Profluent

The models are the company. Profluent's platform pages describe foundation models that write protein sequences, a suite of large language models for protein generation with task specific models built on it (ProseLM for structure conditioned design, Protein2PAM for the DNA motifs an editor recognizes), paired with an in house wet lab whose purpose is to test what the models write.

The published result carries the point: OpenCRISPR-1, a working gene editor whose sequence sits more than 400 mutations from SpCas9 and nearly 200 from any known natural CRISPR associated protein, is not something a search of nature could have returned. Remove the models and there is no product to license and no partnership to enter.

The oversight structure is described and the threshold is missing. Generation is followed by laboratory characterization as a matter of course: the platform pages describe foundation models paired with an in house wet lab so that every design step is tested, partnerships include wet lab validation calibrated to the partner's needs, and the OpenCRISPR work followed that path, with candidates generated, a selected set characterized functionally in human cells, and both on target and off target activity reported against the incumbent enzyme. Nothing the models produce reaches a use without being made and measured.

What is not published is any threshold or rule: no generation to hit ratio across programs, no confidence measure exposed with a value at which a design should not be pursued, and no statement of where the models are unreliable. The oversight is demonstrated in the flagship case rather than specified as a standing practice a partner can hold the company to.

The approach and the models are named and published, without version or update discipline. Two papers carry the method: Large Language Models Generate Functional Protein Sequences Across Diverse Families (Nature Biotechnology, 2023), which established that the approach produces working proteins, and Design of highly functional genome editors by modeling CRISPR-Cas sequences (Nature, 2025), which describes the designed editor and its measurement. The models are named individually rather than presented as one platform: the ProGen family for generation, ProseLM for structure conditioned design, Protein2PAM for editor recognition motifs, with the company describing ProGen3 as billion parameter models trained on more than 3.4 billion sequences using a sparse architecture it reports gives a fourfold speedup.

What is not published is which model version produced a given design, any release note or update practice, or model cards for the models a partner would use; the weights are held rather than released, so a partner relies on the published account of a related system.

A structurally short chain: the models are the company's own and are named. ProGen3, ProseLM and Protein2PAM are built in house and run by Profluent, with the wet lab work done in its own laboratory, so a partner's targets and sequences pass through one organization rather than a chain of suppliers, and no outside model provider is named as receiving partner material.

The party list is not enumerated: no hosting arrangement, no subprocessor list, and no published position on whether partner supplied sequences or assay results improve models used for other partners, which is the commercially sensitive question for any design partner. A second question is specific to what this company ships, since the material that leaves the platform is a functional molecule rather than a report: nothing published describes what governs onward transfer of a design once it is licensed, beyond the ethical use obligations in the OpenCRISPR license.

Peer reviewed molecular evidence with a head to head comparator, and no candidate in humans. OpenCRISPR-1 was published in Nature in 2025 and measured against SpCas9, the tool it would replace: on target editing of 55.7 percent against 48.3 percent and off target editing of 0.32 percent against 6.1 percent across a panel of genomic sites in HEK293T cells, with the designed editor also shown to work in a base editing architecture when combined with generated deaminases. The founding method paper, showing that language models generate functional proteins across diverse families, appeared in Nature Biotechnology in 2023. Adoption is named rather than anonymous: a strategic partnership with Eli Lilly on AI designed recombinases for genetic medicine was announced in April 2026, and the company names commercial relationships across life science tools, agriculture and rare disease research.

What does not exist is a therapeutic in humans: no clinical study of a molecule designed here is reported, and the company states bringing an AI designed therapeutic to patients as an aim.

General assurances that do not answer the questions AI raises, and one disclosed practice that runs against the user. The privacy policy (last updated 1 January 2025) covers the website, its subdomains and the outputs generated through the Protein2PAM server, and lists among its purposes improving the models used on the website, so a user of that tool is told its submissions may feed the models. Retention is stated only as a principle, with the company reserving the right to dispose of any data at its discretion without notice, and de-identification is not addressed.

Nothing published reaches the business a partner actually contracts for. A discovery partnership puts a partner's targets, constraints and assay results into the platform, and no published document states what is retained, whether that material trains models used for other partners, or what happens to it when the work ends; those terms sit in private agreements. Safety engineering is not addressed.

No statement of status and no privacy document that reaches the product. The privacy policy (last updated 1 January 2025) defines its subject as the website, its subdomains and the outputs of the Protein2PAM server, so it reaches a visitor and a user of one free web tool and not the licensing and partnership business a buyer contracts for.

No business associate status is stated anywhere, and no scope position explains whether the company can receive protected health information, for example sequence data derived from patient samples in a partner's program. Silence is not a stated position.

Controls are asserted with nothing independent behind them. The company publishes no SOC 2 report, ISO 27001 certificate, penetration testing cadence or trust center. The privacy policy states that security measures have been implemented to protect personal information and that no measures can guarantee security; that is an assertion about the website and the Protein2PAM server, which is what the policy's scope covers, and there is no outside check behind it.

The published terms of use and privacy policy do not reach the licensing and partnership business, where a counterparty's targets and design constraints would be the sensitive material, and no separate published terms govern that relationship. The privacy policy also reserves the right to dispose of any data at the company's discretion without notice, which is unusual phrasing to find beside a hosted tool.

The product carries no device claim and is scoped accordingly. Profluent designs proteins and licenses them; nothing in its published material presents the models or the designed molecules as regulated products, and the company discloses no clinical pipeline of its own. A therapeutic built on OpenCRISPR-1 or another licensed asset carries its own regulatory position, held by the licensee that develops it, which follows from the licensing model. The one regulatory commitment published is contractual rather than statutory: commercial therapeutic users of OpenCRISPR-1 execute a license with ethical use obligations.

Responsible use is committed to in the licensing terms, with no evaluation of model behavior behind it. The OpenCRISPR page states that the release is licensed for ethical research and commercial uses, and its FAQ explains why a license is required at all for a freely available sequence: commercial therapeutic users execute an agreement carrying obligations to use the tool for ethical purposes only, alongside other terms of use. That is a use control written into a binding instrument rather than a values statement, and it places the obligation on the licensee.

Nothing published examines how the models themselves behave: no model card, no fairness or robustness evaluation with a stated method, and no breakdown of generation quality across protein families, which for a model trained on natural sequence databases is where coverage would be expected to vary. The published characterization of OpenCRISPR-1 includes off target editing measured against SpCas9, which is evidence about one designed molecule rather than about the model that produced it.

A measured error rate is published for the asset a buyer licenses, with its method and comparator. OpenCRISPR-1's characterization appeared in Nature in 2025 (Design of highly functional genome editors by modeling CRISPR-Cas sequences): delivered by plasmid in HEK293T cells across a panel of genomic sites, it edited on target at 55.7 percent against SpCas9's 48.3 percent while editing off target at 0.32 percent against 6.1 percent, roughly a twentyfold reduction in the failure mode that matters most for a gene editor. The company names the figures on its own estate and publishes the protocol, so a licensee can reproduce the measurement rather than take it on trust, and the sequence is obtainable, so anyone can test it independently.

That is a rate for one released molecule, not for the design service: no success rate is published for proteins designed in a partnership, and no warranty, indemnity or remediation commitment attaches to either route. The website privacy policy reserves the right to dispose of data at the company's discretion without notice, which runs the other way.

No clinical record integration exists: Profluent designs molecules and licenses them, and no clinical record system is involved. For a pharma or biotech partner the record systems that matter are its research systems: electronic lab notebooks, laboratory information systems, research data platforms and cloud data environments.

None is named and no integration is claimed: work moves through a partnership rather than through a connection between systems, and the only hosted surface is a web server for predicting the DNA motifs an editor recognizes. For a design partnership, that missing integration is an absence rather than a defect. How sequences, constraints and results move between the two organizations is not described.

An internal platform, with location implied rather than committed. The models are not deployed to customers: a partner sends its goal and its targets and receives designs, with validation run in Profluent's own wet lab, and the alternative route is a licensed molecule, where the recipient simply holds the sequence and no platform access is involved. The one hosted surface is the Protein2PAM server, used through the website.

The company operates from Emeryville, California, and nothing states where the platform runs, where partner material is stored, whether any processing crosses a border, or how one partner's sequences are separated from another's.

One price is published and the rest of the structure is discoverable. The published term is OpenCRISPR-1: the sequence is free, available through the preprint, and a buyer who wants it for commercial therapeutic use executes a license the company describes as simple and lightweight, carrying ethical use obligations and other terms of use rather than a fee, royalty or milestone. For a buyer weighing a gene editor that is a real number, since the comparison is against paying to license conventional Cas9.

The two routes the partnerships page names, a discovery partnership with wet lab validation and asset licensing of ready made proteins, both route to a contact form, and no fee schedule, milestone structure or royalty terms are published for either.

Coverage is named across industries, with validation behind one part of it. The company states that its platform is pointed first at CRISPR and gene editing, and names applications from new therapeutics to industrial enzymes across medicine, agriculture and beyond, with designs spanning gene editors, antibodies and broader enzymes and a partnership on recombinases for genetic medicine announced in April 2026. Modality is proteins rather than small molecules, which is the real boundary.

Validation sits behind the gene editing work, where a designed editor was characterized in human cells and published with its numbers; the antibody, enzyme and agricultural applications are named without published results.