AI designed a protein "mask" for anticancer cytokines from scratch: it keeps the drug inactive in the bloodstream and is cleaved only inside the tumor by the tumor's own enzymes
AI designed a protein "mask" for anticancer cytokines from scratch: it keeps the drug inactive in the bloodstream and is cleaved only inside the tumor by the tumor's own enzymes
David Baker's laboratory and the Dana-Farber Cancer Institute described a system called Prokine in a September 22 preprint: cytokines, signaling proteins that activate immune cells against tumors, are packaged inside a mask that AI generated to fit their structure. The mask weakens the cytokine's binding to its receptor by several hundred thousand fold for one variant, and the enzyme that cleaves the mask is one that tumors produce in excess. In mice, including a pancreatic cancer model, the masked drug suppressed tumors and extended survival as effectively as the unmasked cytokine, but without its toxicity.
In the bloodstream, cytokines activate immune cells in both the tumor and healthy tissues at once, and this indiscriminate activation is the source of their toxicity. Baker's laboratory had previously engineered synthetic versions of these signals: 21h10 (modeled on interleukin-21) and Neo2 (modeled on interleukin-2). 21h10 suppressed tumors in mice but caused weight loss, a consequence of pancreatic inflammation.
Masking cytokines had been attempted before by attaching a receptor fragment or an antibody that a tumor enzyme would later cleave, but the geometry of those masks was inherited from evolution, leaving no way to tailor them to a new target. Baker's team (Baker received the 2024 Nobel Prize in Chemistry for de novo protein design) built a custom mask for each cytokine: RFdiffusion2 proposed the scaffold shape, ProteinMPNN selected the amino acid sequence, and AlphaFold2 verified the fit.
The first masks reduced signaling by a factor of 16 to 17. Redesigning the contact surface yielded Pro21 and Pro2: receptor affinity dropped ~700,000-fold for Pro21 and ~170-fold for Pro2, and after cells were treated with MMP2, an enzyme that tumors produce in excess, signaling recovered to near its original level.
Faster cleavage of the mask did not necessarily improve the outcome. The team tested four linkers (the segment connecting the mask to the cytokine, cleaved by the tumor enzyme) with different sensitivities to MMP2. The most sensitive linker suppressed the tumor but caused toxicity outside it as well, while the least sensitive one failed to produce the full effect. Only the intermediate variant, L4, suppressed the tumor without causing weight loss.
In a separate approach, Neo2 was split in two. Each half was concealed under its own mask and conjugated to a molecule targeting the tumor marker PD-L1 or HER2. Signaling activated only when both halves converged on the same cell and both masks had been cleaved by the enzyme, yielding a "dual-lock" cytokine.
In the KPC.1 pancreatic cancer model, which is typically immune-excluded, targeting PD-L1 allowed a fivefold dose reduction of Pro21 with linker L4, down to the level of unmasked 21h10, and still extended survival in mice as effectively as the unmasked cytokine, without weight loss. In these mice, STAT3 (a marker indicating that a cell has encountered active cytokine) was activated more weakly in the spleen than in mice receiving unmasked 21h10, indicating that less cytokine was reaching tissues beyond the tumor.
The difficulty of this problem is illustrated by the fate of bempegaldesleukin, a slow-activating version of interleukin-2 developed by Nektar and Bristol Myers Squibb: its phase III trial in melanoma in 2022 failed to show a significant effect. The Prokine authors see in their work a general principle of programmable protein encapsulation. The same Dana-Farber group tested the concept in reverse: the NISE shell keeps an unstable drug molecule enclosed for over 50 hours instead of immediate degradation. They consider the approach applicable to other toxic immunotherapeutics, including antibodies.