AI-Designed Cyclic Peptides Are Cracking Undruggable Targets

Cyclic peptides — macrocyclic and bicyclic forms specifically — sit at a chemical sweet spot that small molecules and antibodies both miss. They're large enough to bury inside the flat protein-protein interfaces that conventional small-molecule drug design has struggled with for thirty years (think KRAS, beta-catenin, MYC, the bcl-2 family). They're small enough to permeate cell membranes if you get the chemistry right — something antibodies can never do. And they're more stable against proteases than linear peptides, often by an order of magnitude.

The historical problem hasn't been the molecules. It's been the manufacturing and the design loop. Synthesizing a single drug-like cyclic peptide typically required 20+ chemistry steps, multiple protecting groups, and chromatography purification at every stage. Designing them required brute-force screening of libraries built one molecule at a time, with success rates measured in fractions of a percent.

What changed in the last four weeks: three papers landed showing that both halves of the problem are converging on a different cost curve. AI is closing the design loop. Biocatalysis is collapsing the synthesis cost. And the targets these methods are hitting — CD28 immune checkpoint, WDR5/MLL oncogenic interaction, PCSK9 — are exactly the protein-protein interfaces that defined 'undruggable' for the previous era.

A March 2026 bioRxiv preprint reported an AI-guided strategy for cyclic peptide antagonists of the CD28 immune checkpoint. The lead candidate, CIP-3, binds the CD28 extracellular domain at nanomolar affinity and produces controllable modulation in cellular assays. CD28 is the primary T-cell co-stimulatory receptor, and the current best biologics in this space — abatacept and belatacept — are CTLA-4-Ig fusion proteins. They work, but they're large recombinant biologics with all the dosing, infusion-pump, immunogenicity, and cold-chain logistics that come with that format.

The interesting thing about CIP-3 isn't just the affinity number. It's the design pipeline that produced it. The group used an AI sampling strategy that explored cyclic backbones and sequence space simultaneously, generated candidates, ranked them by predicted binding to the CD28 extracellular domain, and produced a small set of synthesizable hits. The CIP-3 hit rate from that pipeline was vastly higher than the historical screening yield from random library work. The work is preprint-stage and the binding data needs orthogonal validation; the experimental readout, though, looks substantive.

If CIP-3 (or one of its analogs) ends up in clinical development, the practical implication is large. A cyclic peptide CD28 antagonist with sub-cutaneous dosing kinetics would be a different proposition than a quarterly infusion of belatacept for transplant patients or weekly abatacept for autoimmune indications. That's still a long way off, but the design pipeline that produced it is the news.

An April 2026 Nature Communications paper from a Boston-area group described CycloSEL — Cyclic Self-Encoded Libraries — an end-to-end workflow that screens 16-million-member synthetic macrocycle libraries enriched for 'beyond rule of five' drug-like features. The platform uses affinity selection and tandem mass spectrometry for hit identification, eliminating the genetic barcoding step that other macrocycle-screening platforms (DEL libraries, mRNA display) require.

The validation target: carbonic anhydrase IX, a tumor-overexpressed enzyme. The platform produced clean hit enrichment with low false-positive rates. The interesting application came later: applied to the WDR5/MLL oncogenic protein-protein interaction in acute myeloid leukemia, CycloSEL produced a hit with 8 nM affinity that was rapidly optimized to a compound (IC50 128 nM) with excellent serum stability, passive membrane permeability, and anti-proliferative activity in leukemia cell lines.

WDR5 is on every short list of 'undruggable but high-value' targets. The MLL fusion-protein interaction it stabilizes drives MLL-rearranged leukemias, which are aggressive and poorly served by current chemotherapy. Conventional small-molecule discovery hasn't produced a clinical-stage WDR5 inhibitor that works the way medicinal chemists want one to work. A passively permeable macrocyclic peptide hit from a 48-day discovery campaign is the kind of result that changes which projects pharma boards will fund.

Merck's May 7, 2026 Science paper on the biocatalytic cascade synthesis of enlicitide decanoate isn't an AI-design story. It's the other half of the problem: how do you actually manufacture a cyclic peptide drug at commercial scale once the design is solved.

Enlicitide is an investigational oral PCSK9 inhibitor — a macrocyclic peptide that binds the LDL-receptor-PCSK9 interface and stops PCSK9 from degrading LDL receptors, the same mechanism as the injectable PCSK9 antibodies (Praluent, Repatha). Phase 3 CORALreef data published in NEJM February 2026 showed 57% LDL-C reduction at 24 weeks vs placebo. If approved, it would be the first oral PCSK9 inhibitor.

What the Science paper describes is a convergent biocatalytic cascade — a tailored set of engineered enzymes that catalyze selective peptide fragment formation, coupling, and macrocyclization in a protecting-group-free sequence, combined with chromatography-free crystallizations. The result: a kilogram-scale GMP synthesis that cuts step count by more than half versus prior chemistry. Joelle Pelletier's accompanying Science Perspective on May 8 framed the cascade as a template for the entire oral macrocyclic peptide modality. Her argument: the limiting step for oral cyclic peptides has been manufacturing cost, not pharmacology, and biocatalysis collapses the cost curve.

That framing matters. If the chemistry of a 14-amino-acid macrocyclic peptide can be done in 12 enzymatic steps at $X per gram instead of 28 protected-residue steps at $5X per gram, the gating economics on the entire field change. Programs that didn't pencil out at the old chemistry start to make sense.

Layered alongside the AI and biocatalysis news: Chugai's LUNA18 (paluratide) reached kilogram-scale GMP supply at >98.5% purity through a 24-step convergent liquid-phase synthesis, per an Organic Process Research & Development paper presented at TIDES USA 2026 on May 11. LUNA18 is an N-alkyl-rich cyclic undecapeptide that binds RAS proteins — KRAS, NRAS, HRAS — at the inactive-state RAS-GEF interaction interface and is in Phase 1 dose escalation for advanced solid tumors.

KRAS has been the test case for 'undruggable' targets for forty years. The first KRAS small-molecule inhibitors (sotorasib, adagrasib) hit only the G12C mutant and were a major achievement when they reached the clinic in 2021. LUNA18 binds wild-type and multiple mutant forms (G12D, G13D, others) and achieves 21-47% oral bioavailability without special formulation — bioavailability numbers that would have been considered impossible for an 11-amino-acid macrocyclic peptide a decade ago.

The Chugai team's published synthesis route — convergent 24-step liquid-phase, >30% overall yield — is a useful proof point alongside the Merck enzyme cascade. The two approaches solve the same manufacturing problem in different ways: one through engineered enzymes, the other through optimized telescoped chemistry. Both deliver kilogram-scale GMP material at quality that supports late-stage clinical and eventual commercial manufacture.

If LUNA18 produces clean clinical data, it joins a small group of pan-RAS molecules — Revolution Medicines' RMC-6236 is the closest analog — that could meaningfully change the standard of care in colorectal cancer, pancreatic cancer, and other RAS-driven malignancies.

Design and synthesis aren't quite separable problems anymore. A May 2026 Nature Communications paper introduced CycloPepper, a machine-learning-guided platform that predicts cyclization outcomes and accelerates automated synthesis of therapeutic cyclic peptides. The bottleneck the paper addresses: many promising linear sequences fail at the macrocyclization step or yield poorly under standard conditions, requiring expensive iterative chemistry that doesn't show up in the design literature.

CycloPepper trains on a curated dataset of cyclization outcomes — which sequences successfully cyclize under which conditions, which yield well, which fall apart — and integrates with automated synthesis platforms to enable closed-loop design-make-test cycles. It joins CyclicMPNN (a fine-tuned ProteinMPNN derivative for cyclic peptide sequence design) and AfCycDesign (an AlphaFold-derivative for cyclic backbone prediction) as part of the computational stack now available to cyclic-peptide drug discovery groups.

What's emerging is a layered toolkit: backbone-and-sequence design from generative models like RFDiffusion (adapted for cyclic backbones), CyclicMPNN, and AfCycDesign; cyclization-and-yield prediction from CycloPepper-style platforms; synthesis at scale via engineered enzymatic cascades (Merck enlicitide) or optimized telescoped chemistry (Chugai LUNA18); and discovery libraries that screen 8-figure compound counts against single targets in weeks rather than years (CycloSEL).

The individual pieces aren't new. The stack is new, and the combination is what's producing usable drug candidates against targets that didn't have any.

The honest picture isn't all good news. A few open questions:

Clinical translation rates aren't yet visible. Of the AI-designed cyclic peptides currently in the literature, a small but growing fraction is in Phase 1 (LUNA18, VX-670, AVA6000, a few others); none has yet produced a Phase 3 readout in an undruggable-target indication. The proof-of-concept that AI-designed cyclic peptides actually deliver clinical benefit at undruggable targets is still pending.

Validation bias in the design literature is real. Most AI-design papers report success on retrospective tests against known targets, which is qualitatively different from prospective binder discovery against novel targets. CycloSEL's WDR5 result is closer to prospective than retrospective; CIP-3's CD28 result is mixed (CD28 is a well-characterized target, but the specific antagonist scaffold is novel).

Manufacturing scale-up remains expensive even with biocatalysis. The Merck enlicitide cascade cuts cost, but the absolute cost of GMP cyclic peptide manufacture is still 10-100x that of small molecules at equivalent scale. Until that gap closes further, cyclic peptides will mostly compete in indications where the value of the target justifies the manufacturing premium — oncology, rare disease, autoimmune — rather than in cardiometabolic indications where small-molecule alternatives exist.

The regulatory pathway is solid but slow. FDA has a clear track record on cyclic peptide approvals (cyclosporine, dalbavancin, oritavancin, the macrocyclic glycopeptides) but the agency hasn't yet had a high-volume macrocyclic peptide approval pathway. Each new cyclic peptide drug currently navigates the system as a special case rather than as part of a well-worn route.

The trajectory is clear, though. The May 2026 cluster — CIP-3, CycloSEL, enlicitide biocatalysis, LUNA18 GMP synthesis, CycloPepper — represents a meaningful step change in the field's ability to design, screen, synthesize, and develop these molecules. The clinical readouts that confirm the strategy are coming over the next 18-24 months.