"/typst/templates/paper-slides.typ": body-s = 15pt body-m = 16pt head-s = 18pt : paper-slides.with( title: "SAGA: Accelerating Scientific Discovery with Autonomous Goal-evolving Agents", subtitle: "arXiv 2512.21782 · 2025", author: "Biscuit", date: "2026-07-15", ) -slide[ (center + horizon)[ (size: 28pt, weight: "bold", fill: accent)[SAGA] (0.3em) (size: 20pt, fill: accent)[Autonomous Goal-evolving Agents for Scientific Discovery] (2em) (size: 14pt, fill: secondary-text)[Yuanqi Du, Botao Yu, et al. · Cornell, OSU, Yale, EPFL, UC Berkeley, Broad Institute, Deep Principle] (0.3em) (size: 13pt, fill: secondary-text)[arXiv 2512.21782 · github.com/btyu/SAGA (MIT)] (1em) -block[ (size: 13pt, weight: "bold", fill: accent)[亮点：抗生素 + 纳米抗体获湿实验验证 · 横跨 5 个科学领域] ] ] ] 目录 ( columns: (1fr, 1fr), gutter: 2em, align: top, [ ( columns: (0.5fr, 4fr), inset: 6pt, stroke: none, align: (center, left), [ (size: 22pt, weight: "bold", fill: accent)[1]], [ (size: body-s)[动机与核心思想 \ (fill: secondary-text, size: 12pt)[固定目标→自演化目标]]], [ (size: 22pt, weight: "bold", fill: accent)[2]], [ (size: body-s)[SAGA 双层架构 \ (fill: secondary-text, size: 12pt)[Planner → Implementer → Optimizer → Analyzer]]], [ (size: 22pt, weight: "bold", fill: accent)[3]], [ (size: body-s)[三种协作模式 \ (fill: secondary-text, size: 12pt)[Co-pilot / Semi-pilot / Autopilot]]], [ (size: 22pt, weight: "bold", fill: accent)[4]], [ (size: body-s)[五领域实验全景 \ (fill: secondary-text, size: 12pt)[抗生素、纳米抗体、DNA、材料、化工]]], [ (size: 22pt, weight: "bold", fill: accent)[5]], [ (size: body-s)[湿实验验证 \ (fill: secondary-text, size: 12pt)[抗生素 MIC + 纳米抗体 K D]]], [ (size: 22pt, weight: "bold", fill: accent)[6]], [ (size: body-s)[局限与展望 \ (fill: secondary-text, size: 12pt)[lab-in-the-loop 与设计空间自动发现]]], ) ], [ [ (size: body-s, weight: "bold", fill: accent)[核心命题] (0.3em) (size: body-s, fill: secondary-text)[自动化目标函数设计，而非手动指定并优化固定目标 —— 这才是科学发现的瓶颈所在] (0.5em) (size: body-s, weight: "bold", fill: accent)[关键数据] (0.2em) ( columns: (1fr, 1fr), inset: 4pt, stroke: 0.4pt, align: (center, center), [ (size: 12pt, weight: "bold")[5 领域]], [ (size: 12pt, weight: "bold")[3 模式]], [ (size: 12pt)[2 湿实验]], [ (size: 12pt)[4 个新 hit]], [ (size: 12pt)[3 个 binder]], [ (size: 12pt)[48% 提升]], ) ] (0.3em) (size: 12pt, fill: secondary-text)[github.com/btyu/SAGA · MIT License · Python + Docker] ], ) 动机：固定目标的困境 ( columns: (1fr, 1fr), gutter: 1em, align: top, [ (title: "问题")[ (size: body-s)[ 现有科学发现 agent 假设目标函数已知且固定。但：\ - 计算代理与实际实验相关性弱 \ - 优化算法会 exploit 代理误差，产生"高分低质"解 \ - 目标及其权重的搜索空间本身组合爆炸 ] ] (0.3em) (title: "SAGA 的答案")[ (size: body-s)[ 将目标函数设计本身变成自动发现过程：\ 外层循环探索目标空间，内层循环优化候选解。\ 模仿科学家"根据中间结果调整目标"的迭代范式。 ] ] ], [ (0.3em) (title: "思考快与慢")[ (size: body-s)[ (weight: "bold")[Inner Loop = 系统 1]：给定目标，快速搜索最优解 \ (weight: "bold")[Outer Loop = 系统 2]：审视结果，反思并调整目标 ] ] ], ) SAGA 双层架构 (size: head-s, weight: "bold", fill: accent)[外循环：目标演化 + 内循环：候选优化] ( columns: (1fr, 1fr), gutter: 1em, align: top, [ (title: "")[ (size: body-m)[1. Planner（规划器）] (size: body-s)[分解科学目标 → 可计算 objectives（名称、描述、方向、权重）] (size: body-m)[2. Implementer（实现器）] (size: body-s)[自动实现评分函数（搜索 web → 写代码 → Docker 验证）· 不可实现则通知 Planner 重试] (size: body-m)[3. Optimizer（优化器）] (size: body-s)[默认 LLM 进化算法：生成候选 → 批量打分 → 选 top 更新池 · 可插拔替代] (size: body-m)[4. Analyzer（分析器）] (size: body-s)[统计分数趋势 + 深入结构分析 → 生成报告 → 触发下轮目标 · 可判定提前终止] ] ], [ ()[ (size: body-s)[ (weight: "bold")[输入]：自然语言目标 + 上下文 + 初始 objectives + 初始候选 \ (weight: "bold")[循环]：Planner → Implementer → Optimizer → Analyzer → Planner \ (weight: "bold")[输出]：跨迭代最优候选集 ] ] ()[ (size: body-s)[ - Implementer 可自动搜索 web 编写评分函数 \ - Analyzer 写代码分析候选结构/性质 \ - 所有评分函数在 Docker 中运行保证安全 \ - 评估指标对优化过程不可见（hold-out） ] ] ], ) 三种协作模式 (size: head-s, weight: "bold", fill: accent)[从全人工干预到全自动，适合不同研究阶段] ( columns: (1.5fr, 2fr, 2.5fr, 1fr), inset: 6pt, stroke: 0.4pt, align: (left, left, left, center), table.header( [ (size: body-s, weight: "bold")[模式]], [ (size: body-s, weight: "bold")[人参与点]], [ (size: body-s, weight: "bold")[自动化部分]], [ (size: body-s, weight: "bold")[适用场景]], ), [ (size: 13pt, weight: "bold")[Co-pilot]], [ (size: 13pt)[Planner 提案 + Analyzer 报告均可修改]], [ (size: 13pt)[Implementer + Optimizer 自动运行]], [ (size: 13pt)[早期探索、高风险领域]], [ (size: 13pt, weight: "bold")[Semi-pilot]], [ (size: 13pt)[仅 Analyzer 输出由人评审]], [ (size: 13pt)[Planner + Implementer + Optimizer 自动]], [ (size: 13pt)[中期优化、有经验方向]], [ (size: 13pt, weight: "bold")[Autopilot]], [ (size: 13pt)[无人类干预]], [ (size: 13pt)[全部四个模块自主运行]], [ (size: 13pt)[成熟方向、大规模探索]], ) (0.3em) (title: "结果")[ (size: body-s)[ Autopilot 在多个任务中达到或超越 human-in-the-loop 水平。88.8% 的 objectives 由统计驱动，跨独立重复实验高度一致。 ] ] 五领域实验全景 (size: head-s, weight: "bold", fill: accent)[横跨化学、生物、材料 · 优化目标均由 SAGA 自动演化] ( columns: (1fr, 1.2fr, 1.5fr, 1.5fr), inset: 5pt, stroke: 0.4pt, align: (left, left, left, center), table.header( [ (size: body-s, weight: "bold")[领域]], [ (size: body-s, weight: "bold")[任务]], [ (size: body-s, weight: "bold")[核心结果]], [ (size: body-s, weight: "bold")[验证方式]], ), [ (size: 13pt, weight: "bold")[抗生素设计]], [ (size: 13pt)[E. coli 抗菌小分子]], [ (size: 13pt)[4 个 hit · MIC 16 μg/mL]], [ (size: 13pt, fill: accent)[湿实验]], [ (size: 13pt, weight: "bold")[纳米抗体设计]], [ (size: 13pt)[PD-L1 结合剂]], [ (size: 13pt)[3 个 binder · K D 300-400nM]], [ (size: 13pt, fill: accent)[湿实验 (BLI)]], [ (size: 13pt, weight: "bold")[功能 DNA 设计]], [ (size: 13pt)[HepG2 增强子]], [ (size: 13pt)[特异性 +48% · motif +48%]], [ (size: 13pt)[计算 (MPRA)]], [ (size: 13pt, weight: "bold")[无机材料设计]], [ (size: 13pt)[永磁体 + 超硬材料]], [ (size: 13pt)[>90% 含轻元素 · >75% 碳氮硼化物]], [ (size: 13pt)[DFT 计算]], [ (size: 13pt, weight: "bold")[化工过程设计]], [ (size: 13pt)[丁醇/水共沸分离]], [ (size: 13pt)[消除无效单元操作 · 平衡成本/纯度]], [ (size: 13pt)[short-cut 模拟]], ) (0.2em) (size: 12pt, fill: secondary-text)[SAGA 自演化出的 objectives 还能迁移到其他方法（REINVENT4、TextGrad），提升其性能。] 湿实验验证亮点 ( columns: (1fr, 1fr), gutter: 1em, align: top, [ (title: "抗生素：化合物 8")[ (size: body-s)[ (weight: "bold")[MIC]：16 μg/mL (E. coli) \ (weight: "bold")[结构新颖性]：Tanimoto 距离 0.28，远低于 0.4 阈值 \ (weight: "bold")[安全性]：HEK293 + HepG2 细胞系无细胞毒性 \ (weight: "bold")[筛选]：28 个化合物中 4 个抑制率 >80% ] ] (title: "关键")[ (size: body-s)[ (weight: "bold")[总筛选]：28 个 SAGA 设计的化合物 → 4 个 hit → 1 个最好的（化合物 8）\ (weight: "bold")[机制]：SAGA 自动引入代谢稳定性 + 药物相似性目标，避免 reward hacking ] ] ], [ (title: "纳米抗体：3 个 PD-L1 Binder")[ (size: body-s)[ (weight: "bold")[K D]：300-400 nM (BLI 实验) \ (weight: "bold")[结构新颖性]：CDR3 与已知抗体 lt 20% 序列相似性 \ (weight: "bold")[独特结构]：A2 的 CDR3 形成非经典 α-helix \ (weight: "bold")[统计显著性]：复合评分函数 p = 0.03，任何单指标 p > 0.05 ] ] (title: "关键")[ (size: body-s, style: "italic")[SAGA 自动构建的复合目标显著区分了 binder 和 non-binder（p = 0.03），而任何单一 in silico 指标都无法做到。] (0.1em) ] ], ) 局限与展望 ( columns: (1fr, 1fr), gutter: 1em, align: top, [ (title: "技术局限")[ (size: body-s)[ (weight: "bold")[计算依赖]：仅适用于可计算验证的目标，无法接入真实实验反馈\ (weight: "bold")[设计空间预定义]：需人为指定搜索空间（如"小分子"），无法自动决定模态\ (weight: "bold")[模型开销]：依赖 GPT-5 级别模型，API 费用高\ (weight: "bold")[代码不完整] ] ] ], [ (title: "可借鉴之处")[ (size: body-s)[ - 双层架构可迁移到任何科学优化任务\ - 自演化目标函数解决 reward hacking 和代理偏差\ - Co-pilot 模式提供实用的 human-in-the-loop 方案\ ] ] ], ) Q&A (3em) (center, text(size: 18pt, weight: "bold", fill: accent)[谢谢！Q & A]) (0.5em) (center, text(size: 11pt, fill: secondary-text)[arXiv 2512.21782 · Code: github.com/btyu/SAGA · 汇报人: Biscuit])