Biotechnology · global
Shortening the CAR Signaling Region by 70%: Anthropic Collaborations Bring AI Design into the Lab
From streamlining T-cell receptors to improving antibody stability, two collaborations show how AI designs undergo experimental testing. But progress in cell culture still faces multiple hurdles before patients can benefit.
Designing a molecule that can recognize cancer cells is only the starting point for drug development; it must also be efficiently produced by cells or remain stable during manufacturing and storage. On October 27, Anthropic announced progress in collaborations with six organizations. Among them, Generate Biomedicines and BigHat Biosciences put AI-proposed sequences through successive rounds of laboratory testing to address specific challenges in CAR T-cell engineering and antibody optimization, respectively.
Generate focused on the intracellular region of the chimeric antigen receptor (CAR) that transmits activation signals. Shortening this protein segment could also reduce the length of the mRNA needed to encode it, leaving room for delivery or additional functions. The team first selected 186 designs for screening, then tested them in donor-derived primary T cells. Subsequent design rounds yielded a version named Happy Stoat: at just 45 amino acids, it is approximately 70% shorter than the 4-1BB–CD3ζ control signaling region. This does not mean that the entire CAR was shortened by 70%.
Generate’s technical preprint identifies this construct as krgr–ζ1–40 and reports increased cell-surface CAR expression and interferon-γ release following antigen stimulation, with no interferon-γ release detected in the absence of antigen under the conditions tested. In another experiment, targeted lipid nanoparticles carrying CAR mRNA were added to cultures of blood immune cells from five donors. Happy Stoat achieved a similar degree of B-cell depletion at a lower dose. This provides evidence of in vitro delivery and function, but does not yet demonstrate the efficacy of generating CAR T cells inside the human body.
BigHat had Claude work with researchers and an automated laboratory to optimize the binding affinity, stability, and aggregation propensity of VHHs and single-chain variable fragments (scFvs). BigHat said it completed seven design rounds within ten weeks, synthesizing and comprehensively testing 1,232 antibodies. Anthropic’s announcement on the same day instead cited approximately 1,010 variants. The figures differ, and the announcement did not clarify the scope of the counts. Descriptions of the results also differed slightly: BigHat reported up to a ninefold improvement in binding affinity for the best variants, while Anthropic stated more than tenfold. For heat tolerance, BigHat reported that some variants remained stable in solution at temperatures 18°C higher than the starting antibodies.
Experimental feedback also changed how the team interpreted results. According to BigHat, quality checks by its data system, Reccy, found that the thermal stability instrument’s software could produce misleading interpretations of protein unfolding and aggregation. Claude and the researchers reexamined the raw data and proposed adjustments to protein concentration and heating rate, which BigHat then validated experimentally. This process shows that optimization depends not only on how many sequences are proposed, but also on whether measurements are reliable and whether the experimental conditions for each sample can be traced.
What the two cases currently provide is primarily laboratory results disclosed by the companies; Generate’s technical document is also a preprint. Stronger cell activation cannot be directly equated with safer treatment, and antibody binding and heat-tolerance data cannot replace human trials of efficacy and safety. These studies have produced testable biological results from AI-assisted design, but have not yet demonstrated improved patient outcomes. They are also insufficient to determine whether the overall success rate or cost of drug development has changed as a result.