How does the FDA view AI-designed drugs? A comprehensive overview of the regulatory framework, review logic, and industry impact.

Beyond wet-lab experimental data, several other categories of data also require traceability—and each comes with its own pitfalls: Here’s a particularly common trap: if computationally generated data is used to train another AI model, traceability must extend further upstream. For example, if you use protein structures predicted by AlphaFold to train a molecular docking model, it’s insufficient to merely state that 'the training data comes from AlphaFold-predicted structures.' You must also specify precisely which version of AlphaFold was used, what confidence-score threshold was applied, and whether low-confidence predictions were filtered out. Otherwise, you’re effectively 'training one model using the output of another,' causing errors to compound across layers—prompting the FDA to question the entire model’s reliability. VI. Several important questions not addressed in the original text Having reached this point, several critical issues remain unaddressed by the earlier regulatory framework—but are vital to the industry. I’ll clarify them here. 1. How should adaptive or continuously learning AI models be regulated? Many AI models undergo retraining and version updates after deployment, incorporating new data over time. When such an update occurs, must the model be re-validated? Must the FDA be notified? The current framework mentions only 'lifecycle management' and states that 'major updates require re-validation'—but offers no clear definition of what constitutes a 'major update,' nor specifies the procedural requirements following an update. This remains a substantial gray area. For instance, if your model undergoes minor fine-tuning monthly with newly acquired data—and performance improves by just 2%—does that qualify as a major update? Requiring full re-validation for every such change would negate the core advantage of continuous learning; yet ignoring updates entirely risks model drift and associated safety hazards. In practice, the industry currently adopts a 'change-tiering' approach: minor updates are internally documented, while major ones trigger formal re-validation. However, each company sets its own criteria—no standardized FDA guidance exists yet. 2. Can molecules generated by AI be patented? This is a major legal minefield. Both the U.S. Patent and Trademark Office (USPTO) and courts currently hold that 'AI cannot be named as an inventor.' Yet whether molecules *assisted* by AI can be patented by human inventors remains an evolving standard. The prevailing consensus is: if humans perform substantive selection, optimization, and experimental validation on AI outputs, then humans may legitimately claim inventorship and file patents; however, if the molecule is fully auto-generated by AI—with humans merely pressing a button—the resulting patent’s enforceability becomes highly questionable. This issue is pivotal for AI-driven pharmaceutical companies: unstable patent rights over AI-generated molecules directly threaten their core intellectual assets. Most such companies currently emphasize human experts’ contributions in patent applications and position AI strictly as a 'tool'—not an 'inventor.' Yet whether this strategy will withstand future legal challenges remains uncertain. 3. Liability allocation If an AI predicts a molecule to be safe—but clinical trials reveal serious adverse effects—who bears legal responsibility: the pharmaceutical company, the AI software vendor, or the algorithm developer? No definitive legal precedent exists, nor is there any industry-standard contractual framework governing liability allocation. In practice, the pharmaceutical company—as the regulatory applicant—bears ultimate responsibility. It may seek contractual recourse against the AI tool provider—but only if liability terms are explicitly defined in the agreement. If an open-source model is used, there may be no legally liable party at all.

Who owns the patent for a drug developed using AI?

When biotech company Insilico Medicine used computational models to propose a promising anti-fibrotic drug candidate for pulmonary fibrosis, it enthusiastically declared in its press release that the molecule had been 'discovered' by its generative AI platform. Insilico Medicine is among numerous companies leveraging AI to rapidly identify novel drug candidates—aiming to uncover therapeutic compounds humans might never have conceived, thereby accelerating the development of new therapies. Today, generating molecular structures with AI has become as effortless as asking ChatGPT to draft a thank-you note. Yet when the company filed for a critically important patent to protect this entirely new chemical structure, it made no mention whatsoever of AI. Instead, the patent listed five human individuals—including CEO Alex Zhavoronkov—as the drug’s 'inventors.' This stark contrast reveals a curious paradox in intellectual property law: no matter how pivotal AI’s role may be in a given discovery, when it comes to legal inventorship, credit—and rights—belong exclusively to humans. U.S. courts reached this conclusion in a public-interest lawsuit spearheaded by Ryan Abbott, a partner at the Los Angeles law firm Brown, Neri, Smith & Khan. Abbott filed a test-case petition naming DABUS—an AI system—as the inventor. The invention in question was an improved food container whose complex geometric surface enhanced heat transfer and facilitated stacking. As no human participated in its design process, Abbott argued that the AI itself should be recognized as the inventor. This case could have sparked profound philosophical debates: Should AI possess legal rights? Where, exactly, does the 'eureka moment' reside when a superior invention emerges? But in 2022, the U.S. Court of Appeals for the Federal Circuit in Washington, D.C., ruled that such 'metaphysical questions' were irrelevant to the case. The court emphasized that U.S. law defines an inventor as an 'individual'—a term whose ordinary meaning refers exclusively to a human being. Since machines are not people, they cannot be inventors. The matter appeared settled. 'You must have a human inventor; otherwise, there is no invention—and therefore, no patent,' says Sarah Korman, a patent attorney who serves as Chief Commercial Officer and General Counsel at Isomorphic Labs—a company under Alphabet focused on AI-driven drug discovery. Speaking last year at the EmTech conference hosted by MIT Technology Review, Korman affirmed that existing laws will inevitably need to evolve alongside AI advancement. One reason is undeniable: no one disputes that AI *can* 'invent.'

Comment: The Molecule Came from AI. The Inventor Still Has to Be Human

1 day ago

ACS Committee Chair Christine E. McCain comments on intellectual property challenges in AI-driven drug discovery, asserting that human inventors remain essential for patent applications despite AI generating molecular structures.

cen

2026-10-26

Comment: The Molecule Came from AI. The Inventor Still Has to Be Human

Comment: The Molecule Came from AI. The Inventor Still Has to Be Human

FAQ