In mid‑2025, the biotech world reached a major milestone: for the first time ever, a cancer drug created entirely via AI is officially being tested on humans. After years of hype about how AI could transform medicine, we're finally seeing it in action. 

On June 18,  Insilico Medicine reported dosing the first patient in a global, multicenter Phase I trial of ISM3412, a generative-AI-designed MAT2A inhibitor targeting MTAP-deleted solid tumors. Meanwhile, Isomorphic Labs —Alphabet’s spinout powered by DeepMind’s AlphaFold 3—is poised to begin its own oncology trials shortly. 

This breakthrough marks a big shift: from lengthy, empirical drug design pipelines to AI-guided discovery. These aren’t just AI-assisted ideas. These drugs were discovered, designed, and refined using algorithms trained on years of molecular data. 

The result? Potential cancer treatments that could hit targets traditional drug discovery missed or took far longer to find. This isn't just a medical milestone. It could completely change how fast, how affordable, and how personally we treat cancer in the future.

Pipeline Disruption: Insilico and ISM3412

Insilico’s trial investigates ISM3412, an oral, selective MAT2A inhibitor powered by their Chemistry42 platform. This drug is designed to exploit synthetic lethality in MTAP-deleted cancers—including lung, bladder, and pancreatic tumors— where MTAP is common. This means the drug is built to attack certain cancer cells that are missing a key gene called MTAP (a gene that helps normal cells stay healthy). These types of cancers include some lung, bladder, and pancreatic tumors. The missing gene creates a weak spot and the drug is designed to hit it hard. If you hit the cancer in just the right spot (the weak point), the cell dies but healthy cells stay fine. It received FDA IND approval in April 2024  and thus began human dosing a year later. This is one of multiple AI-designed molecules now in early clinical testing.

Emerging Frontiers:  Isomorphic Labs & DeepMind.

Isomorphic Labs, spun out of Google DeepMind, is preparing its first human trials using oncology compounds generated by AlphaFold 3-driven models, which is an AI system with the ability to predict protein structures and model interactions between biological molecules, to accelerate and enhance drug discovery. 

According to company president Colin Murdoch, the approach enables virtual simulation of protein-ligand interactions, compressing traditional discovery timelines from years to months. He also has stated they are "getting very close" to initiating human trials and are currently staffing up for them. The company raised $600 million in its first major funding round, led by Thrive Capital in April 2025 forming partnerships with leading pharmaceutical companies such as Novartis and Eli Lilly.

Why AI Matters in Oncology

  • Traditional drug discovery averages 10–15 years, costs over $2 billion, and faces >90% failure rates. AI tools aim to cut cost and time by rapidly generating and filtering candidates.

  • Platforms like Insilico’s Chemistry42 and DeepMind’s AlphaFold accelerate predictive modeling (predict which molecules might bind to disease targets), drug-target identification and simulation (simulate how drugs interact with the body—without stepping into a lab), and molecular optimization (suggest tweaks to make those molecules more effective and less toxic).

  • Recursion and Xaira also reported multiple AI-originated drug candidates entering human trials—including oncology compounds.

Regulatory Frontiers 

As AI begins designing actual drug candidates, regulators like the FDA (U.S.) and EMA (Europe) are facing new challenges. Unlike traditional drug discovery, where scientists can explain each step, many AI-generated molecules come from complex models that even experts can’t fully explain. And that raises big questions about how we regulate something we don't fully understand.

Let’s break down the challenges and what’s being done about them:

  • Transparency: Many AI systems that design drugs—like AlphaFold or generative models from companies like Insilico Medicine—are based on complex algorithms. These systems analyze massive amounts of data to suggest which molecules might work as treatments. But even the developers can’t fully explain exactly how the AI arrived at a specific drug design. This is what people mean by a “black box.” It's like getting an answer without seeing the math.

 Why this matters:

Regulators like the FDA or EMA are responsible for public safety. They need to understand: why a drug was designed a certain way (rationale), where the idea came from (provenance), and whether it can be reliably recreated (reproducibility).

An Example: If an AI proposes a molecule to treat lung cancer, but no one can explain why it would work—or how to recreate it in another lab—that’s a problem for approval.

  • Addressing Algorithmic Bias: AI systems are trained on historical data. But if most of that data comes from one region or population, the AI could “learn” biased patterns.

Why it matters:

A drug that works great in clinical trials (say, mostly conducted in the U.S. or Europe) might not perform the same way in African, Asian, or Indigenous populations. Most regulators now require that datasets used for AI models are diverse and representative—so the drugs work safely across different age groups, ethnicities, and genders.

  • International Collaboration– Everyone Needs to Agree on the Rules: Drug development is international. A medicine developed in the US could be tested in South Korea and sold in Nigeria.

 The challenge: If each country has a different set of rules for AI in drug development, it creates confusion and slows progress. That’s why global agencies like the WHO, ICH, and OECD are working with countries to create shared standards and develop a harmonized international standard.

  • Ensuring Reliability and Safety: It’s not enough to approve a drug once and forget about it. Regulators need to track: How the drug performs in the real world, whether unexpected side effects emerge, and how future updates to the AI affect safety (yes, AI models can be updated like apps). This process is called post-market surveillance, and AI adds a new layer of complexity. Regulators now need tech experts alongside clinicians to monitor these drugs.
  • Evolving Regulations:
  • Regulatory bodies like the FDA and EMA are actively developing guidelines for AI in drug development. They are already holding public workshops on AI explainability, releasing draft guidance documents for AI use in drug development and collaborating with academia and tech companies to build smarter evaluation tools. They’re focused on data integrity, algorithm transparency, clinical validity.

Ethical Frontiers  

But There Are Risks and Unknowns

Yes, this sounds like the future and it probably is. But experts are also sounding the alarm about:

  • Bias in the algorithms: If you train AI on biased data, it may produce drugs that don’t work for everyone.
  • Lack of transparency: Some AI-designed drugs are based on models that aren’t peer-reviewed.
  • High costs: Even if drugs are discovered faster, will they be affordable? Will low-income countries benefit?

There is also no guarantee these drugs will work because still in very early stages.

Implications for Global Health & Pharma

  • For patients, faster, more precise drug development could mean earlier access to effective treatments.
  • For Research and development, AI may usher in a shift from empirical pipelines to predictive modeling, reducing costs and increasing success rates.
  • For low-resource settings, equitable access hinges on pricing, licensing, and inclusion in global trial networks but not guaranteed.