ai-scientific-discovery-jumper

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

AI for Scientific Discovery

用于科学发现的 AI

Principles and workflows for building AI systems that accelerate scientific research, derived from the development of AlphaFold at DeepMind.

源自 DeepMind 开发 AlphaFold 的经验，介绍用于加速科学研究的 AI 系统原则与工作流。

Core Principles

核心原则

Research Ideas as Multipliers

研究创意作为倍增器

Research and architectural innovation multiply the value of data and compute rather than adding linearly.

Key insight: AlphaFold2 trained on 1% of available data outperformed AlphaFold1 trained on 100% of data. This demonstrates that ideas can provide 100x leverage over data alone.

Implication: When building ML systems for scientific domains, prioritize research iteration over data acquisition. Most compute budget should account for failed research iterations, not final model training.

研究和架构创新对数据与算力的价值产生乘法效应，而非线性叠加。

关键洞察：使用 1% 可用数据训练的 AlphaFold2，表现优于使用 100% 数据训练的 AlphaFold1。这表明，相较于单纯增加数据，创意可以提供 100 倍的杠杆效应。

影响：为科学领域构建 ML 系统时，应优先开展研究迭代，而非获取数据。大部分算力预算应用于可能失败的研究迭代，而非最终模型训练。

Many Mid-Scale Ideas

大量中等规模的创意

Breakthroughs come from accumulating many medium-sized innovations, not one revolutionary idea.

Key insight: In AlphaFold ablation studies, no single component explained more than 10% of the system's improvement. The breakthrough emerged from combining numerous architectural innovations.

Implication: Avoid searching for a single "magic bullet" solution. Instead, systematically develop and integrate multiple complementary improvements.

突破来自许多中等规模创新的累积，而非某个革命性创意。

关键洞察：在 AlphaFold 的消融研究中，没有任何单一组件贡献了系统提升的 10% 以上。突破源于多项架构创新的组合。

影响：不要寻找单个“银弹”方案，而应系统地开发并整合多项互补改进。

Biological Relevance Threshold

生物学相关性阈值

Real-world impact requires crossing a threshold where predictions become useful to domain experts who don't care about ML metrics.

Key insight: Experimental biologists adopted AlphaFold not because of benchmark scores, but because predictions matched their private experimental data at a useful accuracy level.

Implication: Define success by domain expert utility, not benchmark performance. Track adoption signals from target users, not just technical metrics.

要产生现实影响，系统必须跨越一个阈值，使预测结果对不关心 ML 指标的领域专家也变得有用。

关键洞察：实验生物学家采用 AlphaFold，并非因为其基准分数，而是因为其预测结果以有用的准确度与实验生物学家私有的实验数据相匹配。

影响：应以领域专家的实用性定义成功，而非以基准性能定义成功。除技术指标外，还应跟踪目标用户的采用信号。

Workflow: Designing Scientific AI Systems

工作流：设计科学 AI 系统

Step 1: Understand the Domain Problem

第 1 步：理解领域问题

Identify the core scientific question (e.g., "How does a linear amino acid chain fold into a 3D structure?")
Map existing experimental methods and their limitations
Determine what accuracy level constitutes "useful" for practitioners
Identify available data sources and their characteristics

确定核心科学问题（例如，“线性氨基酸链如何折叠成三维结构？”）
梳理现有实验方法及其局限
确定 Practitioners 眼中达到何种准确度才算“有用”
识别可用的数据源及其特征

Step 2: Evaluate Off-the-Shelf Approaches

第 2 步：评估现成方案

Test standard ML architectures on the problem
Document where generic approaches fail or underperform
Identify domain-specific constraints that require custom solutions
Establish baseline performance for comparison

Warning: Off-the-shelf ML approaches are typically insufficient for scientific breakthroughs. Expect to develop domain-specific innovations.

在问题上测试标准 ML 架构
记录通用方案失效或表现不足的具体位置
识别需要定制解决方案的领域特定约束
建立用于比较的基线性能

警告：现成的 ML 方案通常不足以实现科学突破。应预期需要开发领域特定创新。

Step 3: Invest in Research Iteration

第 3 步：投入研究迭代

Allocate majority of compute budget to experimentation, not final training
Design fast iteration cycles to test architectural hypotheses
Track ideas systematically—expect most to fail
Look for compounding improvements from combining innovations

将大部分算力预算分配给实验，而非最终训练
设计快速迭代周期，以测试架构假设
系统地跟踪创意——应预期大多数创意都会失败
寻找通过组合创新产生的累积效应

Step 4: Validate with Blind Benchmarks

第 4 步：使用盲测基准进行验证

Use held-out test sets where answers are unknown to prevent overfitting:

Validation Strategy:
├── Training data: Known structures (e.g., Protein Data Bank)
├── Validation data: Subset of known structures, held out
└── Blind test: Problems with unpublished answers
    └── Example: CASP competition (biennial, ~100 proteins, answers unreleased)

Critical: True capability can only be measured on problems where answers are unknown to everyone, including system developers.

使用答案未知的留出测试集，以防止过拟合：

验证策略：
├── 训练数据：已知结构（例如，Protein Data Bank）
├── 验证数据：从已知结构中留出的子集
└── 盲测：答案尚未公开的问题
    └── 示例：CASP 竞赛（每两年一届，约 100 个蛋白质，答案未公开）

关键点：真正的能力只能在答案对所有人（包括系统开发者）都未知的问题上衡量。

Step 5: Cross the Relevance Threshold

第 5 步：跨越相关性阈值

Identify domain experts willing to test against their private data
Track word-of-mouth adoption signals
Measure whether users validate predictions independently
Monitor for emergent use cases you didn't anticipate

识别愿意使用私有数据测试系统的领域专家
跟踪口碑采用信号
衡量用户是否独立验证了预测结果
监测未预料到的涌现用例

Workflow: Releasing Scientific AI Tools

工作流：发布科学 AI 工具

Accessibility Over Capability

可访问性优先于能力

A tool's impact depends more on ease of access than raw capability.

Example: AlphaFold's database of 200 million pre-computed predictions drove adoption far more than open-source code alone.

工具的影响更多取决于获取难易程度，而非原始能力。

示例：AlphaFold 包含 2 亿条预计算预测结果的数据库，对采用的推动远超单纯开源代码。

Release Strategy Checklist

发布策略检查清单

□ Code release (minimum viable)
  - Open source the model and training code
  - Provide documentation for technical users

□ Pre-computed results (high impact)
  - Generate predictions for common use cases
  - Host accessible database
  - Enable non-technical users to benefit

□ API access (broader reach)
  - Remove computational barriers
  - Enable integration into existing workflows

□ User interface (maximum accessibility)
  - Build tools for domain experts without ML expertise
  - Focus on their workflow, not your architecture

□ 发布代码（最低可行方案）
  - 对模型和训练代码进行开源
  - 为技术用户提供文档

□ 提供预计算结果（高影响力）
  - 为常见用例生成预测结果
  - 托管易于访问的数据库
  - 让非技术用户也能从中受益

□ 提供 API 访问（扩大覆盖范围）
  - 消除计算门槛
  - 支持集成到现有工作流中

□ 提供用户界面（最大化可访问性）
  - 为不具备 ML 专业知识的领域专家构建工具
  - 关注他们的工作流，而非你的架构

Anticipate Emergent Capabilities

预判涌现能力

Power users will discover capabilities you didn't design for.

Example: Protein interaction prediction emerged from AlphaFold 2 days after release—a capability the team hadn't explicitly built or tested.

Action: After release, monitor novel use cases and validate whether they work reliably.

高级用户会发现你未曾设计的能力。

示例：AlphaFold 发布两天后便涌现出蛋白质相互作用预测能力——这是团队并未明确构建或测试过的能力。

行动：发布后，应监测新的用例，并验证它们是否能够可靠运行。

Decision Framework: Data vs Research Investment

决策框架：数据投入与研究投入

When allocating resources between data acquisition and research:

Prioritize DATA when:
├── Problem is well-understood
├── Standard architectures perform reasonably
├── Diminishing returns not yet observed
└── Data collection is feasible and affordable

Prioritize RESEARCH when:
├── Off-the-shelf approaches fundamentally fail
├── Domain constraints require custom solutions
├── Data is expensive or limited
└── Existing approaches hit performance ceiling

Default bias: Lean toward research investment. The AlphaFold case demonstrates that architectural innovation can provide orders of magnitude more value than data scaling.

在数据获取与研究之间分配资源时：

优先选择 DATA，当：
├── 问题已得到充分理解
├── 标准架构表现尚可
├── 尚未观察到边际收益递减
└── 数据收集可行且成本可承受

优先选择 RESEARCH，当：
├── 现成方案出现根本性失败
├── 领域约束要求定制解决方案
├── 数据昂贵或有限
└── 现有方案已达到性能上限

默认倾向：优先投入研究。AlphaFold 案例表明，架构创新能够比数据扩展提供高出一个数量级的价值。

Validation Checklist for Scientific AI

科学 AI 验证检查清单

Before claiming a system works for real-world use:

□ Blind benchmark validation
  - Test on problems with unknown answers
  - Use established domain competitions if available
  - Avoid overfitting to published test sets

□ Ablation studies
  - Remove each component systematically
  - Measure contribution to overall performance
  - Verify no single idea explains >50% of improvement

□ Domain expert validation
  - Find practitioners with private test data
  - Track whether they adopt the tool
  - Monitor word-of-mouth signals

□ Relevance threshold assessment
  - Define "useful" accuracy for domain experts
  - Measure whether system crosses this threshold
  - Distinguish benchmark improvement from practical utility

在宣称系统适用于现实场景之前：

□ 盲测基准验证
  - 在答案未知的问题上进行测试
  - 如有成熟的领域竞赛，应优先使用
  - 避免对已公开发布的测试集过拟合

□ 消融研究
  - 系统地移除每个组件
  - 衡量各组件对整体性能的贡献
  - 确认没有单一创意解释了超过 50% 的提升

□ 领域专家验证
  - 寻找拥有私有测试数据的从业者
  - 跟踪他们是否采用该工具
  - 监测口碑信号

□ 相关性阈值评估
  - 定义领域专家眼中“有用”的准确度
  - 衡量系统是否跨越该阈值
  - 区分基准性能提升与实际效用

Key Concepts Reference

核心概念参考

Concept	Definition	Relevance
Blind benchmark	Evaluation where answers are unknown to all participants	Prevents overfitting, validates true capability
Ablation study	Systematic removal of components to measure contribution	Identifies which innovations matter
Relevance threshold	Accuracy level where predictions become useful to practitioners	Defines real-world success
Equivariance	Neural network property where outputs transform predictably with inputs	Important for physical/geometric domains
Self-distillation	Using model predictions as additional training data	Technique for improving with limited labeled data

概念	定义	相关性
盲测基准	所有参与者均不知道答案的评估方式	防止过拟合，验证真实能力
消融研究	系统地移除组件以衡量其贡献	识别真正重要的创新
相关性阈值	预测结果开始对从业者有用的准确度水平	定义现实场景中的成功
等变性	输入发生可预测变换时，输出也随之变换的神经网络属性	对物理和几何领域尤为重要
自蒸馏	将模型预测结果用作额外训练数据	在标注数据有限时改进模型的技术

Anti-Patterns to Avoid

应避免的反模式

Benchmark Obsession

痴迷基准

Optimizing for published benchmarks while missing what domain experts actually need.

Fix: Define success by practitioner adoption, not leaderboard position.

针对已公开发布的基准进行优化，却忽视领域专家的实际需求。

修复方案：以从业者采用情况定义成功，而非以排行榜排名定义成功。

Data Hoarding

囤积数据

Assuming more data will solve fundamental architectural limitations.

Fix: Test whether 10x more data would plausibly help before investing in collection.

假设更多数据能够解决根本性的架构限制。

修复方案：投入数据收集前，先测试增加 10 倍数据是否确实可能有所帮助。

Capability Without Accessibility

有能力但不可访问

Building powerful tools that domain experts can't use.

Fix: Release in the most accessible format possible—pre-computed results often beat open-source code.

构建领域专家无法使用的强大工具。

修复方案：以尽可能易访问的形式发布——预计算结果通常比开源代码更能推动采用。

Single Big Idea Hunting

寻找单个大创意

Searching for one revolutionary insight instead of accumulating improvements.

Fix: Track many mid-scale ideas; expect breakthroughs from combinations.

寻找一个革命性洞察，而非持续累积改进。

修复方案：跟踪大量中等规模的创意；突破往往来自组合。

Closed Validation

封闭式验证

Testing only on data where you know the answers.

Fix: Participate in blind benchmarks; seek domain experts with private test data.

仅在已知答案的数据上进行测试。

修复方案：参与盲测基准；寻找拥有私有测试数据的领域专家。