人工智能在加速药物发现中展现出潜力，但目前面临数据质量与完整性的重大瓶颈。药物开发成本每九年翻倍增长，平均耗时十至十五年，单个项目成本达十亿至二十五亿美元，失败率超过百分之九十 [1]。AI技术可用于靶点筛选和候选化合物设计，但仍无法可靠预测药代动力学和可开发性 [1]。
影响AI药物发现效能的关键问题在于数据的系统性缺陷。现有公开数据集存在显著的发表偏差，研究人员主要分享成功结果而隐瞒失败实验 [1]。荷兰微生物学家Elisabeth Bik的研究发现，约百分之四的生物医学论文包含重复或篡改的图像 [1]。截至撰文时，尚无药物主要通过AI驱动设计获得美国食品药品监督管理局完全批准，但业界预计未来两至三年内会出现这样的案例 [1]。
未来的解决方向是建立集成的自动化实验室系统，实现计算与物理实验的闭环反馈，以确保AI模型基于完整、高质量的实验数据进行训练与优化 [1]。与此同时，前沿AI模型的训练成本持续高涨，根据斯坦福研究自二零一六年以来每年翻倍增长 [1]。
Artificial intelligence holds promise for accelerating drug development by streamlining compound screening and optimization, potentially reducing both costs and risks inherent in the process [1]. However, the field faces significant obstacles that threaten to undermine AI's effectiveness in this critical application. The pharmaceutical industry currently spends 10 to 25 billion dollars and requires 10 to 15 years on average to bring a drug to market, with failure rates exceeding 90 percent [1]. While AI can assist in target selection and candidate compound design, it remains unable to reliably predict drug pharmacokinetics and developability [1].
A fundamental challenge lies in data quality and integrity. Published datasets suffer from publication bias, as researchers predominantly share successful results while withholding failed experiments [1]. This selective reporting compromises the training data available to AI systems. Research by Dutch microbiologist Elisabeth Bik found that approximately 4 percent of biomedical papers contain duplicated or manipulated images, based on 2016 data [1]. These data integrity issues directly limit AI's capacity to learn robust patterns applicable to real-world drug discovery scenarios. As of the time of writing, no drug has received full FDA approval based primarily on AI-driven design, though such approvals are anticipated within the next 2 to 3 years [1].
Moving forward, the pharmaceutical industry is exploring integrated automated laboratory systems that would create closed-loop feedback between computational models and physical experiments [1]. Additionally, advances in AI itself come with escalating demands: training costs for cutting-edge models have doubled annually since 2016, according to Stanford research [1]. Addressing these technical and economic challenges will be essential for realizing AI's full potential in drug discovery.
