AI制药新纪元：诺奖得主哈萨比斯的“万病皆可医”愿景

内容概要

德米斯・哈萨比斯正全力开发能攻克一切疾病的AI系统，其AlphaFold蛋白质结构预测技术为他赢得了2024年诺贝尔化学奖。他创办的AI制药企业Isomorphic Labs即将公布首批候选药物，该公司已启动19个研发项目，主攻癌症、心血管疾病和免疫学三大方向，并与多家国际制药巨头建立合作。Isomorphic Labs采用DeepMind独家新药设计模型IsoDDE，其性能超越市面主流开源方案，出于生物安全考虑暂未开源。哈萨比斯表示，他的目标是创造终极科研利器，权势与财富不过是副产品。

双语阅读

SIR DEMIS HASSABIS has long had an ambitious plan: to develop artificial-intelligence models capable of curing all dis- eases. His journey towards realizing it has been a circuitous one—few career advisers would recommend teaching a machine to play Atari games as a good first step—but his successes are hard to dispute. As the boss of Google DeepMind, a world-leading AI lab, he shared the Nobel prize for chemistry in 2024 for designing AI models that could predict how proteins fold. Near- ly two years on, how close to his dearest goal does he think he is?

德米斯・哈萨比斯爵士长久以来怀揣着一个宏大愿景：打造能治愈所有疾病的人工智能系统。这条追梦之路充满曲折——鲜少有人会将教机器玩雅达利游戏视为职业规划的起点，但他的成就无可辩驳。作为全球顶级AI实验室谷歌DeepMind的领导者，他凭借蛋白质结构预测模型斩获2024年诺贝尔化学奖。时隔两年，他如何看待自己与这一终极目标的距离？

Speaking to The Economist's "Inside Tech", a video show, Sir Demis says work is proceeding according to plan. Five years ago Google spun off Isomorphic Labs (with Sir Demis as its boss), an AI-powered pharmaceutical firm, with a remit to use DeepMind's protein-structure tech, Alpha- Fold, to find novel medicines. After a few years of tooling up, the lab is about to announce its first candidates. Sir Demis says it has 19 programs spread across three main research areas: cancers, cardiovascular conditions and immunology.

在接受《经济学人》视频栏目《科技内幕》访谈时，德米斯爵士透露项目正稳步推进。五年前，谷歌剥离出AI制药板块，成立同构实验室（Isomorphic Labs），由哈萨比斯执掌帅印，旨在运用DeepMind的AlphaFold蛋白质预测技术开发创新药物。历经数年技术积累，该实验室即将发布首批候选药物名单。据哈萨比斯介绍，公司现已铺设19条研发管线，集中攻坚癌症、心脑血管疾病及免疫学三大治疗领域。

Those programs, which include partnerships with big pharma companies including Eli Lilly, Novartis and Johnson & Johnson, as well as internal projects, are in- tended as the first step towards a generic technology that could tackle any medical condition thrown at it. Once the underlying technologies are developed, says Sir Demis, "Like with AlphaFold, you can ap- ply them extremely quickly". AlphaFold it- self took six years of work to predict its first protein structure, and then one year to follow up with what he describes as the structures of "all 200m proteins known to science". He hopes a similar speedup will happen inside Isomorphic.

这些管线既涵盖与礼来、诺华、强生等制药巨头的联合开发，也包括内部独立研究，标志着向通用医疗平台迈进的首个里程碑——该平台未来将具备治疗各类疾病的能力。德米斯爵士指出，待基础技术成熟后，"如同AlphaFold般，这些工具可迅速推广部署"。AlphaFold从立项到首次成功预测蛋白质结构花了六年，随后仅用一年便完成了他所说的"科学界已知的2亿种蛋白质"结构图谱。他期待同构实验室也能迎来类似的效率飞跃。

The company is benefiting from advance access to DeepMind's work. In February Isomorphic announced that it was working with a new, proprietary version of AlphaFold—the first to be reserved for in- ternal use. IsoDDE, as it is called, can be used to predict various properties of potential drugs, including their binding affinity—a measure of how strongly they link to proteins and a proxy for their eventual efficacy. At that particular task, according to Isomorphic's published results, it out- shines state-of-the-art open-source alter- natives like Boltz-2, developed at the Massachusetts Institute of Technology.

该企业享有DeepMind最新成果的优先使用权。今年2月，同构实验室宣告采用全新AlphaFold专有版本——这是该技术首次仅限内部部署。该体系名为IsoDDE，可预判候选药物的各项参数，涵盖药物结合亲和力，即药物分子与蛋白质的结合强度，这是评估药效的关键指标。同构实验室发布的测试数据显示，在此项性能上，该系统超越了麻省理工学院开发的Boltz-2等现有顶级开源竞品。

IsoDDE covers more ground than AlphaFold, says Sir Demis, as it is able to predict the biochemical interactions of proteins in a way that its predecessor was not designed to do. It is such predictions that will differentiate a commercially valuable "drug design engine"—the "DDE" in the name—from an academic tool. But it isn't just the pursuit of profit that has kept the model behind closed doors. "There are trade-offs in terms of biosecurity and bio- safety. If you just make that freely avail- able, if a bad actor were to get hold of it, they could repurpose it for harmful ends."

据德米斯爵士介绍，IsoDDE的功能广度显著超越AlphaFold，其能够预测蛋白质的生化交互作用，而这正是前代技术的设计盲区。正是这种预测能力，使其成为具备商业潜力的"药物设计引擎"（DDE名称由此而来），与学术工具有本质区别。然而，该模型不对外开源，并非只为商业考量。"这其中存在生物安全与生物安保的取舍。若完全开放，一旦被恶意使用者掌控，恐将用于有害目的。"

He is not alone amongst his peers in unilaterally decreeing what powers humanity can be trusted with. This week Anthropic, another AI lab, said that its latest model, Claude Mythos, would only be accessible to cybersecurity experts, citing the hacking risk a public release could create. Yet it would be fair to ask Sir De- mis—who are you to make that call?

不只是他独自裁决人类应掌握何种技术能力。本周，AI研究机构Anthropic亦发布声明，其最新Claude Mythos模型仅向网络安全领域专家开放，理由是公开部署可能招致黑客威胁。但公众难免质疑：德米斯爵士凭何做出如此决断？

"People have to make their own decisions about the lab leaders," he says. "There's a lot of information out there now about each of the leaders' different approaches...You also need to think about people's motivations, and why they got into AI." His own motivation? To create "the ultimate tool for science". To hear him tell it, power and riches are just an unfortunate side product of being a humble scientist working towards his Nobel prize.

"公众有权对实验室领袖的决策做出评判，"他答道，"当下各领导者迥异的管理哲学已充分透明……人们也需审视从业者的本心，及其踏入AI领域的原动力。"那他的本心？是构建"终极科研利器"。在他看来，权势与财富不过是一位专注科研、矢志摘取诺贝尔奖的学者，在求索途中意外所得的副产品。