提升多模态数据融合效率。
数据端主动调控	聚焦为适配特定模型架构与任务需求，通过突破原始数据的固有局限，决定了数据价值的可释放上限与模型性能的实际达成度。	领域知识嵌入	AlphaFold 2的FAPE（Frame-Aligned Point Error）损失函数：将“刚体变换不变性”这一物理约束，通过训练目标函数的形式显式嵌入。AlphaFold 3的生成过程引导：将AMBER力场等物理化学规则，作为扩散模型生成过程的外部约束和引导。

Table 2

Table3

中文全称	英文全称/缩写	解释说明
多序列比对	Multiple sequence alignment, MSA	多重序列比对是对三个以上生物序列（这里指蛋白质）进行生物学比对的方法。
自由建模	‌Free modeling, FM	不依赖同源模板，直接通过序列信息和物理化学规律进行结构预测。
模板基建模	Template-based modeling, TBM	通过比对已知同源结构进行建模。
数据质量评估	Data quality assessment, DQA	对数据的核心特征进行检验、量化与判定的过程，核心是判断数据是否符合使用场景的需求。
表征学习	Representation learning, RL	通过自动学习数据的高效表示形式来提升机器学习效果。
静态二维距离分布	Distogram	蛋白质结构预测中用于表示残基间距离关系的概率分布模型，通过将连续距离离散化为有限区间，捕捉空间约束关系。
深度残差卷积网络	ResNet	一种通过引入残差连接解决深层网络训练难题的卷积神经网络架构。
自蒸馏技术	Self-distillation, SD	是知识蒸馏的一种特殊形式，指模型通过自我学习提升性能的技术，无需依赖外部教师模型。
框架对齐点误差	Frame-aligned point error, FAPE	一种损失函数，通过局部刚体变换对齐后量化预测与真实原子坐标的差异，具有全局旋转和平移不变性。
卷积神经网络	Convolutional neural network, CNN	是一种前馈神经网络，广泛应用于图像识别和视觉任务，是深度学习中的核心模型之一。
关联性学习	Associative learning, AL	一种通过神经元间的联系增强或减弱实现的学习过程，其核心是通过重复刺激使神经元之间的突触连接强度发生变化。
生成式学习	Generative learning, GL	是机器学习领域的一种方法，通过训练数据规律生成符合数据分布的新实例。
翻译后修饰	Post-translational modifications, PTMs	指蛋白质在合成完成后，通过共价添加化学基团或蛋白质间的化学交互作用改变其结构和功能的过程。
扩散生成模型	Diffusion model	是一种基于概率统计和非平衡热力学原理的生成式人工智能模型，通过模拟数据逐步添加噪声的过程实现样本生成。其核心包括前向扩散（向数据添加噪声）和逆向生成（从噪声中恢复原始数据）两个阶段。
FAIR原则	FAIR	科学数据治理的核心原则，分别指数据可通过元数据发现（Findability）、开放可访问（Accessibility）、跨系统互操作（Interoperability）、可重复利用（Reusability）。
大规模综合序列数据库	Big fantastic database, BFD	大型蛋白质序列数据库，收录数十亿级原始序列，经聚类去冗余后供生物序列分析使用。
宏基因组数据库	MGnify	宏基因组领域专用数据库，整合微生物组序列数据、组装结果及功能注释信息。

References 37

[1]	BAI X C, MCMULLAN G, SCHERES S H. How cryo-EM is revolutionizing structural biology. Trends in Biochemical Sciences, 2015, 40(1):49-57. doi: 10.1016/j.tibs.2014.10.005
[2]	RCSB PDB. 2024 RCSB PDB Advisory Committee Meeting,2024[R]. https://cdn.rcsb.org/rcsb-pdb/general_information/about_pdb/rcsbpdbac24-presentations.pdf.
[3]	LYUMKIS D. Challenges and opportunities in cryo-EM single- particle analysis. Journal of Biological Chemistry, 2019, 294(13): 5181-5197. doi: 10.1074/jbc.REV118.005602
[4]	GAO T, DAMBORSKY J, JANIN Y L, et al. Deciphering enzyme mechanisms with engineered ancestors and substrate analogues. Chemical Catalysis and Chemistry, 2023, 15(19):e202300745.
[5]	NAKANE T, KOTECHA A, SENTE A, et al. Single-particle cryo-EM at atomic resolution. Nature, 2020, 587(7832):152-156. doi: 10.1038/s41586-020-2829-0
[6]	HO C M, LI X R, MASON L, et al. Bottom-up structural proteomics: cryoEM of protein complexes enriched from the cellular milieu. Nature Methods, 2020, 17:79-85. doi: 10.1038/s41592-019-0637-y
[7]	SANJYOT V S, DEEPTARUP B, ARTHUR Z, et al. AlphaCross-XL: A seamless tool for automated and proteome-scale mapping of crosslinked peptides onto three-dimensional protein structures. Molecular&Cellular Proteomics, 2025, 24(17):101057.
[8]	SENIOR A W, EVANS R, JUMPER J, et al. Improved protein structure prediction using potentials from deep learning. Nature, 2020, 577(7792):706-710. doi: 10.1038/s41586-019-1923-7
[9]	JUMPER J, EVANS R, PRITZEL A, et al. Highly accurate protein structure prediction with AlphaFold. Nature, 2021, 596(7873): 583-589. doi: 10.1038/s41586-021-03819-2
[10]	VARADI M, ANYANGO S, DESHPANDE M, et al. AlphaFold DB: A comprehensive database of protein structures predicted using AlphaFold. Nucleic Acids Research, 2022, 50(D1):D439-D444.
[11]	ABRAMSON J, EVANS R, PRITZEL A, et al. AlphaFold 3: High- accuracy structure prediction for complex molecular systems. Nature, 2024, 625(7996):479-488.
[12]	BOUATTA N, SORGER P K. AI-driven structural biology: from predictions to paradigm shifts. Nature Reviews Molecular Cell Biology, 2022, 23(12):745-759.
[13]	孙坦, 张智雄, 周力虹, 等. 人工智能驱动的第五科研范式(AI4S)变革与观察. 农业图书情报学报, 2023, 35(10):4-32. doi: 10.13998/jcnki.issn1002-1248.23-0850
	SUN T, ZHANG Z X, ZHOU L H, et al. The transformation and observations of AI for science(AI4S) driven by artificial intelligence. Journal of Library and Information Science in Agriculture, 2023, 35(10):4-32. doi: 10.13998/jcnki.issn1002-1248.23-0850
[14]	中国科学技术信息研究所. AI for Science创新图谱, 2025[R].
	Institute of Scientific and Technical Information of China. AI for Science Innovation Map, 2025[R].
[15]	YANG Z Y, ZENG X X, ZHAO Y, et al. AlphaFold2 and its applications in the fields of biology and medicine. Signal Transduction and Targeted Therapy, 2023, 8(1):115. doi: 10.1038/s41392-023-01381-z pmid: 36918529
[16]	GUO F, GUAN R C, LI Y H, et al. Foundation models in bioinformatics. National Science Review, 2025, 12(4):nwaf028.
[17]	FAN Z G, YANG Y D, XU M Y, et al. EC-Conf: A Ultra-fast diffusion model for molecular conformation generation with equivariant consistency. Journal of Cheminformatics, 2024, 16(1):107. doi: 10.1186/s13321-024-00893-2 pmid: 39228003
[18]	XU M, YU L, SONG Y, et al. Geodiff: A geometric diffusion model for molecular conformation generation[EB/OL].[2022-03-06]. https://arxiv.org/abs/2203.02923.
[19]	BAEK M, DIMAIO F, ANISHCHENKO I, et al. Accurate prediction of protein structures and interactions using a three-track neural network. Science, 2021, 373(6557):871-876. doi: 10.1126/science.abj8754 pmid: 34282049
[20]	ROSIGNOLI S, PACELLI M, MANGANIELLO F, et al. An outlook on structural biology after Alpha Fold: tools, limits and perspectives. FEBS Open Bio, 2025, 15(2):202-222. doi: 10.1002/feb4.v15.2
[21]	KRISHNA R, WANG J, AHERN W, et al. Generalized biomolecular modeling and design with RoseTTAFold All-Atom. Science, 2024, 384(6693):eadl2528.
[22]	KRYSHTAFOVYCH A, SCHWEDE T, TOPF M, et al. Critical assessment of methods of protein structure prediction (CASP)-round XIII. Proteins, 2019, 87(12):1011-1120. doi: 10.1002/prot.v87.12
[23]	KRYSHTAFOVYCH A, SCHWEDE T, TOPF M, et al. Critical assessment of methods of protein structure prediction (CASP)—Round XIV. Proteins: Structure, Function, and Bioinformatics, 2021, 89(12): 1607-1617. doi: 10.1002/prot.v89.12
[24]	ABRAMSON J, ADLER J, DUNGER J, et al. Accurate structure prediction of biomolecular interactions with AlphaFold3. Nature, 2024, 630:493-500. doi: 10.1038/s41586-024-07487-w
[25]	MCMASTER B, THORPE C, OGG G, et al. Can AlphaFold’s breakthrough in protein structure help decode the fundamental principles of adaptive cellular immunity?. Nature Method, 2024, 21:766-776. doi: 10.1038/s41592-024-02240-7
[26]	JUMPER J, EVANS R, PRITZEL A, et al. Applying and improving AlphaFold at CASP14. Proteins: Structure, Function, and Bioinformatics, 2021, 89:1711-1721. doi: 10.1002/prot.v89.12
[27]	VARADI M, BERTONI D, MAGANA P, et al. AlphaFold protein structure database in 2024: providing structure coverage for over 214 million protein sequences. Nucleic Acids Research, 2024,52(D1): D368-D375.
[28]	LAURENTS D V. AlphaFold 2 and NMR spectroscopy: partners to understand protein structure, dynamics and function. Frontiers in Molecular Biosciences, 2022, 17(9):906437.
[29]	FANG Z, RAN H, ZHANG Y, et al. AlphaFold 3: an unprecedent opportunity for fundamental research and drug development. Precision Clinical Medicine, 2025, 8(3): pbaf015.
[30]	GOWTHAMAN R, PARK M, YIN R, et al. AlphaFold and docking approaches for antibody-antigen and other targets: insights from CAPRI rounds 47-55. Proteins: Structure, Function, and Bioinformatics, 2025, 0:1-6. DOI: 10.1002/prot.26801.
[31]	DAMA International. DAMA-DMBOK: Data management body of knowledge (2nd ed.),2017[R]. Technics Publications.
[32]	LECUN Y, BENGIO Y, HINTON G. Deep learning. Nature, 2015, 521(7553):436-444. doi: 10.1038/nature14539
[33]	AI Business. ScaleUp AI 2022:Google brain’s Andrew Ng says data-centric approach ups AI success[EB/OL].[2022-04-06]. https://aibusiness.com/companies/scaleup-ai-2022-google-brain-s-andrew-ng-says-data-centric-approach-ups-ai-success.
[34]	BERTOLINE L M F, LIMA A N, KRIEGER, J E, et al. Before and after AlphaFold2: An overview of protein structure prediction. Frontiers in bioinformatics, 2023, 3, 1120370. doi: 10.3389/fbinf.2023.1120370
[35]	NIAZI S K, MARIAM Z, PARACHA R Z. Limitations of protein structure prediction algorithms in therapeutic protein development. BioMedInformatics, 2024, 4:98-112. doi: 10.3390/biomedinformatics4010007
[36]	YUAN Q M, CHEN S, RAO J H, et al. AlphaFold2-aware protein- DNA binding site prediction using graph transformer. Briefings in Bioinformatics, 2022, 23(2):bbab564.
[37]	REMMERT M, BIEGERT A, HAUSER A, et al. HHblits: lightning- fast iterative protein sequence searching by HMM-HMM alignment. Nature Methods, 2012, 9(2): 173-175. doi: 10.1038/nmeth.1818

核心维度	关键子维度	AlphaFold 1 (2018) 被动沿用范式（单链结构的统计学习解析）	AlphaFold 2 (2020) 主动构建范式（分子几何的物理建模）	AlphaFold 3 (2024) 生成赋能范式（互作体系的生成式探索）
数据内在属性	质量与可靠性	被动依赖：基于PDB既有标注质量。	主动创造：通过自蒸馏生成高质量伪标签。	功能导向的验证：引入PoseBusters等功能性基准。
	规模与覆盖度	基础存量拓展：依赖核心蛋白质数据库。	序列空间穷举：战略性整合元基因组数据。	多模态全景：跨域整合异构生物分子。
	信息完备性	孤立的静态结构：仅包含单体静态信息。	上下文缺失的静态结构：仍为单体，互作局限性凸显。	功能性的互作网络：整合复合物与功能状态信息。
数据表征范式	信息维度与抽象层次	统计分布表征：对残基距离的离散化概率建模。	确定性几何表征：对三维空间刚体框架的显示建模。	生成式概率表征：对原子坐标点云的概率分布学习。
数据表征范式	物理本质贴合度	信息降维：三维物理问题的二维压缩。	三维几何约束：基于刚体不变性的几何拓扑建模。	无偏见的原子生成：从最基本物理单元出发的生成式建模。
数据端主动调控	数据预处理与标准化	固定流水线：依赖标准化的外部工具链。	自适应优化：引入复杂的内部MSA去冗余策略。	FAIR数据生态：遵循FAIR原则的多模态数据治理与整合。
数据端主动调控	领域知识嵌入	隐式统计关联：知识被动包含于MSA中。	显式几何约束：以“刚体变换不变性”作为训练硬约束。	动态物理引导：多尺度物理知识对生成过程的软引导。

Unveiling AlphaFold’s Iterative Breakthroughs: Data Strategy Insights from a Scientific Perspective

RichHTML

PDF (PC)

Like

Knowledge

Abstract

Cite this article

share this article

Figures/Tables 4

References 37

Related Articles 15

Metrics

Comments

Recommended 0

[1]	HE HaoYu, HOU ChunMei, SUN LiWei, CHI XiuLi, YE XiYan. A Comparative Analysis of Multidisciplinary General-Purpose Scientific Data Platforms: Taking Zenodo and ScienceDB as Examples [J]. Journal of Agricultural Big Data, 2025, 7(2): 193-200.
[2]	DU JiaoYan, ZHANG HongYu, LI AnDing, CAO Yang, CAI GuoJun. A Dataset on Leaf Functional Traits of a Pioneer Species (Pyracantha fortuneana) in the Karst Region of Guizhou, Southwest China [J]. Journal of Agricultural Big Data, 2025, 7(2): 246-245.
[3]	SHEN QiHui. Review and Suggestion of National Scientific Data Collection in the New Era from the Perspective of Data Factionalization [J]. Journal of Agricultural Big Data, 2024, 6(3): 363-372.
[4]	WANG YueYue, CHEN ZuGang, WU XinQian. Knowledge Graph Analysis of Scientific Data Centers Evaluation in China [J]. Journal of Agricultural Big Data, 2024, 6(3): 373-379.
[5]	WANG Jian, ZHOU GuoMin, LIAO FangYu, XU ZhePing, ZHANG JianHua, LIU TingTing. Rule-Based Framework for Scientific Data Security Governance: A New Tool for Understanding the Imbalance and Challenges of Data Protection and Utilization [J]. Journal of Agricultural Big Data, 2024, 6(3): 295-306.
[6]	WANG Jian, ZHOU GuoMin, ZHANG JianHua, XU ZhePing, LIU TingTing. Navigating the Distinctiveness of Research Data Protection: Framework and Mode [J]. Journal of Agricultural Big Data, 2024, 6(3): 307-324.
[7]	LIAO FangYu, LI Jing, LONG Chun, YANG Fan, YUAN ZiMeng. Security Challenges and Countermeasures on Open Sharing of Scientific Data in the Context of Open Science [J]. Journal of Agricultural Big Data, 2024, 6(2): 146-155.
[8]	WEI Xin, KONG LiHua, WANG Yang. Research on Countermeasures of Scientific Data Exit Management in China [J]. Journal of Agricultural Big Data, 2024, 6(2): 156-160.
[9]	LI YiZhan, DONG Lu, WANG DongYao, ZHANG Hong, WANG ZhiQiang, WEI Ren, LI ZeXia. Governance of Scientific Data Sharing within International Organizations and International Science and Technology Project [J]. Journal of Agricultural Big Data, 2024, 6(2): 161-169.
[10]	WANG JuanLe, LI Kai, DUAN BoWen, SU Na. Analysis of Autonomy in Geosciences Data Processing and Analysis Software [J]. Journal of Agricultural Big Data, 2024, 6(2): 230-240.
[11]	XU Qi, HU XiaoYan, ZOU ZiMing, TONG JiZhou. Research on the Security Classification Conceptual Framework of Space Environment Scientific Data [J]. Journal of Agricultural Big Data, 2024, 6(2): 259-268.
[12]	WANG JiaRong, ZHOU CaiQiu, YUAN XinYang, ZHU YiMing, SUN QianRan, YAN Tian, CHEN Gang, QI FaZhi. National High Energy Physics Science Data Security System [J]. Journal of Agricultural Big Data, 2024, 6(2): 269-277.
[13]	FU Yu, JIANG XiaoYi, WEI YangMing, TONG Xing, XU MoGen, WANG Yi. Design and Practice of Marine Scientific Data Security Governance System [J]. Journal of Agricultural Big Data, 2024, 6(2): 286-294.
[14]	ZHANG LiXia, LIU Qi, LI ShaoTing, REN YanJun. A Dataset of the Statistics on Food Consumption and Nutrition- related Health of Rural Residents in Shaanxi Rural Areas in 2022 [J]. Journal of Agricultural Big Data, 2023, 5(4): 118-124.
[15]	WANG ZiYu, JI YueQing, ZHOU Li. Statistical Dataset of Land Economic Survey in Jiangsu Province, China [J]. Journal of Agricultural Big Data, 2023, 5(4): 138-143.