What AI Drug Discovery Could Mean for Rare Disease Families

For years, artificial intelligence in drug discovery sounded like a screen glowing in a clean room: elegant models searching chemical space, ranking targets, and promising to make biology more legible. That chapter is not over. But the more important story is beginning somewhere messier and more consequential. AI is moving from the screen into the laboratory.

The signal is not one announcement. It is the pattern. Eli Lilly has been building an NVIDIA-powered AI factory and supercomputer for drug discovery, and in January 2026 Lilly and NVIDIA announced a co-innovation lab backed by up to $1 billion over five years. The language around that partnership matters. This is not simply a faster search engine for molecules. Lilly and NVIDIA are describing a system in which computational dry labs, agentic wet labs, foundation models, robotics, BioNeMo tools, and manufacturing digital twins begin to learn from one another.

That is a different kind of infrastructure. Traditional drug development has often moved like a relay race: a hypothesis is handed from one team to another, then into a lab, then into a meeting, then back into a model or a new experiment. The emerging AI model is more circular. A system proposes an experiment, a robotic lab runs it, the result becomes new data, and the model returns with a sharper question. Human scientists are still essential, but their judgment moves closer to the center: deciding what question is worth asking, whether the assay is meaningful, and whether a promising signal is real biology or a beautiful mistake.

OpenAI's life-sciences work shows the same drift toward the real world. Its collaboration with Retro Biosciences on GPT-4b micro explored protein engineering around Yamanaka factors and reported large increases in stem-cell reprogramming markers in vitro. Its work with Ginkgo Bioworks pushed even further into closed-loop experimentation: GPT-5 connected to a cloud laboratory, tested more than 36,000 cell-free protein-synthesis reaction compositions across 580 automated plates, and reduced protein-production cost by 40 percent. That is not a chatbot answering questions about biology. That is a model participating in the design of experiments.

Anthropic is moving along a parallel track. Genmab announced a partnership with Anthropic to build Claude-powered agentic AI systems for clinical-development workflows. In April 2026, TechCrunch reported that Anthropic acquired Coefficient Bio, a stealth biotech AI startup with roots in Genentech's Prescient Design group, in a roughly $400 million stock deal. The frontier AI labs appear to understand that biology is not just another market for general-purpose assistants. It is a domain where the winning systems may need to combine language, reasoning, experimental design, private scientific data, and the discipline of laboratories that can prove or disprove the machine's ideas.

For rare disease families, the reason to care is not that AI sounds futuristic. It is that rare disease research has always been constrained by scarcity: too few patients, too little natural-history data, too few validated models, too little funding, and too many plausible paths that cannot all be tested. If AI-connected laboratories can help researchers choose better targets, generate stronger candidates, interpret variants, design better assays, and discard weak ideas earlier, the effect could be profound.

The early clinical numbers are intriguing. A 2024 open-access analysis in Drug Discovery Today found that AI-discovered molecules had an 80 to 90 percent success rate in Phase I trials, well above historic industry averages. That finding should be read with care. Phase I is not approval. It is mostly a test of safety, tolerability, pharmacokinetics, and whether a drug behaves enough like a drug to keep going. The same analysis found Phase II success around 40 percent on a limited sample, closer to historical norms. The wall between a safe molecule and a therapy that meaningfully changes disease is still high.

That wall is especially high in rare disease. A therapy may look promising in a dish and still fail because the disease model is incomplete, the dose cannot reach the right tissue, the endpoint is too blunt, the manufacturing process is fragile, or the trial cannot recruit enough patients. AI can help with each of those problems, but it does not erase them. It can accelerate the questions. It cannot excuse weak evidence.

The best version of this future is not a world in which machines replace the slow work of medicine. It is a world in which fewer families wait while science circles the wrong target. A world in which a foundation with limited money can make sharper bets. A world in which a promising therapy reaches the right experiment sooner, and a bad idea is allowed to die before it consumes years that families do not have.

That is the promise worth watching. Not magic. Not instant cures. A faster, more disciplined conversation between computation and biology. For rare disease families, that may be enough to change what hope feels like.