诺贝尔奖得主、AlphaFold 负责人 John Jumper 探讨蛋白质结构预测的突破、对生物学的影响以及他转投 Anthropic 的决定。
John Jumper, Nobel laureate and lead of AlphaFold, discusses the breakthrough in protein structure prediction, its impact on biology, and his move to Anthropic.
要点 · TL;DR
AlphaFold 在几分钟内预测蛋白质结构，彻底改变了结构生物学。 AlphaFold predicts protein structures in minutes, revolutionizing structural biology.
AlphaFold 是特定实验的狭义预测器，而非全细胞模型。 AlphaFold is a narrow predictor of a specific experiment, not a whole-cell model.
AlphaFold 2 的成功来自许多中等规模的胜利，而非一两个突破。 AlphaFold 2's success came from many mid-size wins, not one or two breakthroughs.
核心观点 · Key points
AlphaFold 在几分钟内预测蛋白质结构，取代了以往需要数年的实验，彻底改变了结构生物学。 AlphaFold predicts protein structures in minutes instead of years, revolutionizing structural biology.
AlphaFold 是特定实验的窄预测器，而非整个细胞的模型。 AlphaFold is a narrow predictor of a specific experiment, not a model of the entire cell.
AlphaFold 2 的成功来自众多中等规模的胜利，而非一两个突破。 The success of AlphaFold 2 came from many mid-size wins, not one or two breakthroughs.
机器学习成功需要十次中九次犯错，并建立局部直觉。 Machine learning success requires being wrong nine times out of ten and building local intuition.
AlphaFold 使资源有限地区的科学家能够进行以前不可能完成的复杂实验。 AlphaFold enables scientists in resource-limited settings to do complex experiments previously impossible.
反共识 · Contrarian takes
AlphaFold 2 与苦涩教训相反；专门的架构是关键，而不仅仅是 Scaling（规模扩张）。 AlphaFold 2 is the opposite of the bitter lesson; specialized architecture was key, not just scaling.
等变性对 AlphaFold 2 的改进仅贡献了 30 分中的 2.5 分，却被过度炒作。 Equivariance contributed only 2.5 out of 30 points to AlphaFold 2's improvement, yet is overhyped.
从 AlphaFold 2 中移除卷积层提高了准确性，这与典型的机器学习预期相反。 Removing convolutional layers from AlphaFold 2 improved accuracy, contrary to typical ML expectations.
AlphaFold 3 的扩散模型首先解决大规模结构，与图像扩散相反。 AlphaFold 3's diffusion model solves large-scale structure first, opposite to image diffusion.
模型并未实现理解；它提供预测和控制，将理解留给人类。 Understanding is not achieved by the model; it provides prediction and control, leaving understanding to humans.
本期章节 · Chapters（共 18）
AlphaFold 的突破与影响AlphaFold's breakthrough and impact
AlphaFold 与蛋白质结构预测简介Introduction to AlphaFold and protein structure prediction
AlphaFold 的影响与应用Impact and applications of AlphaFold
AlphaFold 在理解生物学中的作用AlphaFold's role in understanding biology
从 AlphaFold 2 到 AlphaFold 3 与药物设计From AlphaFold 2 to AlphaFold 3 and drug design
狭义与通用生物学机器Narrow vs. Universal Biology Machine
AlphaFold 架构演进：CNN 到扩散AlphaFold Architecture Evolution: CNN to Diffusion
AlphaFold 2：EvoFormer 与结构模块AlphaFold 2: EvoFormer and Structure Module
AlphaFold 2 中的等变性Equivariance in AlphaFold 2
AlphaFold 的启示：架构与数据Lessons from AlphaFold: Architecture vs Data
架构修改与迭代过程Architecture modifications and iterative process
AlphaFold 3 扩散与图像扩散AlphaFold 3 diffusion vs image diffusion
细节胜于高层标签Importance of details over high-level labels
建设性复杂性与 AGIConstructive complexity and AGI
表征的重要性Importance of representations
通过下一词预测构建智能On building intelligence through next-token prediction
闭幕致辞与 Emmanuel 介绍Closing remarks and introduction of Emmanuel
主持人对 AlphaFold 与混合模型的总结Host's final thoughts on AlphaFold and hybrid models