Eli Lilly Just Bet $2.25 Billion That AI Can Do What CRISPR Cannot

CRISPR has dominated the genetic medicine conversation for a decade , and for good reason. The ability to cut DNA at a precise location and either disrupt or replace a gene sequence transformed what was theoretically possible in treating inherited disease. But CRISPR has a fundamental limitation that its proponents have not always emphasized: it cannot easily insert large pieces of DNA. It can edit existing sequences efficiently, but inserting a complete functional gene , the kind of therapeutic payload required to treat conditions like Duchenne muscular dystrophy, sickle cell disease, or many metabolic disorders , remains extraordinarily difficult with CRISPR-based tools. That limitation is precisely the gap that Eli Lilly just paid up to $2.25 billion to close, through a partnership with a three-year-old AI biotech company that is designing enzymes that do not exist in nature.

What Actually Happened

On April 28, 2026, Profluent Inc. and Eli Lilly announced a multi-program strategic research collaboration to develop and commercialize custom AI-designed site-specific recombinases for genetic medicine applications. Under the terms of the agreement, Profluent will apply its AI protein design models to engineer and optimize recombinases , enzymes that can recognize and act on specific DNA sequences , targeting multiple genomic locations across a range of disease areas with severe unmet medical needs. Lilly will receive exclusive rights to advance selected recombinase candidates through in vivo research, preclinical development, clinical studies, and commercialization. Profluent will receive an upfront payment plus committed research and development funding, with the potential for up to $2.25 billion in development and commercial milestone payments plus tiered royalties on net sales, making this among the largest AI drug discovery partnerships ever announced.

Profluent was founded in 2022 and has spent its first years building what it describes as a protein language model , an AI system trained to understand the relationship between protein sequences and their functional properties at a level that allows it to design entirely novel proteins with specified characteristics. The company raised $106 million in its most recent funding round, co-led by Altimeter Capital and Bezos Expeditions, the family investment office of Amazon founder Jeff Bezos. The Lilly deal is Profluent's first major pharmaceutical partner agreement and its validation as a serious player in the increasingly competitive AI drug design space. The specific technology focus , recombinase design for large-gene insertion , is not Lilly's first investment in this area: in January 2026, Lilly signed a $1.12 billion collaboration with German biotech Seamless Therapeutics also targeting recombinase-based therapies, specifically for hearing loss.

Why This Matters More Than People Think

The conventional framing of this deal , "Lilly bets on AI for gene editing" , undersells both the technical ambition and the strategic logic. The category of genetic medicine that Profluent and Lilly are targeting is not incremental improvement on existing approaches. Site-specific recombinases capable of inserting kilobase-scale DNA sequences at precise genomic locations would represent a genuinely new therapeutic modality , one that opens disease categories that current gene therapy tools cannot effectively address. The difference between editing a single base pair, which CRISPR handles well, and inserting a functional gene of several thousand base pairs, which requires a different enzymatic mechanism entirely, is not a quantitative difference in capability. It is a qualitative difference in what diseases become treatable. If Profluent's AI-designed recombinases work as intended, the addressable disease space for genetic medicine expands significantly.

Lilly's strategic logic for pursuing this technology through AI-designed enzymes rather than through conventional discovery or through competing gene therapy modalities deserves scrutiny. Traditional recombinase discovery involves screening libraries of naturally occurring enzymes and then iteratively engineering them toward the desired target sequence , a slow, expensive, and often unsuccessful process. Profluent's approach uses AI protein language models to design recombinases computationally, starting from the desired functional specification and working backward to a protein sequence that should achieve it. This is not the same as in silico screening of natural variants. It is generative protein design , the AI creates candidates that have never existed in biology. The potential speed and cost advantages over traditional approaches are substantial if the AI-designed candidates demonstrate activity in laboratory and in vivo testing, which remains the critical open question.

The Competitive Landscape

The AI protein design space has become intensely competitive in 2025 and 2026. DeepMind's AlphaFold established that AI can predict protein structure with high accuracy; the next frontier is designing novel proteins with specified functions, which is what companies like Profluent, Generate Biomedicines, EvolutionaryScale, and others are pursuing. Generate Biomedicines, one of Profluent's most direct competitors in generative protein design, was pursuing an IPO in early 2026 as part of an $11 billion wave of AI drug discovery investment in Q1 2026. The field is converging rapidly from multiple directions , generative protein design, AI-assisted small molecule discovery, and AI-powered clinical trial optimization , and the large pharmaceutical companies are building positions across all of them simultaneously.

Within the recombinase-specific category, Lilly's dual-partnership approach , Seamless Therapeutics for hearing loss, Profluent for large-gene insertion across multiple disease areas , reveals a deliberate strategy of building redundancy and optionality in what is still an unproven therapeutic modality. Recombinase-based gene therapy is genuinely promising but has not yet produced approved therapies. Lilly is making early, large-scale bets on a modality it expects to become clinically validated within the next five to eight years. By establishing exclusive relationships with two different recombinase technology providers, Lilly is protecting against platform-specific failure risk while positioning itself ahead of competitors who have not yet entered the space. Novo Nordisk, AstraZeneca, and Roche are notable for their absence from announced recombinase programs , which creates an asymmetric first-mover opportunity for Lilly if the modality succeeds.

Hidden Insight: AI Is Not Just Finding Drugs , It Is Designing Biology That Has Never Existed

The most important thing to understand about what Profluent is doing , and why it is fundamentally different from earlier generations of AI drug discovery , is that it is not using AI to screen, rank, or optimize from within the space of things that already exist in biology. It is using AI to expand the space of possible biology itself. When Profluent trains a protein language model on the universe of known protein sequences and their functional properties, the model develops a representation of the underlying grammar of protein function , not just a lookup table of known proteins, but an implicit model of how sequence determines structure and structure determines function. This model can then be used to generate novel sequence proposals that the model predicts will fold into structures with the desired functional properties, including functions that no naturally occurring protein has ever demonstrated. The recombinases Profluent designs for Lilly may have no evolutionary antecedents. They are genuinely new biological entities, created by an AI model rather than by billions of years of natural selection.

This matters for the regulatory landscape in ways that are not yet fully appreciated. FDA has frameworks for evaluating gene therapies based on vectors that are derived from naturally occurring viral particles. It has frameworks for evaluating gene editing tools like CRISPR based on their behavior in known biological contexts. It does not yet have a well-established framework for evaluating therapeutic proteins that were designed de novo by AI systems with no natural analog. When a recombinase designed by Profluent's AI enters clinical testing , assuming the preclinical data supports it , FDA will be evaluating a biological entity whose properties cannot be predicted from evolutionary precedent or from the behavior of related natural proteins. The regulatory science will need to evolve alongside the technology, and the companies that participate earliest in that dialogue with regulators will have significant advantages in shaping the frameworks that govern the entire field.

There is a second hidden dimension in the Lilly-Profluent deal that points to a broader pattern in how frontier AI is being deployed in drug development. The $2.25 billion in milestone payments is not a purchase price , it is a performance-contingent structure that requires Profluent to produce AI-designed recombinases that actually work in biological systems, clear preclinical toxicology and efficacy hurdles, and ultimately advance through clinical development. The deal structure acknowledges that AI-designed biology, however intellectually compelling, still needs to survive contact with actual human physiology. This is the fundamental challenge that distinguishes AI drug discovery from AI software development: the validation loop is measured in years and requires biological experiments, not just benchmark tests. The companies that can close that loop fastest , from AI design through experimental validation , will define the competitive hierarchy in AI-driven therapeutics over the next decade.

What to Watch Next

The primary indicator to track in 2026 and 2027 is whether Profluent's AI-designed recombinases demonstrate in vivo activity in relevant disease models. Publication of preclinical data , in peer-reviewed journals, conference presentations, or FDA IND filings , will be the first independent validation that the AI design approach produces biologically functional enzymes. Watch specifically for data on large-gene insertion efficiency (defined as the percentage of target cells that receive the correct genetic payload), off-target insertion rates (a critical safety parameter), and immunogenicity profiles (whether the human immune system recognizes the AI-designed protein as foreign). All three metrics need to meet high bars before clinical development becomes viable.

The competitive landscape to monitor is whether Novo Nordisk, AstraZeneca, or Roche announce recombinase programs of their own in the next twelve months. If Lilly's two major recombinase commitments generate positive preclinical signals by mid-2026, expect fast-follower deals from competing pharmaceutical companies that have so far held back from the space. A wave of recombinase deals would validate the modality and significantly increase the commercial stakes for both Profluent and Seamless Therapeutics. Conversely, if early in vivo data from either partnership is disappointing, watch for Lilly to quietly restructure its commitment , the milestone-payment structure gives it financial optionality that a full acquisition or upfront payment would not. Within the next eighteen months, the biological evidence will begin to separate the genuine promise of AI-designed enzymes from the hype that currently surrounds generative protein design.

Eli Lilly is not betting that AI will find a drug faster , it is betting that AI can invent biology that evolution never produced, and that this invented biology will cure diseases that every other approach has failed to treat.

---