AI-designed cyclic peptides enable controllable modulation of the CD28 immune checkpoint

This study presents an AI-guided strategy for designing the cyclic peptide CIP-3, a controllable CD28 antagonist that reversibly suppresses T-cell activation and treats colitis in preclinical models, offering a tunable synthetic alternative to traditional biologic immune checkpoint therapies.

The Gist

Imagine your immune system is a highly trained security team guarding your body. Usually, this team is great at spotting invaders like bacteria or viruses. But sometimes, the security team gets a little too excited and starts attacking your own body, causing diseases like colitis (a painful inflammation of the gut) or rheumatoid arthritis.

To calm this team down, doctors use "brakes" called immune checkpoints. One of the most important brakes is a button on the security guard's chest called CD28. When the wrong signal hits this button, the guard goes into overdrive.

The Problem: The "Heavy Hand" Approach

Currently, the main way doctors try to stop this overactive button is by using antibodies (large, complex biological drugs). Think of these antibodies as giant, heavy-duty handcuffs.

• The Good: They work well to stop the button.
• The Bad: Once you put the handcuffs on, they are very hard to take off. They stay locked for a long time. If you accidentally handcuff the guard too tightly, you can't easily loosen the grip to let them do their job again. Also, because they are so big and complex, they are expensive to make and can sometimes cause the immune system to freak out in dangerous ways (like a "cytokine storm").

The New Solution: The "Smart, Reversible" Peptide

This paper introduces a new, clever solution: AI-designed cyclic peptides.

Here is the analogy: Instead of giant handcuffs, imagine a smart, magnetic key that fits perfectly into the button's lock.

1. How they found it (The AI Architect):
   Scientists didn't guess the shape of this key. They used a super-smart Artificial Intelligence (AI) to design it from scratch. The AI looked at the CD28 button and said, "I need to build a tiny, ring-shaped molecule that fits exactly into the groove where the natural activator binds." It's like using a 3D printer to design a custom key that fits a specific lock better than any existing key.

2. The "Ring" Shape (Cyclic Peptides):
   The key they built is a cyclic peptide. Think of a normal protein as a loose string of beads that flops around. A cyclic peptide is that same string, but the ends are tied together in a tight ring. This makes it stiff and strong, so it doesn't flop around and can hold its shape perfectly against the button.

3. How it Works (The Perfect Fit):
   The lead molecule, named CIP-3, is so well-designed that it snaps onto the CD28 button with incredible precision (nanomolar affinity). It acts like a dummy key that jams the lock. When the real "activator" tries to turn the button, CIP-3 is already sitting there, blocking it. The security guard stays calm.

Why This is a Game-Changer

The paper highlights three superpowers of this new "magnetic key" compared to the old "handcuffs":

• 1. It's Reversible (The "Off Switch"):
  This is the biggest breakthrough. The handcuffs (antibodies) stay on for days. The magnetic key (CIP-3) is temporary. If the security guard needs to be activated again, you can simply wash the key away, and the button is free to work again within hours. This gives doctors total control. They can dial the immune system down when it's angry, and dial it back up when it's too quiet, without waiting days for the drug to leave the body.

• 2. It's Safe (No False Alarms):
  Some old drugs accidentally pushed the button instead of blocking it, causing the security team to go berserk (a dangerous side effect). CIP-3 is purely a blocker. It sits there and does nothing but block; it never accidentally pushes the button.

• 3. It Works in Real Life (The Test Drive):
  The scientists tested this on human cells from healthy people and patients with ulcerative colitis. It worked just as well as the best antibody drugs at calming the inflammation.
  Then, they tested it in mice with colitis. The mice treated with CIP-3 got better: their bellies stopped hurting, their weight stabilized, and their colon length returned to normal. The drug worked exactly as predicted.

The Bottom Line

This paper shows that we can now use AI to design tiny, ring-shaped molecules that act as precise, controllable brakes for our immune system.

Instead of using a sledgehammer (antibodies) that stays stuck in place, we now have a scalpel (the cyclic peptide) that can gently turn the immune system down when needed, and be removed quickly when the job is done. It's a major step toward safer, smarter, and more controllable treatments for autoimmune diseases.

Technical Summary

1. Problem Statement

• Therapeutic Gap: Immune checkpoint therapies are currently dominated by biologic agents (monoclonal antibodies, fusion proteins). While effective, these biologics suffer from prolonged receptor occupancy, limited pharmacologic controllability, high immunogenicity risks, and complex manufacturing.
• Safety Concerns: Modulating costimulatory receptors like CD28 is particularly risky. Excessive activation can trigger severe cytokine release syndromes (e.g., the TGN1412 tragedy), necessitating precise, tunable control over signaling intensity rather than binary on/off blockade.
• Drugability Challenge: CD28 presents a shallow, solvent-exposed protein-protein interaction (PPI) surface lacking deep binding pockets. This topology renders it "undruggable" by conventional small molecules, which typically rely on deep pockets, yet creates a modality gap between small molecules and large biologics.
• Need: There is a critical need for synthetic modalities that can target these shallow PPI interfaces with high specificity, tunable pharmacology, and reversible binding kinetics.

2. Methodology

The study employed an AI-guided discovery framework to design and validate cyclic peptide antagonists for CD28.

• Computational Design:
  • Platform: Utilized the Highplay platform, integrating Monte Carlo Tree Search (MCTS) based reinforcement learning with the HighFold structure prediction model.
  • Target: The extracellular domain (ECD) of human CD28, specifically a flat, solvent-exposed $\beta$-sheet region involved in CD80/CD86 binding.
  • Scaffold: Disulfide-constrained cyclic peptides (Cys-X$_n$-Cys) were prioritized to provide conformational preorganization, reducing entropic penalties and enhancing proteolytic stability.
  • Filtering: Candidates were ranked by predicted interaction energy, interface hydrogen bonding, and structural confidence (i_pLDDT > 0.85).
• Synthesis & Characterization:
  • Top candidates (CIP-1, CIP-2, CIP-3) were synthesized via Solid Phase Peptide Synthesis (SPPS) using Fmoc/tBu strategy.
  • Cyclic structures were formed via air oxidation of cysteine residues to create disulfide bridges.
  • Purity and mass were confirmed via HPLC and MS.
• In Vitro Evaluation:
  • Binding Affinity: Measured using Microscale Thermophoresis (MST) with fluorescently labeled CD28 ECD.
  • Competitive Binding: Assessed via ELISA to determine inhibition of CD28-CD80 interaction.
  • Cellular Signaling: Used a Jurkat T-cell luciferase reporter assay (NFAT-driven) to measure downstream costimulatory signaling.
  • Primary Human Systems: Tested on PBMCs from healthy donors and ulcerative colitis (UC) patients to assess cytokine suppression (IL-2, IFN-$\gamma$) and intrinsic agonist activity.
  • Reversibility: Washout experiments compared the kinetic reversibility of peptide vs. antibody binding.
• In Vivo Evaluation:
  • PK/PD: Assessed plasma stability, liver microsomal stability, and pharmacokinetics (PK) in mice (subcutaneous administration).
  • Efficacy Model: Adoptive T-cell transfer model of chronic colitis in C.B-17 scid mice (induced by CD4$^+$CD45RB$^{high}$ T cells). Mice received daily subcutaneous doses of CIP-3 (1 mg/kg or 5 mg/kg).
  • Endpoints: Disease Activity Index (DAI), colon length, and serum cytokine levels (TNF-$\alpha$, IL-6).

3. Key Contributions

• AI-Driven Discovery: Successfully demonstrated that AI-guided reinforcement learning can design disulfide-constrained cyclic peptides targeting shallow PPI interfaces (CD28) that are traditionally difficult to drug.
• Synthetic Modality: Established cyclic peptides as a viable "middle ground" modality between small molecules and biologics, offering synthetic accessibility with biologic-like surface complementarity.
• Controllable Pharmacology: Introduced a synthetic ligand with rapid pharmacologic reversibility and no intrinsic agonist activity, addressing the safety limitations of current CD28-targeting biologics.
• Translational Proof: Validated efficacy across healthy human donors, disease-relevant patient samples (UC), and an in vivo murine colitis model.

4. Key Results

• Binding Affinity:
  • CIP-3 (the lead peptide) bound human CD28 ECD with nanomolar affinity ($K_d \approx 108$ nM).
  • CIP-2 showed micromolar affinity ($K_d \approx 47$ $\mu$M), while CIP-1 showed no binding, highlighting the sequence-dependence of the design.
• Functional Potency:
  • Competition: CIP-3 competitively disrupted CD28-CD80 interactions with an $IC_{50}$ of 609 nM (approx. 65-fold better than CIP-2).
  • Cellular Signaling: CIP-3 suppressed CD28-dependent T-cell activation in Jurkat cells with an $IC_{50}$ of 443 nM.
• Primary Human Systems:
  • CIP-3 suppressed IL-2 and IFN-$\gamma$ secretion in PBMCs from 10 healthy donors and 5 UC patients in a dose-dependent manner.
  • Efficacy was comparable to the benchmark anti-CD28 antibody (FR104) but with a tunable, graded suppression profile rather than a binary block.
• Safety & Reversibility:
  • No Agonism: CIP-3 did not induce cytokine production in unstimulated PBMCs.
  • Reversibility: Washout experiments showed CIP-3-mediated inhibition was rapidly reversible (85% signaling recovery), whereas antibody-mediated inhibition persisted (30% recovery), confirming transient, non-covalent binding.
• In Vivo Efficacy:
  • CIP-3 demonstrated favorable PK in mice (half-life ~3.9 h, stable in plasma/microsomes).
  • In the colitis model, daily subcutaneous dosing (5 mg/kg) significantly reduced Disease Activity Index (DAI), preserved colon length, and suppressed systemic TNF-$\alpha$ and IL-6 levels compared to vehicle controls.

5. Significance

• Paradigm Shift: This work challenges the dominance of biologics in immune checkpoint modulation by proving that synthetic cyclic peptides can achieve comparable efficacy with superior pharmacologic controllability.
• Safety Profile: The absence of intrinsic agonism and the rapid reversibility of CIP-3 offer a potential solution to the safety risks (cytokine storms) associated with CD28 agonists or long-acting antibodies, allowing for "exposure-dependent" tuning of immune responses.
• Future Therapeutics: The study provides a framework for developing next-generation immunotherapies for autoimmune diseases (e.g., IBD, rheumatoid arthritis) and potentially oncology, where precise tuning of T-cell activation thresholds is required.
• Methodological Advance: It validates the integration of reinforcement learning and structural prediction (Highplay/HighFold) for the de novo design of macrocyclic drugs targeting "undruggable" PPIs.