AI-Designed Base Editors Reach the Clinic: Profluent, Beam, and the $160M ARPA-H Push

In December 2023, the US Food and Drug Administration (FDA) approved Casgevy, the first CRISPR-based therapy, for sickle cell disease. Eighteen months later, Beam Therapeutics reported that its base editing therapy BEAM-302 restored functional protein production in patients with alpha-1 antitrypsin deficiency, a disease caused by a single-letter mutation in DNA. The median AAT level across 29 patients crossed the threshold believed to be protective. Circulating misfolded protein dropped by over 80%. Base editing had moved from the lab to the clinic, and the data was clear.

NUS Medicine researchers publishing in Advanced Science in March 2026 used AlphaFold3 to engineer SsdAtox, a compact base editor two-thirds the size of conventional editors, and developed the Base Editor Performance Index (BEPI) to standardise comparisons across platforms. Compact editors matter because the industry's preferred delivery vehicle, the adeno-associated virus (AAV), has a payload limit of roughly 4.7 kb. Many high-performing editors simply do not fit.

The Profluent catalogue: from base editors to recombinases

Profluent Bio, founded in 2022 and headquartered in Emeryville, California, has built the most visible AI-native platform for protein design in the gene editing field. The company raised $106M in November 2025 from Altimeter Capital and Bezos Expeditions, bringing total funding to $150M, and has since signed four major partnerships in less than 18 months.

Its collaboration with Ensoma (December 2025) targets AI-designed base editors for in vivo hematopoietic stem cell therapies: editing blood stem cells inside the body rather than extracting, editing, and reinfusing them. The Ensoma deal uses its AI-designed deaminases combined with Ensoma's helper-dependent adenovirus (HDAd) delivery system. In April 2026, it signed a second, far larger deal with Eli Lilly, this time for site-specific recombinases that can insert or replace large DNA sequences, not just single-base edits.

As we wrote in July, AI-designed genetic circuits are already rewriting the code of life (When AI Learns to Write the Code of Life). It goes one layer deeper: it designs the enzymes that write those circuits.

The company also partners with Revvity on the Pin-point base editing platform, which integrates AI-enhanced adenine deaminases for commercial research use, and with the Rett Syndrome Research Trust to design base editors targeting the specific mutations that cause the neurodegenerative disorder.

Beam Therapeutics: the clinical proof point

Beam Therapeutics treated the first patient with a CRISPR-derived base editing therapy in 2023, making it the most advanced clinical-stage company in this space. Its lead program, BEAM-302 for alpha-1 antitrypsin deficiency (AATD), reported Phase 1/2 data in March 2026 showing that all three dose groups produced functional AAT above the protective threshold, with misfolded protein reduced by over 80%. It selected a 60 mg go-forward dose and plans to pursue an accelerated approval pathway with the FDA.

In June 2026, it received FDA IND clearance for BEAM-304, a base editing therapy for phenylketonuria (PKU) targeting the R408W mutation with a platform-based approach that can expand to other prevalent PKU-causing variants. The PKU program is designed as a single therapy addressing multiple mutations, a template for the broader mutation-agnostic model.

Reducing the editor without reducing the effect

Size matters in gene therapy delivery. AAV vectors, the most common delivery vehicle for in vivo gene editing, can carry roughly 4.7 kb of genetic payload. Many high-efficiency base editors exceed this limit, forcing trade-offs between editor performance and deliverability.

The NUS Medicine team led by Assistant Professor Jungjoon K. Lee tackled this constraint directly. Using AlphaFold3, they identified a structural bottleneck in SsdAtox, a compact DNA-editing enzyme roughly two-thirds the size of standard base editors. By engineering a single amino acid position (K31) that controls how DNA enters the enzyme's active site, they widened this entrance, improving editing efficiency while keeping the enzyme small enough to fit comfortably inside an AAV capsid, alongside regulatory elements.

The team developed a three-layer bacterial screening platform called Trinity-Screen to validate the optimised variants, and produced the Base Editor Performance Index (BEPI) to standardise efficiency comparisons across different editing platforms. The field generates increasingly diverse editor architectures, and a consistent measure was needed.

ARPA-H enters with $160M

The clearest signal for the field came in July 2026, when ARPA-H announced its THRIVE program (Treating Hereditary Rare diseases with In Vivo precision genetic mEdicines), allocating up to $160M over five years to seven teams. The GEMMABio-Profluent collaboration was the headline award, combining its AI-designed base editors with GEMMABio's delivery infrastructure led by gene therapy pioneer Jim Wilson. The award targets two monogenic liver diseases: homozygous familial hypercholesterolemia (HoFH, LDLR target) and maple syrup urine disease (MSUD, BCKDHB target).

Both targets are delivered as LNP-encapsulated mRNA to hepatocytes, and the program's stated goal is to build a library of base editors that can correct any transition mutation. That is a platform rather than a one-off therapy. The pitch word is scalable: the same AI design pipeline that produced one editor can produce a hundred.

The radical element, as several analysts noted, is not the editor but the regulatory pathway. FDA's February 2026 draft guidance on individualized therapies and the "plausible mechanism" standard, inspired directly by the case of baby KJ who received a personalised base editing treatment in 2024, opens the door to master protocols and label extension to mutations never clinically tested. If approved, this could compress the timeline from editor design to first human dose from years to months.

How base editing compares

Base editing occupies a specific niche: it addresses approximately 60% of known disease-causing single-nucleotide variants, with a safety profile that makes it suitable for tissues where double-strand breaks carry unacceptable risk, such as hematopoietic stem cells and post-mitotic neurons. It cannot insert new genes or correct deletions. Those remain the domain of prime editing and viral gene addition. But for the mutation class it covers, it is the smallest, most deliverable, and most clinically advanced precise editing modality.

The platform licensing model emerging in this field works like this: companies develop proprietary editing technologies and license them to therapeutic developers for per-patient royalties. It creates an economic structure distinct from traditional drug development. Profluent, Beam, and the Broad Institute's Liu lab all participate in some form of this model, where the platform company captures recurring revenue across multiple indications without bearing the full regulatory cost of each one.

Three milestones to track

Three milestones will determine whether AI-designed base editors become a genuine therapeutic platform or remain a disconnected collection of proofs-of-concept. First, its accelerated approval filing for BEAM-302, expected in 2027. A green light would validate the base editing modality for regulators, not just scientists. Second, the first compact editor candidate entering IND-enabling studies from the NUS Medicine SsdAtox lineage or a comparable sub-3 kb scaffold. Delivery fits inside AAV, and regulatory momentum from the FDA's individualized therapies guidance could compress the timeline. Third, the GEMMABio-Profluent ARPA-H program showing that a single AI-designed editor platform can address multiple mutations across different liver diseases. That is the scalability thesis in action.

Sources