▶ 原文链接

基于AI的药物研发中的实验闭环(Lab-in-the-Loop)

来源: ICML 2026 特邀演讲 | Aviv Regev | 2026-07-08(录制日期) 分类: 其他 原文发表: 2026-07-08 纪要生成: 2026-07-10


全集重点


嘉宾/话题简介

Aviv Regev 是基因泰克(Genentech/Roche)研究与早期开发部门的执行副总裁兼负责人,此前曾在麻省理工学院、布罗德研究所和霍华德·休斯医学研究所担任要职。她是单细胞基因组学和计算生物学领域的先驱,也是人类细胞图谱(Human Cell Atlas)项目的联合创始人。本次演讲中,她系统阐述了基因泰克如何构建一个覆盖靶点发现、药物发现和开发全流程的“实验闭环”(Lab-in-the-Loop)系统,通过大规模数据生成与AI模型的迭代互动,应对药物研发中成功率极低的根本性挑战。


分节详述

00:26:24 药物研发的困境与AI的机遇

本节重点

详细精要

💬 精华片段(中文)

“是的,这个空间名义上巨大,但它是结构化的,其真实维度要小得多。没错,AI 对此很擅长。是的,它是多尺度的,所有转变和变换都是非线性的。AI 对此也很擅长。”

"Yes, the space is nominally huge, but it's structured and its real dimensionality is a lot smaller. Well, Bingo, AI is good for that. Yes, it's multi scale and all the transitions, transformations are nonlinear. AI is good for that, too."

00:34:19 核心范式:Lab-in-the-Loop(实验闭环)

本节重点

详细精要

💬 精华片段(中文)

“我们需要做的是将这些部分组合起来,置于我们称之为‘实验室(有时是诊所)在环中’的架构里。从实验开始,收集数据,训练模型,用模型定义下一组实验,并不断重复。”

"What we need to do is to put these together in what we call the lab or sometimes the clinic in the loop, start with an experiment, collect data, train a model, use the model to define the next set of experiments. And repeat."

00:35:24 靶点发现第一步:解码非编码基因组与疾病因果关系

本节重点

详细精要

💬 精华片段(中文)

“过去,人们基于对生物学的最佳理解设计了大约1万个序列。那点数据量不大,模型的预测效果也不好。但如果我们转向‘数量取胜’,测试非常大量的随机序列,这实际上在实验室里更容易做到,会怎样呢?”

"So in the past, people designed about 10,000 sequences based on their best understanding of biology. That's not a lot of data, and the models were not predicting well. But what if we went to, for sheer quantity instead, and we tested very large numbers of random sequences, which is actually much easier to do in the lab."

00:40:15 靶点发现第二步:扰动筛选(Perturb-seq)解析编码基因网络

本节重点

详细精要

💬 精华片段(中文)

“这种可解释模型将扰动分组为具有相似影响的模块,并将受影响的RNA分组为共调控基因。这个回路在免疫细胞生命周期的五个主要部分中,分配了不同的关键蛋白质。”

"The interpretable model that we learn from one such screen is a regulatory model. It groups perturbations into modules with similar impacts, and it groups the impacted Rnas into co regulated genes. This circuit assigns distinct proteins of interest in each of five major parts of the life cycle of these immune cells."

00:42:59 走向多模态:让AI整合RNA与影像视图

本节重点

详细精要

💬 精华片段(中文)

“那么一个有趣的生物学问题是:如果我基于RNA或基于显微镜图像来关联这些扰动,我会得到相同还是不同的答案?...结果证明,我们得到的基本上是相同的。事实上,在RNA中具有相似影响的扰动,在成像中也具有相似影响,反之亦然。”

"So an interesting biological question: If I relate perturbations to each other based on looking at RNA or based on looking at microscopy images, do I actually get a same or a different answer? ... It turns out that we kind of get the same, in fact, perturbations that have similar impacts in RNA also have similar impacts in imaging and vice versa."

00:48:30 应对组合爆炸:用主动学习探索五元组合可能

本节重点

详细精要

💬 精华片段(中文)

“如果我们不是实验性地去做全部18万对组合,而是用AI预测出600对信息量最大的组合去测试,从而得到一个更好的模型,然后用它来完成剩下的工作呢?这就是迭代式‘实验室在环中’让我们能做的事。”

"What if instead of doing 180,000 pairs experimentally, we'd actually use AI to predict 600 pairs most informative to test in order to get a better model. And that can complete the rest. This is what an iterative lab in the loop allows us to do."

00:51:57 从患者数据构建疾病基础模型与AI代理

本节重点

详细精要

💬 精华片段(中文)

“这些数据持续以如此之快的速度增长,在过去的几年里,它们每六个月就翻一番。但这些数据增长得越多,对其进行推理和回答药物研发关键问题的要求就越高,劳苦程度也就越甚。”

"In fact, these kinds of data continue to grow so rapidly that for this past several years, they've been doubling every six months. But the more these data grow, the more demanding it actually becomes to reason over them and answer critical questions. For drug R&D, the worst it is in the toiling."

00:57:44 药物分子设计闭环(一):小分子

本节重点

详细精要

原文发表:2026-07-08  ·  纪要生成:2026-07-10