How Far Has AI Drug Discovery Actually Come? The Reality of $100 Billion in Investment and Zero Approvals
機械翻訳 / Machine-translated
A
AI Friends
1日前・13分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
"AI drug discovery"—the use of artificial intelligence to design new medicines—has attracted enormous attention in recent years. But behind the dazzling headlines, how far has it actually progressed? This article examines the track record and challenges of AI drug discovery.
As of September 2026, not a single drug designed or discovered by AI has received approval from the FDA (U.S. Food and Drug Administration), the EMA (European Medicines Agency), or Japan's PMDA (Pharmaceuticals and Medical Devices Agency).
This is a striking fact. The reason: global investment in AI drug discovery has surpassed $100 billion (approximately ¥15 trillion), and between 2022 and 2026 there was widespread expectation that AI would dramatically transform pharmaceutical development.
Currently, 173 AI-derived programs are in clinical development, 15 of which have already advanced to Phase III (the final stage of clinical trials). In other words, the pipeline—the roster of drugs under development—undeniably exists.
Industry analysts predict that the first approval of a fully AI-designed drug will arrive between 2027 and 2028. However, this forecast has been pushed back every year for the past three years. Initially, approvals were expected in 2024 or 2025.
Interestingly, AI drug discovery demonstrates very high success rates in Phase I trials (early clinical trials that assess safety)—specifically, 80–90%, far exceeding the 40–55% success rate seen with conventional drugs.
However, the picture changes dramatically in Phase II (mid-stage trials that assess efficacy). The success rate drops to around 40%, which is barely different from the 37% recorded by conventional drugs.
The final probability of approval also remains in the range of 8–12%—the same as the industry average.
Why does this happen? The answer lies in the nature of the data on which AI is trained.
AI is trained on preclinical data—experimental results from test tubes and animal studies conducted before a drug is administered to humans. As a result, it is good at generating "drug-like molecules": compounds that are appropriately absorbed in the body, resistant to degradation, and low in toxicity.
Because Phase I primarily tests safety, the "drug-like molecules" that AI produces tend to pass with ease.
Phase II and Phase III, on the other hand, test whether a drug actually works against real human disease. Having drug-like properties is simply not enough; the molecule must produce real effects within the complex machinery of the human body.
Current AI models cannot predict this "real-world efficacy in humans." There is a large gap between preclinical data and human outcomes—a problem known as translational biology. AI is failing to bridge that gap.
Despite the enormous sum of $100 billion invested, why has not a single drug been approved?
According to experts, there is another structural problem: AI systems built for speed often lack the documentation infrastructure required for regulatory submissions.
Obtaining drug approval requires detailed records of the entire development process and proof of data reliability. Yet AI platforms designed with efficiency as the top priority frequently treat this documentation process as an afterthought.
In other words, there is yet another gap—between "AI that can find drugs quickly" and "a development process that produces drugs in a form regulators can actually approve."
For this reason, 2026 is described as a year in which validation and disappointment will arrive in roughly equal measure. Some companies are beginning to deliver results, but a similar number of cases are expected to fall short of expectations.
Not everything is disappointing. There are companies that are genuinely producing results.
The most advanced example is rentosertib, developed by Insilico Medicine. A TNIK inhibitor (a drug that suppresses the activity of a specific protein in the body), it entered Phase III clinical trials on July 7, 2026, as a treatment for idiopathic pulmonary fibrosis (a serious disease in which the lungs become scarred and stiff).
This is the most advanced case of an AI-designed drug reaching the final stage of trials.
Generate:Biomedicines has also reached Phase III with a treatment for autoimmune disease. Iambic Therapeutics has advanced an AI-designed drug for ALS (amyotrophic lateral sclerosis, a serious neurological condition) into Phase II.
What these companies have in common is that they are not using AI as a complete replacement for conventional methods, but instead introducing it gradually as a tool that complements existing processes.
So, was AI drug discovery simply hype? The answer is no—but a realistic assessment is necessary.
AI is genuinely accelerating the early stages of drug discovery—specifically, the computational search for candidate compounds. There are reports of tasks that once took years being completed in months.
At the same time, AI's accuracy drops when data is not standardized or when data volumes are small. AI can only show its true power when large amounts of high-quality data are available.
Furthermore, for modalities beyond small-molecule drugs (small chemically synthesized molecules)—such as antibodies (drugs based on immune-system proteins), nucleic acids (drugs based on DNA or RNA), and cell therapies—current machine learning alone is insufficient. Complex approaches combining molecular simulation, mathematical modeling, and other techniques are required.
Experts emphasize that it is essential to avoid over-relying on AI and instead use it in combination with established scientific knowledge.
AI drug discovery genuinely has the power to dramatically accelerate the early stages of the drug development process. However, many hurdles remain before these candidates become medicines that cure human disease. Correctly understanding the gap between expectation and reality is indispensable for future progress.
This article is a cross-post from AI Friends.
