An artificial intelligence-designed protein binder enabled CAR T cells to control tumor growth more effectively than antibody-derived binders used in approved BCMA-targeting therapies in a mouse model of multiple myeloma.
The study, led by researchers at Memorial Sloan Kettering Cancer Center, New York, examined whether generative protein design could provide an alternative to the single-chain variable fragments commonly used to recognize target antigens in CAR T therapies.
The team generated and screened 1,758 de novo binders against BCMA, CD19, and CD22. Candidates were assessed using protein-binding assays, cell-based CAR activation studies, primary human T cells, and mouse models.
AI-designed binders are built from scratch rather than adapted from existing antibodies. Their small size and ability to target selected epitopes offer potential advantages, but strong biochemical binding does not necessarily translate into effective CAR T cell activity.
The researchers identified three recurring problems: tonic signaling, in which CAR T cells activate without encountering the target antigen; binding sites that were inaccessible on cells; and off-target activity. They then developed computational and experimental rules to identify or correct these liabilities.
Results were strongest for BCMA, a target expressed on multiple myeloma cells. One binder, designated B5, produced antigen-dependent activation and cytokine release in primary CAR T cells. An optimized variant, B5.I0, controlled tumor growth in mice more effectively than CAR T cells incorporating BCMA binders used in approved therapies. The work remains preclinical, and the mouse experiments involved small treatment groups.
The team also developed CARPNN, a neural network trained on experimental results to examine how differences in amino acid sequence affected CAR performance. The analysis linked excessive positive charge with unwanted tonic signaling, and identified high alanine content as a potential barrier to effective binding.
Applying these findings to CD22 enabled the researchers to modify a binder that activated against cells lacking the intended target. Changes to amino acids outside the binding interface retained on-target activity while reducing off-target activation.
The CD19 program was less successful. Designed binders recognized recombinant CD19 but showed limited activity against the protein on cells, which the researchers attributed to obstruction of the targeted epitope by CD81, a natural binding partner.
“We’re changing what was traditionally educated guesswork with more informed, evidence-based design,” said senior author Caleb Lareau in an MSK article.
Newsletters
Receive the latest pharmaceutical news, personalities, education, and career development – weekly to your inbox.
