基于transformer的AI正把药物发现从湿实验转向更多计算：Insilico Medicine在2019年用AI找到靶蛋白并提出分子，18个月就得到候选药物rentosertib，相比常见的4.5年更快。其管线已有40多种AI开发药物；有预测称该领域年投资将从2025年的38亿美元升至2030年的152亿美元，而IQVIA统计2024年公布了12笔交易、总价值100亿美元，另有Eli Lilly与Nvidia合作建设“最强”超级计算机以加速研发。

在“90%进入临床的候选药会失败、成功药物开发成本约28亿美元”的行业经济学下，即便小幅提效也很关键。行业报告显示AI把临床前阶段压到12–18个月（此前为3–5年），且AI设计药物在安全性试验中的成功率达80–90%（历史为40–65%），从而把全流程成功率提高到9–18%（此前为5–10%）；同时AI可在不接触试管的情况下筛选“数百亿”小分子，AstraZeneca称速度提升到原来的两倍，并表示其小分子发现管线中90%以上已由AI辅助。

AI也在缩小并加速临床试验：通过“合成患者/数字孪生”作为匹配对照，AI从既往试验数据预测自然病程，再为每位受试者生成特征一致的虚拟对照。Unlearn.AI在2025年的工作显示，这种方法可将一项早期帕金森病试验的对照组规模减少38%，并在另一项阿尔茨海默病研究中减少23%，在某些情况下甚至可能消除对照组并提高受试者获得治疗而非进入对照组的概率。

Transformer-based AI is shifting drug discovery from wet labs toward computation: in 2019 Insilico Medicine used AI to find a target protein and propose molecules, reaching the candidate rentosertib in 18 months versus a typical 4.5 years. It now has 40+ AI-developed drugs in its pipeline; one projection puts annual investment rising from $3.8bn in 2025 to $15.2bn in 2030, while IQVIA counts 12 deals announced in 2024 worth $10bn, and Eli Lilly partnered with Nvidia to build a “most powerful” supercomputer to speed R&D.

With a 90% clinical-trial failure rate and an estimated $2.8bn cost per successful drug, even modest efficiency gains matter. Reports suggest AI cuts preclinical timelines to 12–18 months (from 3–5 years) and lifts safety-trial success to 80–90% (from 40–65%), raising end-to-end success to 9–18% (from 5–10%); it can also screen libraries of tens of billions of small molecules in silico, and AstraZeneca says this is twice as fast and that over 90% of its small-molecule discovery pipeline is AI-assisted.

AI is also shrinking and accelerating clinical trials via synthetic patients (digital twins) as matched controls, learned from past trials to predict untreated disease trajectories. Work published in 2025 by Unlearn.AI suggests control arms could have been reduced by 38% in an early Parkinson’s trial and by 23% in an Alzheimer’s study, potentially eliminating controls in some cases and increasing the chance participants receive the experimental treatment rather than placebo.

Source: An AI revolution in drugmaking is under way

Subtitle: It will transform how medicines are created—and the industry itself

Dateline: 1月 08, 2026 05:48 上午