科技生物对话第1集：麻省理工学院的Regina Barzilay与主持人Najat Khan

欢迎来到TechBioTalks，我们将在这里探讨重塑AI与医学未来的大胆构想、突破性技术和远见卓识者。首期节目，我非常激动地邀请到我的挚友兼杰出智者Regina Varzle。Regina身兼数职，她是MIT的杰出教授，领导Jameel诊所的AI研究，同时也是麦克阿瑟天才奖得主。

Welcome to TechBioTalks, where we will explore the bold ideas, breakthrough technologies, and visionary people reshaping the future of AI and medicine. For our very first episode, I couldn't be more excited to welcome my dear friend and a brilliant mind, Regina Varzle. Regina wears a lot of impressive hats. She's a distinguished professor at MIT. She leads AI efforts at the Jameel Clinic and is a recipient of the MacArthur Genius Grant.

她还是美国国家医学院和国家工程院院士。基本上在AI与健康的交叉领域有任何大胆构想时，她不仅参与其中，往往已经领先数步。Raghenia团队在医疗AI领域开创了突破性工作，从开发用于乳腺癌和肺癌早期检测的Mirai和Sybil深度学习模型（现已投入使用），到我们曾在Recursion合作开发的分子建模与药物发现AI系统Bolts two。

She is also a member of the National Academies of Medicine and Engineering. Basically, there's a bold idea at the intersection of AI and health. So she's not just working on it. She's likely already a few steps ahead. Raghenia's team has led groundbreaking work in AI for healthcare, from developing Mirai and Sybil deep learning models for breast and lung cancer early detection that are now being deployed to Bolts two, an AI system for molecular modeling and drug discovery that we collaborated on at Recursion.

我们实验室的创新涵盖影像学、化学、个性化肿瘤学等诸多领域，包括预测癌症耐药性等前沿新研究。请大家与我一起热烈欢迎杰出的Regina Barsley教授。

Our lab's innovations span imaging, chemistry, personalized oncology, and so much more, including new work in bold areas such as predicting resistance in cancer. Please join me in warmly welcoming the brilliant Professor Regina Barsley.

非常感谢Najaten，见到你真高兴。经过这么长的介绍环节，我们终于可以开始聊些有趣的内容了。

Thank you very much, Najaten. It's really fun to see you. So, after this very lengthy introduction, we can actually start talking about interesting stuff.

确实如此。我常说每当思考问题或关注领域时，喜欢用对患者的潜在影响来衡量一切。医疗关乎每个人，我们终将成为患者。有句名言总让我深有共鸣：人们往往高估创新技术在短期内的影响，而低估其长期效应——这是我对Roy Amara原话的个性化诠释。

Absolutely. I've always mentioned that whenever I think about a problem or an area for us to focus on, I like to bookend everything in terms of the impact it can have on patients. Healthcare touches all of us, and we'll all be patients at some point in our lives. And a quote that always resonates with me is the fact that we generally tend to overestimate the effect of innovation, technology in the short run and underestimate the effect in the long run. I have personalized it a little bit from the quote from Roy Amara.

那么Raghena，第一个问题让我想起我们在强生初遇时，你作为科学顾问委员会成员，是组建AI团队的关键推动力。2018-2019年间你的团队研究MRI结合AI实现乳腺癌早期检测，这个当时非常新颖的构想我记得你演示时遭遇了不少质疑。

So Raghena, with that first question, I remember the first time we met when we were both at J and J. You were on the Scientific Advisory Board and you were a crucial driving force in building the AI team there. Your team, this is back in 2018, 2019, was working on MRI, AI to ensure we can actually enable early detection in breast cancer. It was such a novel idea. Still I remember when you were presenting it, there were a lot of naysayers and non believers.

你的团队从概念到实践取得巨大进展——正如你说过，开发算法并非最难，真正的挑战在于部署推广。能否分享一下这个算法的应用历程？

Your team has made so much progress going from idea, and you've said this before, developing the algorithm is not the hardest work, it's deploying and getting adoption. So can you just tell us a little bit about the journey that algorithm has been on?

确实，我依然记得那次会议，当时人们对AI持高度怀疑态度，要知道，做AI远不如现在这么时髦。尤其是当你进入制药行业时，他们觉得这很有趣，但并非主业。但这个构想其实超越了检测范畴。就我们对癌症的认知而言，癌症不像流感那样几天内发作，它实际上是个非常漫长的组织病变过程。

So yeah, I still remember our meeting and I do remember that at the time people were really skeptical, you know, doing AI was not as sexy as it is today. So, especially when you came to pharmaceutical industry and they thought this is fun, but it's not our main business. But the idea actually goes beyond detection. As to what we know from cancer, cancer doesn't develop like flu in several days. It's actually a very long process sometimes of tissue transformation.

如今检测通常在肿瘤长到足够大时才进行——那时可能有数十亿个细胞，放射科医生才能在影像上看到。我们确实能观察到极早期的发展阶段，但要等到人类肉眼可见还需要很长时间。事实上，大量研究表明，以乳腺癌为例，即使患者每年筛查，在确诊女性群体中，有些人的影像上其实已有癌症迹象却被漏诊，因为特征非常模糊——当然如果不明显医生就不会做活检以免过度干预。所以我们不该依赖肉眼判断谁有风险或未来可能患病，而应该借助机器——当你能用数十万乃至数百万张已知患者结局的影像训练时，机器比人类更具优势。

And today the detection is done when the tumour is large enough, which is maybe billions of cells, when the radiologist can see it on the image. So we definitely see very early stages of development and it takes a long time until human can see it. And in fact, there are a lot of studies that demonstrate, for instance, for breast cancer, even when the patient is screened every year, if you look at women who diagnosis here, for one set of them, actually on the image the cancer was there and radiologist listed, because it was very ambiguous and of course if it is not very clear then they will not biopsy you because they don't want to interfere. So as a result we shouldn't be relying on our eyes to detect who is at risk and who is likely to develop the disease in the future. We really want to use a machine because machine have a benefit over human when you can train it on, you know, hundreds of thousands or millions of images when you know the outcome of the patients.

机器能识别这些人类肉眼难以察觉的癌症细微征兆。比如我们开发的Mirai，这是我得意门生亚当·亚拉（现任伯克利和UCSF教授）的博士成果。我们还开发了适用于肺癌的Siebel系统，具备相同功能。

And the machine can identify these very subtle signs that are difficult for human eye to predict cancer. And we've seen Mirai. This is a tool that was a PhD work of my dear student who is now a professor at Berkeley and UCSF, Adam Yalla. And we also developed Siebel, works for lung cancer. It actually does the same functionality.

它能分析当前医院医生认为良性的影像，预测患者未来六年内的患病概率，其表现远超任何放射科医生——医生甚至不会尝试做这种未来预测。关键在于：在临床中发现高风险患者后该怎么办？仅仅告知风险是远远不够的。

It looks at the image, which today is considered benign by the doctors in the hospitals and gives you the likelihood of developing disease for up to six years. And it does it, you know, better than any radiologist can do. Radiologists are not even trying to do this task of predicting the future. And the question is, what do you do in the clinic when you identify high risk patients? It's not enough just to say to them you are high risk.

我们需要推进到下一步：或许用MRI等更好的方式筛查，或增加筛查频率，对于乳腺癌甚至可以采用化学预防措施。我们正致力于推动这些临床转化，不仅在美国，还在拉美（智利、墨西哥、巴西）、印度、台湾、克罗地亚、意大利等多地展开——因为每个医疗体系的转化路径都独具特色。技术引入是通用的，但如何落地整合？这需要众多参与者的深度认同，是个漫长征程。

You really want to take it to the next step and maybe screen them with better modality like MRI, or maybe you want to screen them more frequently, or like in the case of breast cancer, want to put them into chemo preventative tool. So what we are trying to do is actually do these translations. And we're not only doing it in The United States, we are doing it across the world, in Latin America, like Chile, Mexico, Brazil, India, Taiwan, Croatia, Italy, many countries, because this translation is really specific to a system. So it's only in a generic way how you, you know, bring in the technology, but how do you incorporate it? And it's really a journey because it needs to have a very serious buy in from many, many players.

目前Mirai已分析超过200万张影像，经过充分验证且持续使用中。所以临床AI的关键问题不在于算法本身，而在于获得算法后如何惠及患者。

Today, Mirai ran over 2,000,000 images, which is quite tested and it continues to be used. So this is the big question to me in clinical AI, is not the algorithms, is it the next step when you have the algorithm, how do you bring it to the patient?

Regena，我想快速追问几点：你强调增强医生能力、优化治疗很重要，但整合到不同国家医院工作流程这部分也极为关键。能否谈谈如何推动规模化应用？如何促进这类技术的采纳？另外如何避免模型漂移问题？

You know, Regena, just maybe a couple of quick follow-up questions on that. I mean, I love how you emphasize the importance of, A, supercharging the physicians, enhancing treatment for patients, But then also this piece around needing to integrate it in the various workflows of the various hospitals and countries you're going into. Can you speak a little bit to like, how do you drive scalability? How do you drive adoption for something like that? And then also, how do you avoid any sort of model drift?

你们如何确保结果持续相关以维持采用率？

How do you ensure that the results are consistently relevant to have sustained uptake?

我们欢迎科学家和临床医生进行报告，并有大量人员独立验证和分析。这是我们确保其良好运行且稳健工作的方式。我们已在众多地方实施，数据与Siebel非常相似——数十万张图像处理成果。关于你提到的第二点——如何进行转化？我认为关键在于获得主导转化的临床医生的认同。

We welcome the scientists and the clinicians to report, and there are lots of people who independently validate it and analyze it. So this is our approach towards ensuring that it's doing well and it's working robustly. Again, we're doing it across many, many places. We've seen very similar numbers, similar with Siebel, you know, hundreds of thousands of images works Now towards the second point that you asked about, you know, how do you do the translation? I think a big part of it, it's really to have a buy in from clinician who are leading the translation.

因为仅仅说‘有个很棒的工具，去用吧’是不够的。必须有人思考：如何将其融入现有工作流程？如何让使用它的临床医生感到顺手？如何与监管机构沟通？这需要解决一系列问题。

Because it is not enough to say, oh, there is this great tool and use it. Somebody needs to think, how do I bring it in my current workflow? How do I make it comfortable for clinicians who are using it? How do we approach the regulators? So there is this whole slew of questions that need to be answered.

部分问题已由我的临床合作者解答，部分仍需解决。Najat，接下来我要说的你会很熟悉：NCCN是美国癌症护理组织，负责制定所有癌症的最佳护理标准。虽然它不是唯一机构，但非常重要。目前他们采用了一款AI工具——我知道你参与了Arterra前列腺癌用药决策工具的部署。这就引出一个问题：为何在AI时代，我们的临床指南仍未提及或使用这类工具？

Some of them are already answered by my clinical collaborators, some of them still have to be answered. And Najat, you would be very familiar with what I'm going to say next, that NCCN, which is a cancer care organization in The United States that decides across all the cancers, you know, what are the best standards of care. It's not the only one, but very important one. Currently has a single AI tool that I know that you were involved in deployment, which is Arterra's tool for deciding, you know, patients that need particular medications for prostate cancer. So this is a question: how come we already, like in the era of AI and our clinical guidelines, do not mention any of these tools, do not use it?

这正是我们需要努力改变观念、证明工具价值的持续性课题。

So this is another kind of ongoing topic that one should really work hard to change their mind and to show the value of these tools.

谢谢Regina。我想对观众说明，实际应用中的挑战在于建立信任——那种能持续体现在治疗指南中的信任。以Arterra为例，被纳入NCCN指南后，最近还开始获得医保报销，这进一步提升了患者可及性。可见从算法开发、部署、迭代到纳入指南，最终惠及患者的完整闭环。

Thank you, Regina. I mean, it's just for the viewers, the challenge with actually having adoption, building that trust, right? And building trust in a way that's sustained in the treatment guideline pattern. So Arterra's example, being in NCCN Guidelines, and recently, they're also starting to get reimbursed, which also increases access for patients. So you see the entire end to end loop from algorithm development, deployment, redeployment, incorporating in guidelines, and then actually getting it in the hands of patients.

确实，这是值得骄傲的时刻。记得我的团队在强生时期刚开始这个项目时，遭到无数质疑。但当你看到转化实现时，那些新增的白头发都值了。好了，我们现在要转向另一个领域。

Yeah, that is one of those proud moments you look back. And I remember the first time when my team started working on that, this was when I was at J and J. So many naysayers. But when you see the translation happen, say, you know, those extra gray hairs were worth it. All right, we're going to shift gears to another area.

要知道，如今市面上有太多药物，我们甚至都不清楚它们的作用机制。当然，成功率很低，百分之十远远不够。而你我都有一个共同基因特质——极度缺乏耐心，总想快速改变现状。每当我思考对生物与化学的认知时，闭眼就像看到一张大地图，有些区域是已知的，其余则笼罩在黑暗中。就拿Boltz two项目来说，团队推进速度之快就是很好的例证。

There are so many drugs even, you know, in the market today where we don't understand the mechanism of action. Of course, success rates are low, ten percent is not good enough. And you and I, we have one very common gene, which is we're highly impatient, the sense of urgency to change things quickly. When I think about the understanding of biology and chemistry, I close my eyes, it's like a big map and there's spots that we know and the rest is dark. And so, you know, when we talk about, let's just say, Boltz two as an example, like how quickly the team was able to do that.

但我想说它还具有这样的潜力——我们需要同步理解生物学原理（这点接下来会谈到），从而能更快、更好、更早、更经济地筛选那张地图上的空间，并提高其可及性。这其实贯穿了你所有工作的共同主题：开源、提高可及性，让人们按需定制以扩大影响力。能否谈谈你们如何在极短时间内推动四万用户采用，同时从计算角度保持高效？请多分享些这段历程。

But also it has the potential, I would say, and we need to understand biology too, which is we're going come to next. But to be able to triage that map, that space much faster, better, earlier, more cheaply and make it more accessible, which is a common theme in everything you've done. Open sourcing, making things accessible so that people can customize as they see fit to enhance impact for all. Can you speak a little bit to the, you know, 40,000 users really, really quickly driving adoption and so forth and doing it in a way that was very efficient from the compute perspective as well. So a bit more on the journey.

不得不说，能与你们的团队合作真是莫大的荣幸。你们是个非常、非常出色的团队。

It was such an honor, have to say, to collaborate with your team. Such a fantastic, fantastic team.

首先我想表达感谢。还记得我们关于Bolt two的初次交谈——在加州的那次会议上，我当时说我们有个绝妙构想，但缺乏算力支持，也无法获取资源，还需要讨论和实证验证。后来我和Najat将这份友谊转化成了合作，这正体现了友情的价值。

So first of all, I want to start by saying thank you. And I remember how we started conversation about Bolt two. We met in California in a meeting, and I was kind of saying, we have this great idea, but we don't have compute and, you know, we cannot access it. And we also need a discussion and empirical validation. And then we, Najat and I, this is a value of friendship that you can translate it into collaboration.

你们团队不仅提供了算力支持，还为Bolt two投入了大量工程工作。但必须说明，这个构想的发起者功不可没——Jeremy Wolwan、Gabriela Corso和Saro Passaro。前两位是我刚毕业的博士生，他们去伦敦创办公司了，我衷心祝福。Saro Passaro是合作者，他曾到访MIT。

And you guys were able to jump in and to help both with the compute and with, you know, a lot of engineering works that went into BOLT two. But I should say that all the credit for BOLT two as their instigators of the idea goes to Jeremy Wolwan, Gabriela Corso, and Saro Passaro. The first two were my PhD students who just graduated and they went to London to start balls to sell the company, I wish them all the best. Saro Passaro was a collaborator. He came and visited MIT.

我们和ANCA的学生们将在生物学领域深入探讨。其实很多人都明白，理解亲和力至关重要——这是理解生物机制的关键，仅获得结构姿势远远不够。遗憾的是AlphaFold three不具备这种能力。现有工具的精度太低，根本不足以用于生物预测应用。你刚才提到个很有趣的观点。

But we, you know, and the students from ANCA, we're going to talk very in biology and we can we and many others understand how important it is to understand actually the affinity. This is really key to understanding the biology, understanding what's going on, just having the pose is not enough. And unfortunately, you know, AlphaFold three doesn't have this capacity. And what, you know, everybody who look at the existing tool know that their accuracy is so low, they really are insufficient to do, you know, to use in biological application in predictive capacity. And you said something very interesting.

确实如此。我们对生物学某些领域已知晓，但仍有大片未知黑暗地带。我不认为AI能取代生物学研究来填补所有空白，但像Bolt这类工具能帮助我们建立桥梁——即便某些生物岛域尚不明确，也能进行连接填补。不过对于某些真正未知的小领域，仍需通过实验来澄清，这样系统才能学习。

It is true. There are certain things that we know about biology and that there is a lot of dark space in between. And I do not believe that, you know, with AI we're going to eliminate biology and just do everything we can fill up the space. But what AI can really help us with and the tools like Bolt's tool and others, they can kind of say, okay, even though I don't have like clear biology island here, I can build like bridges and fill out. But there are these areas, this is much smaller areas where I really don't know, so you need to give me experiments to clarify it and then I can learn.

要填补整个图谱，你需要两者的结合，但人工智能确实能指明该进行哪些实验。正如你所说，与其手工操作或半随机选择一些点，AI能在某种程度上真正引导你找到有帮助的方向。但在这一特定案例中，有两项智力层面的突破使BOLT得以实现。其中一项与BOLT相关——我们实验室受到AlphaFold三和其他创新的启发，付出了协同努力，构建了极其丰富的生物分子相互作用表征体系。这种表征方式即便是在整体折叠任务中训练的，也能为你提供思考相互作用强度的基础。

So to infill the whole map, you need a combination of both, but AI can really shed the light of which experiments to do. Rather, as you said, artisanal or sometimes semi random, you just select some point, AI can really guide you where it is in certain way it can help. But in this particular case, were two things that like intellectually sort of enable BOLT to. One of them is related, you know, in the BOLT one, there was really a concerted effort inspired by AlphaFold three and other innovations that we did in the lab to have very rich representation of biomolecular interaction. This representation, even though you train it like on whole folding task, It gives you a foundation to start reasoning about, you know, the strengths of interactions.

但当然，你需要有关于相互作用强度的数据。在这方面，我必须对Samuropasara和Pudelila的团队表达无尽赞誉——他们在如何智能化收集数据上付出了巨大努力。这非常有趣，因为如果你只是把所有关于亲和力的现有数据简单堆砌，会产生太多无法学习的噪声。人们尝试过但都失败了。所以一方面你需要非常智能的数据管理，因为这些数据来自不同实验室、不同类型的实验。

But of course, you need to have data for the strengths of interaction. And here I cannot say, you know, enough good words for Samuropasara, Pudelila, huge effort on how to intelligently collect the data. And this was really interesting, because if you just take all the data that is available on Affinity and they just put it together, there will be so much noise you cannot learn. People try to learn it and they couldn't. So on one hand you need to have very intelligent curation of data because it comes from different labs, from different type of experiments.

你需要整合数据，并建立能够处理这些数据源固有噪声的学习机制。在Saro的带领下，团队在递归工程团队的大力支持下真正做到了这一点。这个工具现已成型，预测亲和力成为我们此前无法实现的新能力——过去我们只能进行耗时耗资巨大的模拟运算。现在任何感兴趣的人都可以直接下载并运行这个工具。

You need to put it together and you need to have a learning mechanism that can handle the noise that is inherent on these data sources. Led by Saro, the team really managed to do it and with a lot of support from, you know, the recursion engineering team, you know, the tool came together and now predicting affinity really it's really a new capacity that we couldn't use prior, that we really needed either to do simulation, which was super, super exciting in time and in terms of real cost. But now anybody who has an interest can just download the tool and run it.

前几天我和人聊到，在Recursion我们不是单独使用它——它已集成到我们的平台中，被应用于每一个项目。正如你所说，凭借我们拥有的专有数据，我们还能进行定制和更新。

You know, was talking to somebody the other day. I mean, at Recursion, we're not using it. It's in our integrated platform. We're using it across every single program. And to your point, we can also customize and update because of all of the proprietary data we have as well.

我们整合了生物学和化学数据。有人曾用BOLT作类比，这个比喻非常贴切。过去我们要去图书馆翻找正确的书籍和研究资料，现在直接上网就能获取所需。这种早期筛选能力不仅如你所说能降低时间和成本，还提高了准确性。

We integrate the biology and the chemistry data. I was talking to somebody and they mentioned it as sort of an analogy for BOLT's too, which resonated a lot. Before we used to go to the library and figure out the right book and within that the right research and so forth. And now we can just go online and figure out what we need to. So the ability to triage early, not only reduces time and cost, which is you're spot on and the accuracy of it, right?

FEP级别的预测过去因为成本高昂通常要拖到很后期才做。但这还能让你更客观——这就是摆脱手工操作的意义所在。我们都有经验，但这能让你在更广范围内搜索，可能会发现意料之外的东西，这才是创新的美妙之处。

FEP level predictions, which people would generally do much, much later because it was expensive. But it also makes you more unbiased. This is the whole piece of not being artisanal, right? We all have experiences, but it allows you to have a broader scope to search. And you might find things you don't expect and that is the beauty of innovation.

归根结底这就是研究的魅力。我想说，这次合作非常愉快。每次在LinkedIn上看到有人以稍不同的方式使用它，都让我会心一笑——这正是我们共同提升的方式。Regina，非常感谢你在这项合作中的贡献。

That's the beauty of research at the end of the day. So I just wanted to say, you know, it's such a pleasure to collaborate on this, but just every time I see something pop up on my LinkedIn, that somebody is using it in a slightly different way, just brings a smile. This is how we lift all boats. So a huge thank you, Regina, on the collaboration there.

看到人们真正拥有远见卓识的地方真是令人惊叹，你能克服所有不同的障碍——比如寻找计算资源就是其中之一，当然还有许多其他挑战——关键在于当你怀有渴望时能走多远。正如你所说，纳詹，全球各地实验室的人们，无论是在企业还是大学里，都能直接采用这些技术，并在他们的内部数据研究中进行微调。有些地方会反馈说这太神奇了，带来了他们原本无法实现的新成果；而另一些人则会指出在某些分子类型上它并不奏效，这其实很棒。

It was truly remarkable to see where people really have vision, how you can undergo through all the different barriers that were there and, you know, scouting for computation is one of them, but there were of course many others, you know, how far you can push when you have desire. And as, you know, as you already said, Najan, the fact that people in all labs around the world and both in companies and in the universities can just take it and, you know, fine tune them on their internal data study. And you know, in some places they'll say, yeah, this is amazing, it brought me to this new result that I could not have done, and somebody says, you know, on this type of molecules it doesn't work, which is great.

这也是一种学习。

Which is also learning.

我们需要明确哪些领域有待改进，或许现在就会有人接手这个领域并思考如何做得更好。最后我想补充的是，在更广泛的社区中——超越Recursion和MIT——有许多贡献者从不同角度参与其中，他们认为社区挺身而出创造我们都能受益的工具确实具有重要价值。

You know, where are the areas that we need to be improving, and maybe somebody now picks up that area and thinks how to do it better. And yet the last point that I want to say that there were contributors on, you know, in the broader community beyond Recursion, beyond MIT, that contributed various aspects and thinks this is really a value that community steps up to create the tools that we all can benefit from.

我完全同意。你提到的关于拥有远见的观点——我们共同完成了许多事情，并且能将这种远见转化为实际影响。比如这个湿实验室与干实验室结合的愿景，你在整个历程中谈到如何真正实现实验操作、搭建桥梁、建立越来越可预测的模型，这些模型能预测并推荐你应该进行的实验。

I totally agree. And, you know, you said something about like having a vision. I mean, we have done so many things together and also just being able to take that vision and create impact. You know, the vision of this wet dry lab, which is so you talked about through the journey in terms of really being able to do that experimental, to build those bridges, to build and having models that are more and more predictable. The models then you know, predict and recommend what experiments you should do.

这正是Recursion的愿景所在，不仅是愿景，更是实践中正在发生的事情。我认为这是技术与生物技术、化学技术、科学、医学融合的平衡之道，对两个学科都保持尊重。人们总问我‘纳吉特，谁会胜出？’我觉得，谁能理解这种文化、这种愿景，并真正推动其产生影响力，

That's really where the vision at Recursion is the same, not just the vision, actually what's happening in practice. And I think that is that balanced approach of tech and biotech and chemistry, tech and science, tech and medicine coming together. And I think respect for both disciplines. And people always ask me, Najat, who's going to win? I'm like, who can figure out that culture, that vision, and then actually driving to impact?

谁就会成为这个领域的领导者。更多精彩还在后面。好的，在最后环节，我想问问接下来是什么，蕾吉娜？那会是什么？

That's going to be the leader in the space. So more to come. Okay. In the last segment, now I want to go to what's next, Regina. What is that?

我们已经讨论了一些正在探索的化学领域，部分模型正变得通用化。在你考虑下一个前沿领域时，你认为应该专注于什么方向？或许可以透露下你们实验室已经在深入研究的方向？

We've talked a little bit about chemistry and areas that are being figured out. Some of these models are getting commoditized. What is that next frontier as you think about an area to work on? Maybe a sneak peek as to where you're already properly working on, looking under the hood of Regina's lab.

最近有些特定话题特别吸引我的注意力，挥之不去。这源于我为ARPA H（现称ADAPT）撰写的一项资助计划。最让我振奋的是这个项目的愿景——旨在帮助转移性癌症患者。我不知道听众中有多少人接触过这类患者，但必须说这种处境极其艰难。当患者是原发性癌症时，尚有治疗方案可供选择，经历系列治疗后可能迎来隧道尽头的曙光——实现临床治愈。但对转移性患者而言，他们注定无法摆脱癌症，只是在与时间赛跑。他们的生存期充满变数，要承受何种副作用，如何与疾病抗争都是未知数。

There are specific topics that I recently that kind of caught my attention and doesn't leave it, was initiated by a grant that I wrote for ARPA H, which is called ADAPT. And I was most excited based on the vision of this programme, which aimed to help metastatic patients. So I don't know how many of the listeners seen those people, but, you know, this is a terrible condition to be, because, you know, when you have an initial cancer, you know, there is a treatment, you go through this growing treatment, but this is like, you know, the light at the end of the tunnel, you can be cancer free. But the people who are metastatic, it is clear that, you know, they are not going to get out of cancer, that they're just running race against time. And it's really questionable for how long they would live, you know, what kind of side effects, how they're going to be dealing with this disease.

这确实非常残酷。就像被疾病囚禁，它主宰着你的一切。我们知道对于原发性癌症，NCCN通常有明确的治疗指南，无论你去顶级医院还是社区诊所都有章可循。但转移性癌症却存在巨大差异：初始治疗或许有指南，但更多时候充满变数——谁能接受免疫治疗？是否有效？采用何种方案？何时终止治疗？存在无数疑问，治疗过程宛如手工定制，结果也千差万别。

It's really, really dramatic. It's like being jailed in your disease and which determines everything. Now we also know that depending, you know, when you go for primary cancer, most of the time NCCN has very clear guidelines what So treatment you're going to it doesn't matter if you go like to the top hospital or you go to community clinic, there is a guideline. However, when patients get to the metastatic cancer, there's actually a lot of variation, because maybe for initial treatment there is a guideline, but most of the time it's kind of there is a lot of possibilities, you know, who gets immunotherapy, is it going to be effective, what are the treatments, you know, when do you stop it? You know, there are so many questions and this is truly artisanal and the outcomes are very different.

为何会这样？因为转移性癌症突变速度极快，对吧？病情瞬息万变。即便初期用药有效，但不知何时（这个时间根本无法预测）患者就会产生耐药性，治疗随即失效。接着该换哪种药？新药又能维持多久？

Why does it happen? Because when the cancer is metastatic, metastatic, it mutates very fast, correct? There's a lot of changes happen. So if somebody put you on the drug at the beginning, within some time, and we don't even know how long it will take, the patient is going to develop resistance and then the treatment is ineffective. And then which one do you give and for how long it's going to be effective?

这个理念其实超越癌症范畴，因为耐药性普遍存在于多种疾病。关键在于：在初始治疗时就能预测患者会产生何种耐药性、何时出现，从而选择最佳治疗方案。目标是让患者在副作用最小的情况下获得最长生存期，并规划好二线、三线治疗方案，尽可能延长病情控制时间。这本质上是个机器学习问题——基于患者状况预测治疗持续时间与最终结局。因为某些耐药状态尚可应对，而有些则无计可施，我们必须避免患者陷入后者。这需要融合化学（理解药物作用机制）与生物学（阐明耐药原理）来实现精准预测。

So the idea is, can your and this is more general than cancer, because you know resistance happens in many different diseases then you predict when you put the patient in the first treatment what kind of resistance the patient is going to develop on the treatment, when they're going to develop this resistance, and select the best possible cure. So that the patient stays for the longest time without minimal side effects, then what will be the second line and the third line? So that you kind of control this condition for longest possible time. So it's pretty much a machine learning problem because you have a condition and you try to predict for how long this patient is going to stay on the therapy and what will be the end, because there are some, you know, states of resistance that we can treat and there are some that you cannot, so you don't want to bring them to this point. So this is a place where you need to bring both chemistry, because you need to understand what the drug is doing, and also the biology to make this type of prediction and to understand the mechanisms.

这正是项目最激动人心的部分——将我们最热衷的转化医学付诸实践。目前MD安德森癌症中心和希望之城（我想还有UNC）正在开展临床试验，通过机器学习算法为患者选择二线治疗方案，当然这些算法会经过严格验证。想到我们能真正改变患者命运，并破解当前传统技术难以企及的癌症耐药性与生物学机制，就令人振奋不已。

So this is like the most exciting part of this project, talking about translation, our favourite topic, that actually there are ongoing clinical trials in MD Anderson's City of Hope, and I think it's UNC, where they recruit patients to participate in these trials, where the selection of the second line treatment will be done through this machine learning algorithm, which of course would be tested and validated. It's exciting that we can really make a difference in somebody's life and also understand something about resistance and biology of cancers that today we have very limited understanding using conventional techniques.

Regina，你研究的正是最根本的问题——无论是结合亲和力预测还是其他方面。在化学领域，虽然人们为研发二线、三线药物付出巨大努力，但看看中位无进展生存期，往往只能延长几个月。总有人问我原因，虽然不能说是唯一因素，但核心问题之一就是无法预判耐药机制的出现。

Regina, you're working on the most fundamental issue, like in both binding affinity predictions, so, so fundamental. And I would say in chemistry, like if you look at second line, third line, right, people are working really hard to make drugs. But sometimes you look at the median progression free survival, it's a few month benefit incremental. And people always ask me why. It's because you cannot, well, I shouldn't say that's the only reason, but one of the core reasons is not being able to anticipate what resistance mechanisms will emerge.

新突变、代偿通路等等。看看生物学教材里的通路图，简直像漫画般简略。认真想想，我们确实需要做得更好。这也正是我在Recursion公司感到振奋的原因。

New mutations, compensatory pathways and so forth. And if you look at books in biology, I mean, pathways are still drawn out. It's a caricature. I mean, just for a second, if you think about that, we need to do better. And I think, you know, this is what also gets me excited at recursion.

这些生物学图谱中，若能更好地理解那些我们之前讨论的暗区，理清通路间的连接方式——这些如同高速公路般的网络，就能预见关键决策点，从而更早研发出更好的药物。我认为你们的工作非常出色。要知道，癌症影响着数百万患者，但其他疾病领域同样存在耐药性问题。这是普遍现象。虽然我们习惯按治疗领域划分，但老实说，生物系统的关联性远超我们的认知。

It's these maps of biology, where if you better understand those dark spots, right, we were talking about before, and you can connect how pathways, these highways are connected, then you can actually start to see what are the forks in the road you want to anticipate and build better drugs earlier. And I think the work that you're doing is fantastic. And look, cancer affects so many millions of patients, but there's resistance in so many other disease areas as well. It's true for anything. You know, we try to silo it by therapeutic areas, but quite frankly, biology is more connected than we give it credit for.

Regina，这次对话太棒了。不得不说，虽然我们交流很多，但这次谈话仍让我学到新东西。最后快速问个问题：观看视频的观众中，有人正尝试进入AI领域、推动技术应用或是全球各地学习的学生。能否给一条建议和一条警示？

So Regina, love this conversation. And I have to say, we've talked so much, but I learned even in this conversation new things. Just the last, very quick. There are going be people watching this that are trying to get into the AI space, trying to drive adoption or learning students all over. What is one piece of advice and also one piece of caution?

你见证过无数技术革新，总是能先人一步预见趋势。对于今天的观众，你会给出怎样的见解？

You've seen so many innovations turn over and you're somebody who sees around the corners early on. So what would that be for those that are watching today?

与五到十年前不同，如今有大量工具和框架大幅降低了入门门槛，尤其在临床领域。无需深厚基础，你就能使用这些工具了解其运作原理。不必成为计算机专家，也能利用现有工具处理图像、尝试各种框架。我想强调的是：即便你非科班出身，也能参与其中——未必是开发新工具，但可以成为工具使用者，或在机构内推动技术落地。希望大家明白，我们都能以不同方式做出贡献。

So today, in contrast to maybe five or ten years ago, there are a lot of tools and a lot of frameworks that make entrance much easier, especially in clinical spaces. So with very little kind of background learning, you can be able to utilise this tool to see how they work. You don't need to be computer scientists to utilize existing tool, take some images, to try lots of frameworks and things to try. So I would definitely say you shouldn't even if you are, again, not educated in this area and you're not a computer scientist with like very little, you can be part and it doesn't mean that you will be a developer of new tools, but you can be the user of new tools or play a role in adoption of these tools in your organization. So I want to give this encouragement, the message that we all can contribute in different ways.

关于警示，我们必须清醒认识技术的边界。最近有人告诉我，某美国供应商向世卫组织声称拥有无需人类介入就能治愈癌症的AI系统。我不禁想问：难道他们现在有能自主完成手术和放疗的机器人外科医生？这类空中楼阁式的幻想比比皆是。因此既要保持信心，也要严格审视工具的实际能力，用逻辑思维和实证依据来判断其真实性。

Now, in terms of the word of caution, again, we should be very cognizant of what we can and cannot do. And also, you know, in the world where everybody comes and says, you know, somebody yesterday was telling me that they had a conversation with WHO where one of the American providers claimed that they have an AI system that can cure cancer without human presence. I was just trying to imagine, do they have now a surgeon that can do the surgery and radiation? So, you know, there are a lot of kind of castles in the air that or emperor without clothes. So one should, you know, be positive, believe in your power, but be very critical of what actually the tool do and rely on your logic and on the provided evidence where it really works.

生命科学工作者往往兼具希望与怀疑精神——做实验时若缺乏质疑，任何波动都会被误认为是诺贝尔奖级发现。将这种思维带入AI领域能帮助我们明辨是非。这已成为新常态，我们都能以不同方式受益并参与其中。

And people in life sciences, I think, have some hope and skepticism because whenever you're doing the experiment, you need to be skeptic. Otherwise, you know, every blip would sound like a next Nobel Prize. So bringing this idea into, you know, the connection to AI, I think it's something that can help us to navigate. And again, this is what becomes our new reality, and we all can benefit from it in different ways and definitely be part of this new reality.

说得好极了。积极尝试，但在发表重大主张前务必确保有实质依据，对吧？要有真材实料支撑观点。太精彩了。Raghena，谢谢你。

Great, great points. Lean in, try it out and make sure you have substance before you start making big statements, right? Have substance to back it up. I love it. Raghena, thank you.

这是我们第一期播客。你能来真是太棒了。再次感谢你抽空参与。

This is our first podcast. It's so great to have you here. Thank you again for the time.

真的非常享受和往常一样与你交谈。谢谢，Regena。

Really, really enjoyed talking to you as usual. Thanks, Regena.