Designing Therapeutic Enzymes with AI: Profluent and BioMarin Partner to Seek New Drugs for Metabolic Diseases

For some inherited metabolic diseases, the key to treatment is restoring impaired enzyme function. But designing proteins that can perform their intended tasks and are sufficiently stable often requires repeated modification and screening. On September 30, Profluent announced a collaboration with BioMarin to introduce AI protein design into this development process and seek new therapeutic enzymes for diseases with a clearly defined genetic basis.

The two companies have specific roles: Profluent uses protein language models to generate and optimize candidate designs, while BioMarin conducts screening to assess which molecules merit further development. BioMarin also confirmed the collaboration in an official social media post, positioning it as part of early-stage research and development for metabolic diseases, with the aim of selecting promising candidate molecules to advance to evaluation in preclinical models.

Protein language models use amino acid sequences as the material for their designs. According to information publicly released by both companies, Profluent’s models use a dataset called the Profluent Protein Atlas, which contains more than 115 billion distinct protein sequences. This figure describes the scale of the sequence data; it does not mean that an equivalent number of proteins have been experimentally validated, nor can it be directly translated into a drug development success rate.

The collaboration will explore a broader range of enzyme designs. Both Profluent’s announcement and HLTH’s October 1 report state that the research will seek candidate molecules with the required activity and stability: the former concerns whether an enzyme can perform its intended function, while the latter is an important condition for whether a protein can remain in a usable state. After the models propose designs, BioMarin’s experimental screening will still determine which ones merit advancement.

The information publicly available so far does not identify specific diseases, enzyme targets, or candidate drug names, nor does it provide activity measurements, stability comparisons, or efficacy data from disease models. Preclinical testing is a possible next step after screening; this does not establish that any candidate drug has already completed the relevant validation. The announcement also provides no new evidence of safety or efficacy in humans.

The significance of the collaboration lies in connecting computational design with a drug company’s experience in enzyme development, giving AI-proposed molecules a defined path for experimental evaluation. Whether it can shorten development timelines or produce enzymes superior to existing approaches remains to be answered through subsequent comparisons. Even if candidate molecules pass initial screening, whether they can function in the body, how they should be administered, and whether they trigger immune responses remain questions that must be addressed as they move toward clinical development.