The origins of immunotherapy for cancer date back to the late 1800s, when William B. Coley noticed an interesting phenomenon: some of his cancer patients who became infected with viral or bacterial diseases would spontaneously clear their tumors. This observation pre-dated any granular knowledge of the immune system, which prevented the advancement of the field until almost 100 years later. After the discovery of T cells and their significance in 1967, the critical cytokine interleukin-2 (IL-2) was discovered, later cloned, and eventually administered in the clinic. High-dose IL-2 proved, for the first time, that pharmacologic stimulation of T-cell growth signals could eradicate metastatic melanoma and renal-cell carcinoma; yet its success came at the cost of dose-limiting capillary-leak syndrome, cytokine-driven hypotension, and the paradoxical expansion of immunosuppressive T-regs. These divergent clinical outcomes were later understood to stem from the interaction of IL2 with different IL2 receptors. High-affinity engagement of the α- and βγ-ILR2 complex drives dose-limiting toxicities while also dampening anti-tumor immunity. By contrast, exclusive binding of the βγ-IL2R expands cytotoxic CD8⁺ and NK populations while largely sparing T-regs, unlocking more potent yet safer immunotherapy.
In this tutorial, we leverage recent advances in de novo protein design to pursue the high level goal of designing a selective IL2R binder, which mediates therapeutic effect while minimizing toxicities. To achieve this, we chain together three algorithms on Superbio into a single workflow. RFDiffusion first generates a series of 3D backbone binder structures, using hotspot residues derived from the binding interface between IL-2 and βγ-ILR2. Those backbones feed directly into ProteinMPNN, which yields optimized amino acid sequences that fold into these backbone templates. Finally, Boltz-1 rescoring ranks the binder candidates by predicted interface interactions and conformational stability to both α- and βγ-ILR2. This enables the filtering of designs most likely to achieve βγ affinity with negligible α binding. The result is an end-to-end, fully open-source pipeline that can progress from a receptor PDB to experimental expression constructs in under an hour, illustrating how modern AI stacks can accelerate selective binder engineering.
In the endogenous context, IL-2 is capable of binding the IL2R α-chain (CD25) and βγ-heterodimer at the same time. CD25 has a much higher affinity for IL-2, usually outcompeting the βγ-chain for binding. Below, the 2B5I PDB shows IL-2 bound to the all three of it’s receptor chains at once.
Step 1: Obtain Target RFDiffusion PDB and parameters.
RFDiffusion generates new protein binders against a given target, when given a PDB structure, restricting the model to designing binders with desired properties. As we are interested in βγ-heterodimer as the sole target, we will first isolate these chains from the 2B5I PDB file.
Now that we’ve obtained our target structure, we can determine RFDiffusions essential parameters for binder design. For a discussion of contig map and hotspot residues, see our previous tutorial.
Based on our analysis of the 2B5I structure and the IL-2/βγ-receptor interface, we’ve determined several interaction points that should be maintained for a βγ-selective IL-2 mimic. We will use the following residues as hotspots. Note: the ‘B’ and ‘C’ precursor refer to the chain identity, not the amino acid - just as the RFDiffusion program expects.
‘B9,B35,B36,B63,B64,B65,B67,B68,B69,B127,B128,B132,B182,C19,C20,C24,C25,C28,C32,C67,C68,C71,C120,C148’
We will then choose a) which regions of the target protein we’d like the model to ‘see’, and b) the desired length of our binder. This gives us a contig map of:
‘B1-196/0 C1-191/0 50-80’