AI drug design has moved from a small-molecule story to a peptide story. Generative models — RFDiffusion adaptations for cyclic backbones, AfCycDesign, ProteinMPNN-derived sequence designers — now generate cyclic peptide candidates that survive experimental validation against targets that resisted traditional medicinal chemistry. Synthesis chemistry has advanced in parallel: one-pot ligations, photo-redox macrocyclizations, and biocatalytic cascades close the loop between in-silico design and tractable scaffolds.
Covered here: the May 2026 CycloPepper Nature Communications paper — a machine-learning-guided platform that predicts cyclization outcomes and accelerates automated synthesis — and CIP-3, an AI-designed cyclic peptide CD28 antagonist with nanomolar affinity reported on bioRxiv in March 2026. Earlier coverage included CyclicMPNN for stable cyclic peptide sequence generation and the broader Drug Discovery World May 2026 review on the algorithmic stack and its industrial implications.
Stories here cover the model architectures, the experimental validations, and the platforms moving AI peptide design out of academic labs. See #cyclic-peptide, #machine-learning, and #drug-discovery.
Researchers from the Chinese Academy of Sciences' Xinjiang Institute of Ecology and Geography and the Chinese Academy of Medical Sciences reported in ACS Omega on October 23, 2026 that NA-CATH, a cathelicidin peptide first found in the Chinese cobra, was the best of five antimicrobial peptides tested on clinical multidrug-resistant Klebsiella pneumoniae and Acinetobacter baumannii in the lab, but killed all treated wax moth larvae. Screening nearly 50,000 computer-designed variants produced NA-7, which kept antibacterial activity, improved survival of Klebsiella-infected larvae compared with the parent peptide, and caused lower acute toxicity and fewer kidney changes in mice. The work is preclinical.
Researchers at Houston Methodist Hospital and the National and Kapodistrian University of Athens reported in Frontiers in Microbiology on October 19, 2026 that CIT8, a 13-amino-acid antimicrobial peptide previously tuned by machine learning against MRSA, also kills drug-resistant Gram-negative bacteria. It inhibited 15 of 16 Acinetobacter baumannii strains at 4 µg/mL, including carbapenem-resistant clinical isolates, and at five times that dose cut about 50 million bacteria per mL to undetectable levels within 15 minutes by breaking their membranes. CIT8 and a mirror-image version also disrupted biofilms with relatively low toxicity to human cells; the work is in the lab.
Researchers at Shanghai Jiao Tong University reported in Nature Communications on October 17, 2026 a deep-learning model that designs antimicrobial lipopeptides, peptides with a fatty-acid tail, by accounting for how the tail changes activity. Starting from 2,657 bacterial peptide sequences mined from 96,911 genomes, the model screened 6,197 virtual lipopeptides; of 46 the team made, 44 killed bacteria, a 95.7% hit rate. The lead, LP630, worked against Gram-positive and Gram-negative bacteria and mycobacteria and treated infected wounds in mice.
Researchers from Merck's discovery chemistry group and Texas Tech University reported in npj Drug Discovery on October 12, 2026 PepOpt, a machine-learning framework for peptide lead optimization that combines chemical-structure and sequence descriptions in simple linear models. The authors report that it predicts activity accurately, can handle noncanonical amino acids absent from its training data when their building blocks are represented, breaks each prediction into contributions from individual residues, and can search peptide design spaces of more than 10^50 possibilities without testing every sequence. The paper presents a design tool, not a new drug.
Researchers at Fujian Medical University reported in Bioconjugate Chemistry on October 8, 2026 that they used the PepMimic AI platform to design TR23, a peptide designed by mimicking a protein binding interface, and labeled it with gallium-68 for PET imaging. The tracer bound TROP-2, a protein common in many solid tumors and the target of the antibody-drug conjugate Trodelvy, with nanomolar strength, and gave clear, specific tumor signals in pancreatic, prostate, and thyroid cancer models in mice, with signal strength tracking TROP-2 levels in tissue. The work is preclinical.
Researchers at Zhejiang University of Technology in China and Egypt's National Research Centre reported in the Journal of Medicinal Chemistry on October 7, 2026 a deep-learning model, SMAMP, that reads peptides as full chemical structures, so it can handle unnatural amino acids and end-group changes that sequence-based models miss. Of 19 very short peptides designed with its help, 13 killed bacteria, and the model's predicted potency against E. coli tracked lab measurements (correlation 0.764). Two leads, HP-1 and HP-2, were strongly bactericidal with low toxicity to mammalian cells, led to less resistance than standard antibiotics in lab tests, protected wax moth larvae from infection, and helped wounds heal in mice.
Researchers at Lembas Bio and Tel Aviv University reported in ACS Nutrition Science on October 6, 2026 an AI-guided design pipeline, built on AlphaFold2 structural modeling, for peptides that act on nutrient-sensing receptors of the gut cells that release the hormone GLP-1. After four rounds of design and screening, lead peptides raised GLP-1 secretion more than 27-fold in a mouse gut cell line, exceeding an optimized small-molecule GLP-1 secretagogue, and 28 days of oral dosing reduced food intake and body weight in diet-induced obese rats, with semaglutide as an active comparator. The authors present the work as a route to nutritional ingredients, and the results so far come from cells and rats.
South Korea's D&D Pharmatech and LG AI Research announced June 17 a joint development agreement for AI-designed next-generation oral peptide drugs. LG AI Research will use its EXAONE Discovery pharmaceutical AI platform, which analyzes disease-causing protein structures to design optimal peptide sequences, while D&D contributes its oral peptide delivery experience and obesity-pipeline programs. D&D's shares closed up 16.59% on the announcement. The deal extends D&D's late-cycle pipeline momentum, following the April 2026 $1.3M Pfizer research contract on oral obesity peptides after Pfizer's Metsera acquisition, and the June ADA showcase of Korean monthly-GLP-1 programs.
A Medscape ADA wrap on June 8 framed the post-GLP-1 era taking shape across the meeting: amylin analogs (petrelintide, cagrilintide, eloralintide) targeting GI-intolerance, triagonists (retatrutide), antibody-peptide conjugates (Amgen's maridebart cafraglutide/MariTide), and preclinical acceleration through AI peptide-design platforms. Each angle aims at the population gap that current GLP-1 monotherapy leaves: the patients who quit for tolerability, the ones who plateau, and the ones who need additional metabolic effects beyond appetite.
On June 4, 2026, Alnylam Pharmaceuticals announced a partnership worth up to $2 billion with Inceptive, the AI biotech focused on programmable RNA and peptide medicines, to apply machine learning to peptide discovery. The deal is the second major back-loaded peptide-AI collaboration to land in the same week, following Regeneron's expanded Parabilis tie-up tied to Phase 3 LAG-3 setbacks.
A venom-peptide discovery system published in Pharmaceuticals (MDPI) in 2026 paired phage display with a machine-learning model that predicts mutation-tolerant residues, building a library from roughly 482 venom-derived scaffolds. Screened against CD47, DLL3, IL33, and P2X7R, the resulting VCX library yielded strong binders for all four targets. Venom peptides are stabilized by multiple disulfide bonds and naturally evolved to hit GPCRs and ion channels, giving them structural stability that conventional peptides often lack.
A Nature Microbiology paper (published May 22) reported a generative artificial-intelligence approach for discovering antimicrobial peptides against multidrug-resistant bacteria. The method uses transfer learning to give large language models domain-specific knowledge for high-throughput mining and generation of novel AMP candidates. The work joins the 2026 AI-AMP wave — ProteoGPT's 94.4% hit rate, the CAMPER mechanistic-AI MRSA platform, ancient-microbiome AMP mining, and the May 19 nano-AMP delivery review — that is collectively moving the antimicrobial peptide field from computational prediction toward clinical candidates. The convergence matters because antimicrobial resistance is projected to cause up to 10 million deaths annually by 2050, and the conventional small-molecule antibiotic pipeline has thinned to the point where membrane-targeting peptides with low resistance-development propensity are among the most credible near-term alternatives. The generative-AI design stack plus nanoparticle delivery addresses the two historical AMP bottlenecks — discovery throughput and the toxicity/stability/manufacturing gap — in parallel.
20n Bio announced a $7.5 million Seed+ round May 20 led by a strategic industry investor with participation from a London-based life-sciences VC. Proceeds advance the company's high-throughput cyclic peptide platform — which screens libraries of up to 10 trillion sequences — and integrate AI into the discovery workflow. 20n is a Bayer Co.Lab resident and has a January 2026 long-term partnership with Yantai Lannacheng Biotechnology to develop next-generation radionuclide drug conjugates, leveraging Lannacheng's clinical translation and manufacturing for precision oncology programs.
Fadi Shehadeh, Biswajit Mishra and collaborators published CAMPER (Constraint-driven AMP Engineering with Ranking) in Nature Communications 2026, integrating machine learning with mechanistic biological features to design peptides that target MRSA persister cells. The lead candidate, WP-CAMPER1, kills S. aureus MW2 at a minimal inhibitory concentration of 4 µg/mL. A 2% topical formulation reduced bacterial burden 2.5 log10 in a murine prophylactic skin infection model; the D-enantiomer WP-CAMPER1-d achieved 1.37 log10 reduction in established biofilm infections.
Beijing Frontier Research Center for Biological Structure (FRCBS) at Tsinghua University organized an international Peptide Design Competition this year with about 300 participants from around the world designing peptide candidates targeting NK2R — a G-protein-coupled receptor involved in energy metabolism and appetite regulation. The competition is structured as a benchmarking exercise to test how well AI-driven structural predictions hold up under experimental scrutiny, with participants submitting designs that are then synthesized and tested for binding and functional activity. NK2R is a target of growing interest in the obesity-pharmacology field as the GLP-1 receptor space saturates and pharmaceutical R&D groups look for the next-generation metabolic-disease receptor that could complement GLP-1/GIP/glucagon agonism. The framework — AI design followed by wet-lab validation — is positioned as a community blueprint for de novo peptide discovery at large.
April-May 2026 produced an unusual concentration of peer-reviewed advances against antimicrobial resistance, with peptide therapeutics anchoring much of the progress. The Frontiers in Bioinformatics March 17 review documented an AI/LLM pipeline (ProteoGPT) that produced 17 active peptides out of 18 designed (94.4% hit rate) in 48 days — collapsing the traditional discovery timeline. A Nature Microbiology generative-AI approach produced novel AMPs against multidrug-resistant bacteria with anti-inflammatory effects and minimal cytotoxic risk. Other April-May 2026 breakthroughs: Houston Methodist's CAMPER engineered against MRSA, the Manchester team's alternative ligase pathway to penicillins, and the Indian Institute of Technology Roorkee antibacterial peptide-drug conjugate against NDM-1/IMP-1 metallo-beta-lactamase pathogens. The combined wave signals the peptide-design stack has matured enough to compete directly with small-molecule antibiotic development.
A Frontiers in Bioinformatics review published March 17, 2026 from Tope Abraham Ibisanmi and colleagues at UNSW Sydney documents how computational antimicrobial peptide discovery has collapsed from decades to weeks. The review covers big-data mining, molecular dynamics simulations, and AI methods that capture complex sequence-activity relationships and predict novel AMPs from genomic and metagenomic data. The headline example: one large language model approach produced 18 de novo peptides of which 17 were active (94.4% hit rate) over a 48-day discovery cycle. The framing complements the broader AMP-as-AMR-response thesis with Aifeity, the University of Bonn, and Cesar de la Fuente at Penn — and lands as Cesar de la Fuente's Penn lab launches new generative AMR molecules into ESKAPE-pathogen testing.
A Nature Communications paper (2026) introduces CycloPepper, a machine-learning-guided platform for predicting cyclization outcomes and accelerating automated synthesis of therapeutic cyclic peptides. The model addresses one of the persistent bottlenecks in cyclic-peptide drug development: many promising linear sequences fail at the macrocyclization step or yield poorly under standard conditions, requiring expensive iterative chemistry. CycloPepper trains on a curated dataset of cyclization outcomes and integrates with automated synthesis platforms to enable closed-loop design-make-test cycles. The work joins CyclicMPNN (a fine-tuned ProteinMPNN derivative for cyclic peptide sequence design) and AfCycDesign as part of a fast-maturing computational stack for cyclic-peptide therapeutics.