NEWS.

A protein language model ranked 689 cone-snail peptides, and one of them, named SS1, blocked the alpha-7 nicotinic receptor. Optimized, it works at nanomolar doses, and cryo-EM caught it sitting inside the receptor at 3.3 angstroms. The AI pick and the microscope agreed.

An audited language-model pipeline read clinical notes from 29 million patients and put 18-month GLP-1 persistence at 42 to 43 percent, about 13 points below what prescription records implied.

ARCADIAMP, a generative AI that writes peptide sequences the way image models write pixels, produced ten antimicrobial candidates. Eight killed bacteria, and the lead one, Arcinin, cleared drug-resistant ESKAPE pathogens while sparing human cells and healing infected wounds in mice.

Wistar Institute researchers injected obese mice with a plasmid encoding cleavage-resistant GLP-1 and GIP analogues, then used AI-guided design to build a dual-agonist version. One dose drove durable weight loss.

A deep-learning model screened 19.3 million fragments cut from prion proteins and flagged 1,179 antibiotic candidates. Of 75 the Penn team synthesized, 59 killed bacteria and two cleared infections in mice.

CIP-3, an AI-designed cyclic peptide reported May 30 in Advanced Science, binds CD28 at nanomolar affinity with no intrinsic agonist activity, suppresses T-cell activation in primary human cells, and matches a benchmark anti-CD28 biologic on cytokines in ulcerative colitis PBMCs. The kinetic profile is the part biologics do not have: occupancy tracks exposure rather than persisting for days.

ApexGO couples a transformer variational autoencoder to a Bayesian optimization loop and proposes scaffold-local edits to peptide antimicrobials. Starting from 10 templates the team chemically synthesized 100 derivatives; 85 percent cleared the in-vitro hit bar, 72 percent improved on their parent against Gram-negative pathogens, and two molecules cleared Acinetobacter baumannii infection in mice at potencies matching or exceeding a last-resort antibiotic.

A new classifier called MARC reads cysteine-rich venom peptide sequences and assigns each to a sodium, potassium, calcium, or non-ion-channel bucket. Trained on 5,165 sequences from sea anemones, snakes, scorpions, spiders, cone snails, and terebrid sea snails, it flagged 28 new terebrid teretoxins as potassium-channel candidates and validated one (Cje1.9) by docking and molecular dynamics against KcsA and MthK.

A BMC Microbiology paper mined 9 Arctic Mid-Ocean Ridge hydrothermal-vent biofilms via metagenomics, metatranscriptomics, and machine-learning prediction, identifying 873 unique antimicrobial peptide sequences with no match in existing databases. 25 of 51 contributing phyla showed active expression. 16.7% predicted active against clinical pathogens, with Acinetobacter baumannii most susceptible. Of four synthesized leads, OLKFNNDA_52_10 showed moderate S. aureus activity with low human-cell cytotoxicity.

HighFold-MeD2, an enhancement of the open-source Boltz-2 structure-prediction model, handles cyclic peptides containing backbone N-methylated and d-amino acids, the modifications that drive cyclic-peptide drug-class metabolic stability and oral bioavailability. The model uses Chemical Component Dictionary representations rather than hard-coded non-canonical amino acid entries, making it extensible without retraining. It outperforms HighFold-MeD, AlphaFold3, and Boltz-2 baseline on cyclic-peptide structure prediction.

RoBERTcr is a 152-million-parameter language model trained specifically on T-cell receptor sequences. A new Bioinformatics paper reports it beats Meta's ESM and other general protein language models on TCR-peptide binding prediction, and learns the binding interface from sequence alone.

POTFlow, a flow-matching generative model conditioned on a known lead peptide, designed candidates targeting ATP5A in glioblastoma. The peptides selectively cut tumor cell viability and prolonged survival in patient-derived xenograft mice, closing the design-to-validation loop on a hard cancer target.

Two new research papers tested whether AI can design drug molecules. General-purpose chatbot-style AIs collapse on specific targets. A specialized tool built for one narrow peptide job found 76 working candidates where the broad approach found zero.

Microsoft researchers pulled every lab-tested AI drug design result into one scorecard. Binders work. Antibodies are inconsistent. Enzymes are harder. One target keeps beating the AI.