Dual-objective peptide generator pairing an MMD-VAE sampler with antimicrobial and cell-penetration predictors to design multimodal antibiotics.
No providers recorded yet. Browse all providers
Cell-penetrating peptides and antimicrobial peptides have been studied as two separate classes for decades: one gets cargo across a mammalian membrane, the other kills bacteria. A pathogen such as Staphylococcus aureus that replicates inside a keratinocyte needs both at once, and conventional antibiotics that cannot reach the cytoplasm at a tolerable dose are why intracellular infections persist and recur. ApexDuo treats dual function as a filtering problem over a generative model: sample enormously from a peptide sequence generator, then keep only the sequences that two independently trained predictors — one for antimicrobial potency, one for cell penetration — both score highly.
The model comes from the Machine Biology Group at the University of Pennsylvania (Angela Cesaro, Fangping Wan, Marcelo D. T. Torres and Cesar de la Fuente-Nunez), posted to bioRxiv in December 2024 and revised in October 2025. It reuses the lab's APEX multi-strain minimum inhibitory concentration (MIC) regressor as its antimicrobial oracle, as ApexGO does, and sits alongside the group's other peptide designers such as ApexAmphion. What distinguishes ApexDuo is the second objective: nothing is selected unless it is also predicted to enter a human cell.
The generator is a maximum mean discrepancy variational autoencoder (MMD-VAE). Peptides are one-hot encoded, embedded, and passed through a gated-recurrent-unit encoder with two stacked attention layers that pool residue features into a latent vector; the decoder is a matching recurrent network with a per-position output layer. It was trained for 1,000 epochs at a learning rate of 0.0001 on UniRef50 sequences of at most 50 residues — 1,418,011 training and 354,503 validation peptides. Generation draws 50 million latent codes from a kernel density estimator fitted to the codes of in-house antimicrobial peptides and public cell-penetrating peptides.
The antimicrobial predictor is APEX, updated here with 14,963 MIC measurements over 1,016 in-house peptides and 37 strains alongside the 14,738 DBAASP values used originally. The cell-penetration predictor shares the MMD-VAE encoder architecture plus a classification head, and runs as a 25-model ensemble trained on 1,217 cell-penetrating peptides from CPPsite 2.0, SkipCPP-Pred and MLCPP against 5,293 non-penetrating or toxic peptides, the latter drawn from SATPdb so the classifier discounts toxicity. Selection kept designs with median predicted MIC at or below 32 µmol L⁻¹, cell-penetration probability at least three standard deviations above the mean of the 50 million, no more than 30 residues, and pairwise sequence similarity below 0.5. Fifty-two soluble peptides, named Turingcins, were synthesized: 47 (90.4%) were active against at least one of twelve strains with MICs of 0.78–50 µmol L⁻¹, and most had cytotoxic or hemolytic thresholds several times their MIC. Turingcin-46 reduced intracellular S. aureus and Listeria monocytogenes in keratinocytes and macrophages, and cut bacterial loads by more than two orders of magnitude in a mouse skin-abscess model and by 1.2–1.8 orders across organs in a peritonitis model.
ApexDuo is built for antibiotic discovery against intracellular pathogens, where the bottleneck is not finding a bactericidal sequence but finding one that reaches the cytoplasm without killing the host cell. The pretrained generator and both predictors are published, so a group with peptide synthesis and microbiology capacity can sample around its own prior sequences, score them for both properties, and shortlist a panel to make. The same two-filter pattern generalizes to other property pairs; the authors point to immunomodulation, endosomal escape and subcellular targeting as candidates.
The result that matters here is not the generator, which is a conventional VAE, but the demonstration that intersecting two independent property predictors over a large enough sample yields molecules with genuinely dual function, validated in animals rather than only in silico. The authors name the main constraint themselves: no training data annotated specifically for intracellular activity, and limited public cell-penetrating peptide data, so the predictors approximate the property actually wanted. Only peptides with confirmed antimicrobial activity entered the intracellular assays, leaving the purely cell-penetrating output untested. The work is a preprint awaiting peer review, and while pretrained weights for all three components are on GitLab, the repository is covered by a University of Pennsylvania non-commercial permission notice restricting use to non-profit research. Generative antimicrobial design also carries dual-use considerations that warrant scrutiny.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.