CIP-3
A 13-residue cyclic peptide that blocks the CD28 immune checkpoint, designed by a reinforcement-learning pipeline that generated 420 candidate sequences of which three were made. Its defining property is not potency but reversibility — it washes out in hours, where antibodies against the same target persist for weeks.
CIP-3 is an experimental ring-shaped peptide, designed by a computer program, that blocks a switch called CD28 which immune cells need in order to switch on. Dialling that switch down calms the immune system, which could help in diseases where it attacks the body. Unlike antibody drugs aimed at the same switch, CIP-3 wears off within hours. It has only been tested in cells and mice — there are no human trials.
What is CIP-3?
CIP-3 is a cyclic peptide antagonist of CD28, the costimulatory receptor that supplies the second signal a T cell needs to activate fully. It was reported by Kuncewicz, Upadhyay, Zhu, Duan and Gabr in Advanced Science in 2026 (epub 30 May, issue August 2026), from the Molecular Imaging Innovations Institute at Weill Cornell Medicine with collaborators at Macao Polytechnic University and the University of Gdansk. The sequence is CTENWRRYDGPQC — thirteen residues, roughly 1.6 kDa, closed by a disulfide bond. CD28 blockade is a clinically validated idea: abatacept, a CTLA4-Fc fusion protein approved in 2005, works by soaking up CD28's ligands. CIP-3 approaches the axis from the other side, binding CD28 itself. What makes it interesting is not affinity but kinetics. After washout, CIP-3-treated T cells recovered to about 85% of normal signalling while cells treated with the benchmark antagonist antibody FR104 recovered to only about 30% — the paper's central claim is controllable, exposure-dependent immune modulation rather than the prolonged receptor occupancy that characterises biologics. CIP-3 is preclinical: mouse and in vitro work only, with no investigational new drug application, no registered trial and no human data of any kind.
What CIP-3 Is Investigated For
CIP-3 is a preclinical proof of concept, and reading it as anything more is the main way this molecule gets misrepresented. The target is well validated — CD28 costimulation blockade underpins abatacept in rheumatoid arthritis and belatacept in kidney transplantation — and the reported data are internally coherent: a dissociation constant near 80–110 nM by two orthogonal methods, a CD28–CD80 competition IC50 of 609 nM, a Jurkat reporter IC50 of 443 nM, cytokine suppression in primary cells from ten healthy donors and five patients with ulcerative colitis, and dose-dependent benefit in a T-cell transfer colitis model in mice. Three things deserve emphasis because they are routinely garbled. First, the comparison with the benchmark anti-CD28 antibody FR104 was made in human peripheral blood mononuclear cells, not in the colitis model — the animal study had no antibody comparator arm, and the authors list head-to-head benchmarking as an open limitation. Second, only CD28–CD80 binding was tested; the authors explicitly flag CD86, potentially the dominant physiological ligand, as untested, along with selectivity against CTLA-4. Third, nothing in the paper concerns oral delivery: CIP-3 is dosed by daily subcutaneous injection with a terminal half-life of about 3.9 hours. The honest framing is that this is a methodology demonstration for designing cyclic peptides against a protein-protein interface long considered undruggable, not a clinical candidate.
History & Discovery
CD28 has a long history as a target and a short one as a druggable surface. The two-signal model emerged from work in the 1980s showing that T cells recognising antigen without costimulation become anergic rather than activated, and the B7-CD28 axis was mapped through the 1990s. The therapeutic payoff came indirectly: rather than blocking the receptor, Bristol-Myers Squibb blocked its ligands with CTLA4-Ig, producing abatacept, approved in 2005 for rheumatoid arthritis, and later belatacept for kidney transplantation. Targeting CD28 itself proved harder and, once, catastrophic. In March 2006 the CD28 superagonist TGN1412 put all six healthy volunteers in a first-in-human trial into intensive care within hours. The mechanism was eventually traced to a species difference — CD28 is not expressed on the CD4+ effector memory T cells of the animals used for preclinical testing — and the episode reshaped first-in-human dosing practice across the industry. Antagonist antibodies followed a more careful path: FR104 and lulizumab pegol both completed first-in-human studies without cytokine-release syndrome, and the field's current state is captured by the CTOT-24 trial reported in 2026, in which five of eight kidney transplant recipients on selective CD28 blockade discontinued early. The chemistry problem is separate from the biology. CD28's interface is a flat, solvent-exposed dimer surface, which is why the Gabr lab's own 2026 small-molecule paper describes the receptor as having long been regarded as undruggable, restricting modulation to biologics. Cyclic peptides are the natural intermediate — large enough to cover a protein-protein interface, small enough to synthesise and modify. CIP-3 was reported in Advanced Science in 2026 by Kuncewicz, Upadhyay, Zhu, Duan and Gabr, with Kuncewicz as first author holding joint appointments at Weill Cornell's Molecular Imaging Innovations Institute and the University of Gdansk, and Zhu and Duan at Macao Polytechnic University, whose group built the computational platform. The method was Highplay — Monte Carlo tree search reinforcement learning paired with the HighFold cyclic-peptide structure predictor — generating 420 candidate sequences from the CD28 sequence, of which three were synthesised. CIP-1 did not bind, CIP-2 bound micromolar, CIP-3 bound at roughly 100 nM. The same laboratory has published a parallel small-molecule CD28 programme, including an ultra-large virtual screen of around 48 million compounds, so CIP-3 is one output of a sustained effort on the target rather than an isolated result.
How It Works
A T cell needs two signals before it will switch on: one from its antigen receptor recognising a threat, and a second through CD28, which acts as a go-ahead. Remove the second signal and the T cell stalls rather than activating. CIP-3 is a small ring-shaped peptide that sits on CD28 and blocks that go-ahead. Because it is small and clears quickly, the block can be dialled up or allowed to wear off within hours.
CD28 binds CD80 (B7-1) and CD86 (B7-2) on antigen-presenting cells, supplying the costimulatory signal that, alongside T-cell receptor engagement, licenses full activation, IL-2 production and clonal expansion. A T cell receiving the receptor signal without costimulation enters anergy instead — the two-signal model established by Jenkins and Schwartz in the 1980s. Abatacept and belatacept work upstream of the receptor, binding CD80 and CD86 so that endogenous CD28 cannot. Belatacept differs from abatacept by two substitutions that raise avidity for both ligands, with the larger gain against CD86. CIP-3 inverts the strategy, binding CD28 directly. The theoretical attraction is selectivity: CTLA4-Ig agents necessarily block the ligands that CTLA-4 also uses for inhibitory signalling, whereas a CD28-specific antagonist leaves that regulatory arm intact — the rationale behind the antagonist antibodies FR104 and lulizumab pegol. The reported pharmacology is as follows. Binding to the CD28 extracellular domain gave a dissociation constant of approximately 108 nM by microscale thermophoresis and about 80 nM by isothermal titration calorimetry with 1:1 stoichiometry. CIP-3 inhibited the CD28–CD80 interaction with an IC50 of 609 nM and blocked signalling in a Jurkat CD28 NFAT-luciferase reporter with an IC50 of 443 nM. In primary human PBMCs from ten healthy donors, and in cells from five patients with ulcerative colitis, it suppressed IL-2 and IFN-γ dose-dependently; in the ulcerative colitis cells, IL-2 inhibition at 1 µM was statistically indistinguishable from FR104 at 10 µg/mL. It showed no intrinsic agonism on unstimulated cells. The reversibility result is the paper's distinguishing finding: after washout, CIP-3-treated cells recovered to roughly 85% of stimulated control signalling, while FR104-treated cells recovered to about 30%. Pharmacokinetics in mice are consistent — terminal half-life 3.9 hours, plasma concentrations above the roughly 500 nM functional threshold for 10 to 12 hours at 5 mg/kg and 4 to 6 hours at 1 mg/kg. The design route is worth stating precisely, because it is the most novel element and is frequently described wrongly. Candidates were generated by the Highplay platform — Monte Carlo tree search reinforcement learning with a transformer policy-value network, using the HighFold structure-prediction model as reward proxy, targeting PDB structure 1YJD chain C. Sequences were filtered on interface pLDDT, hydrogen-bond counts and interface hydrogen bonds. From 420 designed sequences, three were synthesised. This was one computational pass, not an iterative design cycle, and RFdiffusion — the best-known de novo design method — was cited as background but not used. No co-crystal structure exists; the binding mode is computational modelling, which the authors flag as requiring mutagenesis and high-resolution structural work to confirm.
Evidence Snapshot
Human Clinical Evidence
None. No investigational new drug application, no registered clinical trial, no human exposure. The only human material studied is ex vivo: peripheral blood mononuclear cells from ten healthy donors and five patients with ulcerative colitis.
Animal / Preclinical
A single published study, concrete but limited. In the CD4+CD45RBhigh T-cell transfer model of chronic colitis in SCID mice, daily subcutaneous CIP-3 at 1 or 5 mg/kg over 30 days (n=8 per group) reduced disease severity dose-dependently, restored colon length and lowered TNF-α and IL-6. There was no antibody comparator arm in the animal work — the comparison with FR104 was made in cultured human cells — and the authors note that disease scoring was semi-subjective and unblinded, with histopathology, immune-cell profiling and benchmarking against established agents still to be done.
Mechanistic Rationale
Strong for the target, unproven for the molecule. CD28 costimulation blockade is validated in humans by abatacept in rheumatoid arthritis and belatacept in kidney transplantation. CD28 itself has long been considered undruggable by small molecules because of its flat, solvent-exposed dimer interface, which is what makes a designed macrocycle an interesting approach rather than a redundant one. What is not established is selectivity against CTLA-4, activity against CD86-mediated engagement, or any translation beyond mice.
Research Gaps & Open Questions
What the current literature has not yet settled about CIP-3:
- 01Everything clinical. There is no investigational new drug application and no registered trial, so Phase 1 data is not pending — it has not been sought.
- 02Selectivity against CTLA-4 and other B7-family receptors, which the authors name as necessary future work. A molecule blocking CTLA-4 alongside CD28 would produce the opposite immunological effect to the one intended.
- 03Activity against CD28–CD86 engagement. Only CD80 was tested, and the authors flag CD86 as potentially the dominant physiological ligand.
- 04A structural confirmation of the binding mode. There is no co-crystal; the interaction is computational modelling, and the authors call for mutagenesis and high-resolution structural work.
- 05Head-to-head comparison against FR104 or abatacept in an animal disease model. The published comparison is in cultured human cells only, and the animal study had no antibody arm.
- 06Whether a roughly four-hour half-life is workable for a chronic autoimmune indication, or whether the reversibility that distinguishes the molecule also makes it impractical to dose.
- 07Whether oral delivery is achievable at all, which has been asserted for this molecule but never tested.
- 08Which indication it would be developed for. Colitis was the model, but no clinical positioning has been proposed, and the human comparator field — abatacept, belatacept, the antagonist antibodies — is crowded and only partly successful.
- 09Whether AI-designed cyclic peptides translate. No molecule from this class has entered a registered human trial, so the methodology's clinical track record is empty rather than poor.
Forms & Administration
Subcutaneous injection in all published in vivo work — daily dosing at 1 or 5 mg/kg in the colitis model. No other route has been tested and no formulation work has been published. Claims that CIP-3 is or will be orally available are extrapolation from the cyclic-peptide field generally, not findings about this molecule: its terminal half-life in mice is about 3.9 hours and the paper makes no oral bioavailability claim.
Dosing & Protocols
The ranges below reflect protocols commonly discussed in the literature and by clinicians — not a prescription. Actual dosing for any individual should be determined by a qualified healthcare provider who knows the patient.
Typical Range
Not established in humans. In the published mouse colitis study, 1 or 5 mg/kg subcutaneously.
Frequency
Daily subcutaneous injection in the animal work, consistent with a terminal half-life of roughly 3.9 hours.
Timing Considerations
No specific timing requirements: can be administered at any time of day, with or without food, and is not tied to exercise timing. Consistency matters more than the specific clock — dose at roughly the same time each day (or same day each week, for weekly protocols) to keep exposure steady.
Cycle Length
Not established.
Protocol Notes
There is no human dosing information because there has been no human exposure. The numbers above are from a single mouse study and are recorded for completeness, not as guidance. One pharmacokinetic detail is worth understanding because it shapes what this molecule could become. In mice, a 5 mg/kg subcutaneous dose kept plasma concentrations above the roughly 500 nM functional threshold for 10 to 12 hours, and a 1 mg/kg dose for only 4 to 6 hours. That is the trade-off at the centre of the design: the rapid reversibility the authors present as the molecule's advantage over antibodies is the same property that would require frequent dosing in a chronic autoimmune disease. Whether that is a feature — allowing immune suppression to be dialled back quickly if infection or another complication arises — or a practical obstacle is exactly what a clinical programme would have to establish. The pharmacokinetics were also measured in immunocompetent C57BL/6 mice while efficacy was tested in immunodeficient SCID mice, so the exposure-response relationship within the disease model is inferred rather than directly measured. The authors say so explicitly.
CIP-3 is a preclinical research compound. It is not approved anywhere, has no investigational new drug application, and is not in any registered clinical trial. It is not available for human use through any legitimate channel.
Timeline of Effects
Onset
Not characterised in humans. In cell-based assays the block is immediate on exposure, and in mice plasma concentrations peak 0.5 to 1 hour after subcutaneous dosing.
Peak Effect
Not characterised in humans. The mouse colitis study dosed daily for 30 days and tracked disease severity across that period rather than identifying a peak effect.
After Discontinuation
Fast — and this is the molecule's defining claim rather than an incidental property. In washout experiments, CIP-3-treated cells recovered to roughly 85% of stimulated control signalling, while cells treated with the benchmark antagonist antibody FR104 recovered to only about 30%. The terminal half-life in mice is about 3.9 hours, and plasma levels fall below the functional threshold within 4 to 12 hours depending on dose. This is the opposite of the pharmacology of CTLA4-Fc fusion proteins and antagonist antibodies, where receptor occupancy and immune suppression persist for weeks after the last dose.
Common Questions
How is CIP-3 different from abatacept?
They work on the same axis from opposite ends, and at very different scales. Abatacept (Orencia) is a CTLA4-Fc fusion protein of roughly 92 kDa given by infusion or subcutaneous injection; it binds CD80 and CD86 on antigen-presenting cells so that CD28 cannot reach them. CIP-3 is a 13-residue cyclic peptide of about 1.6 kDa that binds CD28 itself. The practical difference the paper emphasises is duration: a peptide with a roughly four-hour half-life can be titrated and stopped, while a fusion protein or antibody occupies its target for weeks. Blocking CD28 directly rather than its ligands also has a theoretical advantage — it leaves CTLA-4's inhibitory signalling intact, which CTLA4-Ig agents necessarily disrupt, and that argument is the basis for the whole selective-CD28-antagonist field.
Why does a CD28 drug make immunologists nervous?
Because of TGN1412. In March 2006, six healthy volunteers in a first-in-human trial of that anti-CD28 monoclonal antibody developed a systemic inflammatory response within 90 minutes of infusion and were critically ill within 12 to 16 hours with pulmonary infiltrates, renal failure and disseminated intravascular coagulation. Preclinical testing had entirely failed to predict it; the explanation established later was that CD28 is not expressed on the CD4+ effector memory T cells of the species used for safety testing, so the relevant cells could not be stimulated in animals. The crucial distinction is that TGN1412 was a superagonist — designed to activate CD28 without T-cell receptor engagement — whereas CIP-3 is an antagonist. The CIP-3 paper takes the history seriously enough to test for it directly, reporting no cytokine rise in unstimulated human PBMCs across 0.01 to 10 µM and no detectable IL-6, TNF-α, IFN-γ or IL-2 elevation in healthy mice at 5 and 10 mg/kg.
What does 'AI-designed' actually mean here?
Something specific and documented, which is not always true of that phrase. The team used the Highplay platform, combining Monte Carlo tree search-based reinforcement learning with the HighFold cyclic-peptide structure-prediction model, generating candidate sequences directly from the CD28 amino-acid sequence and ranking them on predicted interaction energy, interface hydrogen bonding and structural confidence. From an initial library of 420 designed sequences, three were prioritised and synthesised: CIP-1 showed no detectable binding, CIP-2 bound in the micromolar range, and CIP-3 was the hit. Note what this is not — it is a single computational pass followed by synthesis and testing, not iterative design-build-test cycles, and it does not use RFdiffusion, which the paper cites only as background. A linear, non-cyclised version of the same CIP-3 sequence failed to suppress cytokines at all, so the macrocycle is doing real work.
Could it be taken as a pill?
Nothing in the published work addresses that, and claims that it could are extrapolation rather than reporting. CIP-3 was administered by daily subcutaneous injection in every in vivo experiment, has a terminal half-life of roughly 3.9 hours in mice, and the paper makes no oral bioavailability claim. Oral delivery is a genuine aspiration for the cyclic-peptide field generally — cyclisation and backbone modification do improve metabolic stability and membrane permeability, and a handful of cyclic peptides have achieved oral dosing — but it is an open research problem, not a property CIP-3 has been shown to have.
Has any AI-designed drug actually reached patients?
Some have, which makes the picture more interesting than either the hype or the dismissal suggests. An analysis of AI-discovered molecules found roughly 80 to 90% Phase 1 success — substantially above historical industry averages — but about 40% in Phase 2, which is unremarkable. That pattern is the honest summary of the field: AI is demonstrably good at producing molecules with drug-like properties and has not yet shown it can pick targets or predict efficacy better than anyone else. A generative-AI-discovered TNIK inhibitor, rentosertib, completed a randomised Phase 2a trial in idiopathic pulmonary fibrosis in 2025. Separately, a computationally designed self-assembling protein nanoparticle forms the basis of an approved COVID-19 vaccine in South Korea. What has not happened is an AI-designed cyclic peptide reaching approval, or as far as the registries show, a registered human trial.
Who CIP-3 Is NOT For
- •All human use. CIP-3 is preclinical, with no safety data, no established dose and no regulatory status — this is the only contraindication that matters in practice.
- •Active malignancy would be a contraindication for any CD28 antagonist, since attenuating T-cell costimulation works against tumour immune surveillance.
- •Active infection, for the same reason in the opposite direction — immune attenuation impairs clearance.
- •Live vaccines, which are generally avoided during immune attenuation.
- •Pregnancy and breastfeeding — no data of any kind.
- •Paediatric use — not characterised.
Drug & Supplement Interactions
No drug interaction work has been done, and none is meaningful for a compound with no human exposure. What can be said is mechanistic and class-level. Combining a CD28 antagonist with other immunosuppressants would be expected to compound immune attenuation and infection risk, and the CTOT-24 trial is a concrete illustration — selective CD28 blockade combined with IL-6 receptor antagonism in kidney transplant recipients produced early discontinuations for acute rejection and severe neutropenia in five of eight patients. Live vaccines are conventionally avoided during costimulation blockade. The interaction that would matter most is pharmacological rather than pharmacokinetic. CD28 antagonism and checkpoint-inhibitor cancer immunotherapy work in opposite directions — PD-1 blockade is thought to act substantially by restoring CD28 costimulation, a relationship reinforced by 2026 work showing CD28 signalling is organised by intracellular condensates that PD-1 counteracts. A CD28 antagonist would be expected to blunt the effect of anti-PD-1 therapy, which is a reason the two would not be combined and an argument for the reversibility CIP-3 is designed around.
Safety Profile
Common Side Effects
Cautions
- • Preclinical only. There is no investigational new drug application, no registered trial, and no human exposure.
- • CD28 is a target with a catastrophic clinical precedent. The 2006 TGN1412 trial put six healthy volunteers in intensive care within hours. That molecule was a superagonist and CIP-3 is an antagonist, but the episode is the reason any CD28-directed agent faces an unusually high bar for first-in-human design.
- • Animal data on CD28 are poorly predictive by default. CD28 is not expressed on CD4+ effector memory T cells in the species conventionally used for preclinical safety testing, which is precisely why TGN1412's preclinical package missed the risk — a caution that applies directly to CIP-3's mouse colitis data.
- • Selectivity against CTLA-4 and other B7-family receptors has not been tested, and the authors identify it as necessary future work. A molecule that blocked CTLA-4 as well as CD28 would have the opposite immunological effect from the one intended.
- • Only CD28–CD80 engagement was assessed; CD86 may be the dominant physiological ligand and was not studied.
- • Sustained CD28 blockade carries the general risks of immune attenuation — infection and, in principle, impaired tumour surveillance.
- • The pharmacokinetic work was done in immunocompetent C57BL/6 mice while efficacy was tested in immunodeficient SCID mice, so the exposure-response relationship in the disease model is inferred rather than measured.
- • Colitis disease-activity scoring in the published study was semi-subjective and not blinded, by the authors' own statement.
What We Don't Know
Everything clinical. No dose, route, schedule, safety profile or efficacy signal in humans exists. Beyond that: whether CD86-mediated CD28 engagement is blocked as effectively as CD80-mediated engagement; whether the molecule is selective against CTLA-4; how it compares head-to-head against FR104 or abatacept in an animal disease model rather than in cultured cells; whether the binding mode inferred from computational modelling survives a co-crystal structure; and whether a four-hour half-life is a feature or a dosing problem in a chronic autoimmune disease.
Legal Status
United States
Preclinical research compound. Not FDA-approved, not under an investigational new drug application, and not in any registered clinical trial. It has no legal status as a medicine of any kind.
International
Not approved or authorised as a medicine anywhere.
Sports & Competition
Not named on the WADA Prohibited List, and CD28 appears nowhere in the 2027 List. That absence is not permission: S0 prohibits at all times any pharmacological substance not addressed elsewhere on the List and with no current approval by any governmental regulatory health authority for human therapeutic use, naming drugs under pre-clinical or clinical development among its examples. A preclinical research peptide falls squarely inside that definition. There is also no performance rationale — suppressing T-cell costimulation does nothing useful for an athlete and would increase infection risk.
Regulatory status changes over time. Verify current local rules with a qualified professional.
Myths & Misconceptions
Myth
CIP-3 matched an anti-CD28 antibody in the colitis model.
Reality
It was not compared with one there. The colitis study tested CIP-3 at 1 and 5 mg/kg against vehicle, with no antibody comparator arm, and the authors list benchmarking against established therapeutic agents as future work. The comparison with the antagonist antibody FR104 was a separate experiment in cultured human cells, where CIP-3 at 1 µM suppressed IL-2 in ulcerative colitis patient PBMCs to a degree not statistically different from FR104 at 10 µg/mL. Those are different experiments in different systems, and merging them overstates what was shown.
Myth
Being a small cyclic peptide means CIP-3 can be taken orally.
Reality
The published work neither tests nor claims oral delivery. Every in vivo experiment used daily subcutaneous injection, and the terminal half-life in mice is about 3.9 hours. Oral bioavailability is a real and active goal for cyclic peptides as a class, and progress is being made, but it is a property that has to be engineered and demonstrated molecule by molecule — not one that follows automatically from being cyclic or being small.
Myth
AI-designed peptides are likely to work because the design is computationally optimised.
Reality
The measured record argues for modest expectations. Across AI-discovered molecules generally, Phase 1 success runs around 80 to 90% — genuinely better than historical averages, which is what you would expect if the technology is good at generating molecules with drug-like properties — while Phase 2 success is around 40%, essentially the industry norm. In other words, computational design helps with the part of the problem that is about chemistry and not the part that is about biology. For CIP-3 specifically it is worth noting that of the three designed peptides synthesised, one did not bind at all.
Myth
CIP-3 is in early clinical trials.
Reality
It is not in trials of any kind. There is no investigational new drug application, no registered study and no human exposure; the published work is in vitro assays, human cells in culture, and mice. The paper was first published online in May 2026 and appeared in an August 2026 issue. Descriptions of the molecule as 'preclinical to early clinical' blur a distinction that matters a great deal to anyone assessing how far along it is.
Myth
CIP-3 will replace abatacept for autoimmune disease.
Reality
Abatacept has two decades of approved use across rheumatoid arthritis, psoriatic arthritis and juvenile idiopathic arthritis, supported by placebo-controlled trials with radiographic endpoints. CIP-3 has one paper. Beyond the evidence gap, the two are not straightforwardly substitutable: abatacept blocks CD80 and CD86, CIP-3 binds CD28, and CIP-3's selectivity against CTLA-4 has not been tested. The molecule's interest lies in demonstrating that a designed macrocycle can engage a protein interface considered undruggable, not in displacing an established biologic.
Myth
Any drug targeting CD28 risks repeating the TGN1412 disaster.
Reality
The direction of the pharmacology is what matters. TGN1412 was a superagonist built to activate CD28 without T-cell receptor engagement, which is why it triggered massive cytokine release. Antagonists do the opposite, and the record since bears that out: FR104 and lulizumab pegol both completed first-in-human studies without cytokine-release syndrome. The CIP-3 paper tested for agonism specifically, finding no cytokine rise in unstimulated human PBMCs or in healthy mice. The durable lesson of TGN1412 is narrower and still applies — animal models underpredict human CD28 responses, because the species used for safety testing do not express CD28 on the CD4+ effector memory T cells that drove the reaction.
Published Research
28 studiesCD28 costimulation is organized by signaling condensates and counteracted by PD-1
Regulatory T cell modulation with selective CD28 blockade and IL-6 receptor antagonist in kidney transplant recipients: Results of CTOT-24
Chandran S et al., Am J Transplant 2026. A sobering measure of where CD28 antagonism actually stands in humans two decades on: of eight treated kidney transplant recipients, five discontinued early, mostly for acute T-cell-mediated rejection or severe neutropenia, and the authors conclude that alternative strategies are needed.
Cyclic Peptides in Modern Drug Discovery: Trends and Therapeutic Directions
AI-Enforced Ultra-Large Virtual Screening Discovers Potent CD28 Binders
Peptides as programmable molecular scaffolds: from chemical synthesis and engineering to translational medicine
AI-Designed Cyclic Peptides Enable Controllable Modulation of the CD28 Immune Checkpoint
Kuncewicz K, Upadhyay S, Zhu R, Duan H and Gabr MT, Advanced Science 2026;13(48):e75892 (epub 30 May 2026). The primary and only CIP-3 paper. Reports a Kd near 80–110 nM by two orthogonal methods, CD28–CD80 disruption at 609 nM, reporter-assay IC50 of 443 nM, cytokine suppression in healthy-donor and ulcerative-colitis PBMCs, and dose-dependent benefit in a mouse T-cell transfer colitis model. The headline property is rapid reversibility — about 85% signalling recovery after washout versus roughly 30% for the benchmark antibody FR104. Note that the FR104 comparison is in cultured human cells; the animal study had no antibody comparator.
Overcoming the undruggable barrier: Structure-guided discovery of a potent small molecule CD28 antagonist with translational potential
CD28 and ICOS in immune regulation: Structural insights and therapeutic targeting
A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial
How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons
Jayatunga MKP et al., Drug Discovery Today 2024. The only quantitative answer to what AI-designed drugs actually achieve in the clinic: roughly 80–90% Phase 1 success, well above historical industry averages, but about 40% in Phase 2 on a limited sample — comparable to everyone else. The pattern says AI is good at making drug-like molecules and has not yet shown it can predict efficacy.
De novo design of protein structure and function with RFdiffusion
Trends in peptide drug discovery
Selective Costimulation Blockade With Antagonist Anti-CD28 Therapeutics in Transplantation
Vanhove B et al., Transplantation 2019. Sets out the argument CIP-3 inherits: selective CD28 antagonists were developed on the rationale that preserving CTLA-4-mediated regulatory mechanisms would give better control of alloreactivity than blocking CD80/86 does, and would be compatible with regulatory T-cell function.
First-in-Human Study in Healthy Subjects with FR104, a Pegylated Monoclonal Antibody Fragment Antagonist of CD28
Poirier N et al., J Immunol 2016. FR104 is the benchmark antibody against which CIP-3 was compared in human cells, so this is the reference point for the whole comparison. It blunts CD28 costimulation while sparing CTLA-4 and PD-L1 coinhibitory signals, and was well tolerated in healthy subjects with no evidence of cytokine-release syndrome.
Costimulation Blockade in Autoimmunity and Transplantation: The CD28 Pathway
Pharmacokinetic, Pharmacodynamic, and Safety Profile of a Novel Anti-CD28 Domain Antibody Antagonist in Healthy Subjects
CD28 Costimulation: From Mechanism to Therapy
Esensten JH et al., Immunity 2016. The standard modern account of CD28 biology and its therapeutic record, and unusually candid about that record: translation of CD28 understanding into immunomodulatory therapeutics has been uneven, with both successes and failures.
Belatacept and Long-Term Outcomes in Kidney Transplantation
The storm has cleared: lessons from the CD28 superagonist TGN1412 trial
Monoclonal antibody TGN1412 trial failure explained by species differences in CD28 expression on CD4+ effector memory T-cells
Eastwood D et al., Br J Pharmacol 2010. The explanation for why the preclinical package missed the catastrophe: unlike humans, the species used for preclinical safety testing do not express CD28 on CD4+ effector memory T cells, so those cells could not be stimulated. Directly relevant to interpreting any mouse CD28 result, including CIP-3's colitis data.
A phase III study of belatacept-based immunosuppression regimens versus cyclosporine in renal transplant recipients (BENEFIT study)
'Cytokine storm' in the phase I trial of monoclonal antibody TGN1412: better understanding the causes to improve preclinical testing of immunotherapeutics
Cytokine storm in a phase 1 trial of the anti-CD28 monoclonal antibody TGN1412
Suntharalingam G et al., NEJM 2006. The report that defines the risk landscape for every CD28-directed drug since. Within 90 minutes of a single intravenous dose, all six healthy volunteers had a systemic inflammatory response; within 12 to 16 hours they were critically ill with pulmonary infiltrates and lung injury, renal failure and disseminated intravascular coagulation. TGN1412 was a superagonist rather than an antagonist, but the episode is why antagonist design and first-in-human dosing in this class are approached the way they are.
Effects of abatacept in patients with methotrexate-resistant active rheumatoid arthritis: a randomized trial
Abatacept for rheumatoid arthritis refractory to tumor necrosis factor alpha inhibition
Treatment of rheumatoid arthritis by selective inhibition of T-cell activation with fusion protein CTLA4Ig
The B7-CD28 superfamily
Antigen presentation by chemically modified splenocytes induces antigen-specific T cell unresponsiveness in vitro and in vivo
Quick Facts
- Class
- AI-Designed CD28 Cyclic Peptide Antagonist
- Tier
- D
- Evidence
- Preliminary
- Safety
- Limited Data
- Updated
- Oct 2026
- Citations
- 28PubMed
Also known as
Tags
Related Goals
Evidence Score
Clinical Trials
View Clinical TrialsLinks to ClinicalTrials.gov for reference. Listing does not imply endorsement.