从 GAN 到扩散：AI 在蛋白质建模中的演变

From GANs to Diffusion: AI's Evolution in Protein Modeling

谢尔盖·叶杜诺夫 Sergey Edunov · Latent Space · 2026-06-30 · 约 109 分钟 · 原视频 ↗

打开互动全文版（中英对照 + 朗读 + 问答）→

本期速览 · Overview

Evan Fineberg 和 Sergey Udov 讨论从 GAN 到扩散模型在蛋白质结构预测中的转变，他们的物理学背景，以及 Genesis Molecular AI 改进 AI 用于药物发现的方法。

Evan Fineberg and Sergey Udov discuss the shift from GANs to diffusion models for protein structure prediction, their physics backgrounds, and Genesis Molecular AI's approach to improving AI for drug discovery.

要点 · TL;DR

• 亚埃级精度对 AI 驱动的药物发现至关重要，而不仅仅是 RMSD 小于 2 埃。
  Sub-angstrom accuracy is essential for AI-driven drug discovery, not just RMSD < 2Å.
• 药物发现 AI 的扩展与 LLM 扩展类似：预训练、后训练和推理时扩展都很重要。
  Scaling in drug discovery AI mirrors LLM scaling: pre-training, post-training, and inference-time scaling.
• 来自物理模拟的合成数据可增强有限的实验数据用于训练。
  Synthetic data from physics simulations augments limited experimental data for training.

核心观点 · Key points

• 药物发现需要高分辨率 3D 结构预测，亚埃级精度对有用的结合姿态预测至关重要。
  Drug discovery requires high-resolution 3D structure prediction, with sub-angstrom accuracy for useful binding pose predictions.
• 药物发现 AI 的 Scaling（规模扩张）模仿 LLM 的 Scaling（规模扩张）：预训练、后训练和推理时 Scaling（规模扩张）都很重要。
  Scaling in drug discovery AI mirrors LLM scaling: pre-training, post-training, and inference-time scaling all matter.
• 基于物理的模拟生成的合成数据是扩充有限实验数据以训练模型的关键。
  Synthetic data from physics-based simulations is key to augment limited experimental data for training models.
• 智能体协调多个 AI 工具以自动化药物发现，使药物化学家更高效。
  Agents orchestrate multiple AI tools to automate drug discovery, making medicinal chemists more efficient.
• 与制药和湿实验室合作伙伴的关系实现了快速的设计-制造-测试-分析循环，以持续改进模型。
  Partnerships with pharma and wet-lab collaborators enable rapid design-make-test-analyze cycles for continuous model improvement.

反共识 · Contrarian takes

• AlphaFold 并未解决药物发现；静态结构不足以预测结合和动态。
  AlphaFold did not solve drug discovery; static structures are insufficient for predicting binding and dynamics.
• 常见的基准 RMSD < 2Å是不够的；实际药物设计需要亚埃级精度。
  Common benchmark RMSD < 2Å is insufficient; sub-angstrom accuracy is needed for real drug design utility.
• 自动化合成实验室在处理新颖分子时遇到困难；人类专业知识对前沿化学仍然至关重要。
  Automated synthesis labs struggle with novel molecules; human expertise remains crucial for cutting-edge chemistry.
• GPU 短缺是药物发现 AI 的主要瓶颈，而不仅仅是 LLM。
  GPU shortage is a major bottleneck for drug discovery AI, not just for LLMs.
• LLM 架构相对乏味；药物发现模型提供了更多样化和有趣的架构挑战。
  LLM architectures are relatively boring; drug discovery models offer more diverse and interesting architectural challenges.

本期章节 · Chapters（共 38）

• 引言与背景 Introduction and Backgrounds
• 分子机器学习：进展与挑战 Machine Learning for Molecules: Progress and Challenges
• 十年前蛋白小分子药物发现 State of protein-small molecule drug discovery a decade ago
• Pearl简介与核心洞见 Introduction to Pearl and its key insights
• 小分子搜索如大海捞针 Small molecule search as needle in haystack
• 训练数据挑战与合成数据 Training data challenges and synthetic data
• 路线图：类比LLM的扩展 Roadmap: scaling analogy with LLMs
• 推理时扩展：扩散头与物理引导 Inference time scaling: diffusion head and physics guidance
• 领域先验与数据挑战 Domain-specific priors and data challenges
• 聚焦中小分子发现 Focus on small and medium molecule discovery
• 药物发现阶段概览 Overview of drug discovery stages
• 精准医疗与不可成药靶点 Precision medicine and undruggable targets
• 低垂果实之外的机会 Opportunity beyond low-hanging fruit
• 更广的药物发现与AI角色 Broader drug discovery and AI's role
• 药物发现尚未解决 Drug discovery not solved
• 管线与合作伙伴关系 Pipeline and partnerships
• 一埃阈值 One angstrom threshold
• 法医类比与模型验证 Forensics analogy and model validation
• 姿态作为抽象 Pose as an abstraction
• 超越姿态：熵及其他贡献 Beyond pose: entropic and other contributions
• 高分辨率对药物发现的重要性 Why high resolution matters for drug discovery
• 评估危机与领域转型 Eval crisis and transition in the field
• 从势网到生成建模的演进 Evolution from potential net to generative modeling
• ADMET性质及其挑战 ADMET properties and their challenges
• ADMET早期工作与MoleculeNet Early work on ADMET and MoleculeNet
• 未来出版物与OpenBind结果 Future publications and OpenBind results
• 计算、ML与湿实验数据整合 Integration of computation, ML, and wet lab data
• Genesis与Insight合作 Partnership between Genesis and Insight
• 分子优化挑战 Challenges in Molecular Optimization
• 制药合作与AI转变 Pharma partnerships and AI shift
• 转向智能体系统 Shift to agentic systems
• 人机交互与扩展 Human-agent interaction and scaling
• 智能体可靠性与人类监督 Agent reliability and human oversight
• Pearl模型在公开基准上的结果 Pearl model results on open benchmark
• Pearl处理动态靶点的原因 Why Pearl handles dynamic targets
• 基准性能与现实价值 Benchmark Performance and Real-World Value
• 瓶颈：GPU短缺 Bottleneck: GPU Shortage
• 行动号召：招聘与独特架构工作 Call to Action: Hiring and Unique Architecture Work

阅读全文双语转录 →