Biotechnology · global
Living Cells Put AI Designs to the Test: 12,000 Proteins Face CAR-T Functional Testing
A global protein design competition placed computer-generated recognition molecules into human T cells, testing everything from proliferation to cancer-killing ability. The results reveal a threshold between triggering an immune response and precisely killing target cells that still must be crossed through experiments.
Designing an “adapter” that enables immune cells to recognize cancer cells takes more than showing on a computer that it appears capable of binding to its target. The adapter must function properly inside cells and turn recognition signals into an effective attack. A global AI protein design competition tested 12,000 designs in human T cells, subjecting algorithms to a test of biological function; the related analysis was published in *Molecular Systems Biology* on October 8.
The competition, called Bits to Binders, received designs from 28 teams with participants from 42 countries. Each miniature binding protein consisted of 80 amino acids, targeted CD20, and served as the recognition region of a chimeric antigen receptor T cell (CAR-T). This region acts as the immune cell’s sensor: after encountering a cell carrying the target antigen, it must also activate signaling downstream of the receptor to drive a T-cell response.
The organizers used a two-stage validation process. First, they screened a large pool of CAR-T designs together, observing which sequences increased as T cells proliferated after encountering CD20-positive cancer cells. A total of 707 designs showed significant CD20-associated enrichment, with team hit rates ranging from 0.6% to 38.4%. These figures reflect the results of an initial functional screen and cannot yet be taken directly as the proportion of designs that effectively kill cancer cells.
Next, the research team selected the ten leading designs and measured their functions individually. Seven outperformed negative controls in CD20-specific cytokine production, proliferation, and expansion in cell numbers; in the CD20-specific killing assay, four showed a significant advantage. The gap between screening and individual validation demonstrates that stimulating T cells and increasing their numbers does not guarantee that they can effectively and selectively eliminate their targets.
These experiments also provided clues for improving the design process. According to Twist Bioscience’s description of a webinar on the same topic, researchers retrospectively analyzed factors related to DNA synthesis, expression, and target-specific activation. After filtering criteria were applied, the proportion of successful designs nearly doubled. However, this was a retrospective analysis using existing results; whether it can improve the hit rate of the next batch of entirely new designs still requires testing in new experiments.
Open data extend the competition’s value beyond the rankings. BioML Society stated that participants agreed not to claim intellectual property rights over their submitted designs and methods. The third version of the data record on Zenodo includes experimental readouts, computational sequence metrics, and analysis code, and provides access to immunophenotyping data and the project’s code repository. Other researchers can therefore compare different design approaches and trace the stage at which candidate molecules lose function.
This remains an in vitro functional evaluation, rather than a new therapy available to patients. The paper notes that the designs were compared within the same CAR architecture and that the T-cell experiments used cells from a single donor. Whether the results can be generalized to other cell sources or receptor configurations remains to be confirmed. This study establishes a standard of validation more closely aligned with therapeutic use: for AI designs to advance toward the clinic, they must progressively demonstrate that the molecules can bind, the cells can respond, and the attack is sufficiently precise.