AI驱动药物发现:Isomorphic Labs如何把药研搬上计算机Powering AI-Driven Drug Discovery with Isomorphic Labs President Max Jaderberg
The Takeaway:药物发现最大的机会,是从现实世界中缓慢的经验试错,转向能达到实验精度的预测计算模型,从而在进入临床试验前大幅提升成功率并减少迭代轮次。
Max Jaderberg是Isomorphic Labs总裁及创始团队成员。这家由Alphabet支持的公司目标是解决所有疾病。他此前在DeepMind工作多年,参与早期生成模型、强化学习和AlphaGo/AlphaStar等项目,并在AlphaFold证明深度学习能在生物学有意义空间泛化后,于2021年共同创立了Isomorphic。
他指出制药行业新药批准数量长期持平,尽管论文数量暴增、分子模态增多,但成本从数亿美元升至15亿美元、周期从8-9年拉长到15年。原因在于整个行业仍高度依赖经验过程:提出假设、合成分子、在真实实验中测试,循环往复。生物学和分子设计都是超高维空间,试错只能触及低垂果实,复杂度再往上就难以触及。
AI的转折在于能把疾病生物学理解、分子设计全部搬到计算机上,建立达到实验精度的世界模型。一旦模拟足够准,现实世界就从“学习迭代”变成“验证”,失败率(尤其是临床毒性)会显著下降。动物模型对人类临床毒性的预测力大约只有30%,他更看好体外终点、器官芯片、类器官与AI模型的结合,直接建模人类毒性。数据策略上,他们主动生成专有数据以追求泛化,而不是依赖有偏的历史仓库。
公司内部管线聚焦肿瘤、免疫与炎症,同时与药企合作攻克对方十年未破的难题。模型本身是疾病、靶点、模态无关的。他强调:“如果五年后疾病生物学的预测理解没有发生巨大阶跃变化,我会非常失望。”更精准、更低毒性的分子还将打开真正的多药联用可能,因为目前很多组合受限于药物相互作用和肝毒性。
患者安全始终第一,但风险偏好因适应症而异——晚期癌症与慢性炎症完全不同。随着设计理解加深,成功率与安全性统计上的差异有望改变监管对药物上市的看法。
Max Jaderberg是Isomorphic Labs总裁及创始团队成员。这家由Alphabet支持的公司目标是解决所有疾病。他此前在DeepMind工作多年,参与早期生成模型、强化学习和AlphaGo/AlphaStar等项目,并在AlphaFold证明深度学习能在生物学有意义空间泛化后,于2021年共同创立了Isomorphic。
他指出制药行业新药批准数量长期持平,尽管论文数量暴增、分子模态增多,但成本从数亿美元升至15亿美元、周期从8-9年拉长到15年。原因在于整个行业仍高度依赖经验过程:提出假设、合成分子、在真实实验中测试,循环往复。生物学和分子设计都是超高维空间,试错只能触及低垂果实,复杂度再往上就难以触及。
AI的转折在于能把疾病生物学理解、分子设计全部搬到计算机上,建立达到实验精度的世界模型。一旦模拟足够准,现实世界就从“学习迭代”变成“验证”,失败率(尤其是临床毒性)会显著下降。动物模型对人类临床毒性的预测力大约只有30%,他更看好体外终点、器官芯片、类器官与AI模型的结合,直接建模人类毒性。数据策略上,他们主动生成专有数据以追求泛化,而不是依赖有偏的历史仓库。
公司内部管线聚焦肿瘤、免疫与炎症,同时与药企合作攻克对方十年未破的难题。模型本身是疾病、靶点、模态无关的。他强调:“如果五年后疾病生物学的预测理解没有发生巨大阶跃变化,我会非常失望。”更精准、更低毒性的分子还将打开真正的多药联用可能,因为目前很多组合受限于药物相互作用和肝毒性。
患者安全始终第一,但风险偏好因适应症而异——晚期癌症与慢性炎症完全不同。随着设计理解加深,成功率与安全性统计上的差异有望改变监管对药物上市的看法。
The Takeaway: The biggest opportunity in drug discovery is moving from slow empirical trial-and-error in the real world to predictive computational models that reach experimental accuracy, so far more design happens in silico and success rates jump when molecules finally enter patients.
Max Jaderberg is President and a founding team member of Isomorphic Labs, the Alphabet-backed company whose stated goal is to solve all disease. He spent years at DeepMind working on early generative models, reinforcement learning and game-playing systems such as AlphaGo and AlphaStar. Once AlphaFold showed that deep learning could generalize in a meaningful biological space, the team spun out Isomorphic in 2021 to build a dedicated culture and move fast on the harder problem.
He notes that new drug approvals have stayed roughly flat for decades even as scientific papers exploded and new modalities proliferated. Costs rose from a few hundred million to $1.5 billion and timelines stretched from eight or nine years to fifteen. The root cause is that medicine-making remains largely empirical: form a hypothesis, make the molecule, test it in the physical world, repeat. Both biology and molecular design are combinatorially vast; trial-and-error simply cannot search the higher-complexity regions once the low-hanging fruit is gone.
The shift Isomorphic is driving is to pull the entire discovery loop onto computers. Build world models that achieve experimental accuracy so that the real world is used mainly for validation rather than iterative learning. Clinical failures, especially toxicity, are where most value is destroyed because problems only become visible in patients. Animal models predict human clinical toxicity only about 30 percent of the time. Better routes combine in-vitro endpoints, organoids and organs-on-chips with AI predictors aimed directly at human outcomes. On data, the team generates proprietary datasets deliberately rather than relying on biased historical repositories, and has already observed strong scaling laws.
Internally they focus on oncology, immunology and inflammation while also partnering with pharma on the hardest problems those companies have worked for a decade. The underlying models are disease-, target- and modality-agnostic. Jaderberg’s benchmark: “I’d be really disappointed if we fast-forward even five years’ time [and] the predictive understanding of disease biology has not step-changed in a massive way.” Safer, more precise molecules should also unlock true multi-drug combinations that today’s toxicity and interaction profiles make impractical.
Patient safety remains non-negotiable, yet risk appetite correctly varies by indication. As design understanding improves and success rates become statistically different, the regulatory conversation itself is expected to evolve.
查看原文 →
Max Jaderberg is President and a founding team member of Isomorphic Labs, the Alphabet-backed company whose stated goal is to solve all disease. He spent years at DeepMind working on early generative models, reinforcement learning and game-playing systems such as AlphaGo and AlphaStar. Once AlphaFold showed that deep learning could generalize in a meaningful biological space, the team spun out Isomorphic in 2021 to build a dedicated culture and move fast on the harder problem.
He notes that new drug approvals have stayed roughly flat for decades even as scientific papers exploded and new modalities proliferated. Costs rose from a few hundred million to $1.5 billion and timelines stretched from eight or nine years to fifteen. The root cause is that medicine-making remains largely empirical: form a hypothesis, make the molecule, test it in the physical world, repeat. Both biology and molecular design are combinatorially vast; trial-and-error simply cannot search the higher-complexity regions once the low-hanging fruit is gone.
The shift Isomorphic is driving is to pull the entire discovery loop onto computers. Build world models that achieve experimental accuracy so that the real world is used mainly for validation rather than iterative learning. Clinical failures, especially toxicity, are where most value is destroyed because problems only become visible in patients. Animal models predict human clinical toxicity only about 30 percent of the time. Better routes combine in-vitro endpoints, organoids and organs-on-chips with AI predictors aimed directly at human outcomes. On data, the team generates proprietary datasets deliberately rather than relying on biased historical repositories, and has already observed strong scaling laws.
Internally they focus on oncology, immunology and inflammation while also partnering with pharma on the hardest problems those companies have worked for a decade. The underlying models are disease-, target- and modality-agnostic. Jaderberg’s benchmark: “I’d be really disappointed if we fast-forward even five years’ time [and] the predictive understanding of disease biology has not step-changed in a massive way.” Safer, more precise molecules should also unlock true multi-drug combinations that today’s toxicity and interaction profiles make impractical.
Patient safety remains non-negotiable, yet risk appetite correctly varies by indication. As design understanding improves and success rates become statistically different, the regulatory conversation itself is expected to evolve.