Biomedical Research · global
Can AI-Designed Proteins Direct Immune Cells? Thousands of CAR-T Designs Undergo Functional Testing
The Bits to Binders competition put 12,000 designs into human T cells, with a small number of candidates showing selective killing ability. The results bring protein design into cellular function testing while revealing the gap between computational predictions and therapeutic use.
Enabling a protein to recognize cancer cells is only the starting point for cell therapy. It must also function properly on the surface of immune cells and transmit appropriate signals to trigger an attack. The Bits to Binders study, published on October 8 in *Molecular Systems Biology*, tested 12,000 AI designs in CAR-T cells, shifting the focus of evaluation from whether molecules can bind a target to whether they can trigger the desired biological response.
A total of 28 teams submitted designs to the competition, with participants from 42 countries. Their task was to design miniature binding proteins with a fixed length of 80 amino acids to recognize CD20—a protein found on the surface of normal B cells and many B-cell malignancies. The researchers incorporated these designs into the same chimeric antigen receptor (CAR) scaffold and then introduced them into human T cells to test whether they could serve as target-recognition components.
The study first used pooled screening, tracking the cells corresponding to each design through their sequences. After culture, 6,811 designs met the sequencing recovery threshold. Of these, 707 were significantly enriched in the presence of CD20-positive lymphoma cells compared with conditions without those cells; hit rates across teams ranged from 0.6% to 38.4%. This indicates that some designs could support cell proliferation, but the number of enriched designs cannot be treated directly as the number of candidates confirmed to kill tumors.
The researchers then tested ten leading designs individually, seven of which were validated for CAR-T function. An accompanying journal commentary further noted that four showed significant selective killing ability in comparisons between CD20-positive and CD20-negative cell lines. However, the overall performance of these designs generally remained below that of antibody-derived CAR controls. As candidates were narrowed down from large-scale screening to individual functional confirmation, the meaning of “success” also became more concrete.
In their commentary, Memorial Sloan Kettering Cancer Center researchers Abdul Vehab Dozic and Caleb A Lareau cautioned that the pooled screen compared culture conditions with and without lymphoma cells. More definitive evidence that the response was driven by CD20 would require additional controls, such as cells with the same genetic background differing only by the removal of CD20. The experiments also used only one CAR scaffold and one T-cell donor, so the sequence features identified may not translate directly to other receptors or cells from different people.
Failed designs also provided clues. After analyzing the sequences retrospectively, the research team proposed filtering criteria related to DNA synthesis, protein expression, and cell activation. Applying these criteria could nearly double the proportion of designs showing antigen-dependent proliferation. However, this was a retrospective analysis derived from existing results; whether it can improve the success rate of the next batch of entirely new designs still requires confirmation in prospective experiments.
The team has made experimental readouts, computational sequence metrics, and analysis scripts publicly available through Zenodo, along with links to public code repositories, allowing other researchers to examine the stage at which designs failed. The value of this work lies in establishing a traceable benchmark for cellular function in AI protein design. Whether the candidates can safely and durably control tumors in the human body remains unanswered by this competition.