rethinkPeptides Search
Menu
Study breakdown

AI Mines 40 Million Venom Peptides and Finds Potent New Antibiotics — 91% Hit Rate in Lab Testing

evidence
The takeaway

Deep learning analysis of 16,123 venom proteins generated over 40 million encrypted peptide candidates, with 53 of 58 tested showing potent antimicrobial activity and lead compounds reducing bacterial infections in mice.

91% hit rate (53/58)

Of 58 AI-predicted venom-encrypted peptides tested experimentally, 53 showed potent antimicrobial activity — an extraordinarily high success rate for drug discovery screening.

What the researchers found

From 16,123 venom proteins, the APEX deep learning model generated 40,626,260 venom-encrypted peptides (VEPs) and identified 386 candidates structurally and functionally distinct from known antimicrobial peptides. These VEPs had high net charge and elevated hydrophobicity — properties that enable bacterial membrane disruption.

Of 58 VEPs selected for experimental validation, 53 (91%) displayed potent antimicrobial activity. Structural studies showed the peptides adopt α-helical conformations in membrane-mimicking environments. Mechanistic assays confirmed they work by depolarizing bacterial membranes. In vivo, lead VEPs significantly reduced A. baumannii bacterial burdens in a mouse infection model without notable toxicity.

Why it matters

This study demonstrates a paradigm shift in antibiotic discovery: instead of the traditional approach of screening individual compounds one at a time, AI can rapidly search millions of peptide candidates from nature's existing chemical diversity. The 91% validation hit rate is extraordinary for drug discovery (typical hit rates are under 1%). Venom peptides have been shaped by millions of years of evolution to be potent and selective, making them an ideal starting library for AI-guided antibiotic development.

How the study worked

Comprehensive global venomics datasets were mined using machine learning. APEX, a deep learning model combining peptide-sequence encoding with neural networks, predicted antimicrobial activity from over 40 million computationally generated venom peptide fragments. Structural analysis included conformational studies in membrane-mimicking environments. Experimental validation tested 58 top candidates for antimicrobial activity, membrane depolarization, and in vivo efficacy in an A. baumannii mouse infection model.

What this study cannot tell us

This is a preprint (bioRxiv), not yet peer-reviewed. The in vivo testing used a single bacterial species (A. baumannii) in an acute infection model; broader pathogen coverage and chronic infection models would strengthen the findings. The pharmacokinetics, stability, and manufacturing scalability of the lead VEPs are not discussed. The transition from mouse efficacy to human clinical trials involves substantial additional development. Cost of peptide antibiotic production compared to small-molecule antibiotics remains a practical challenge.

How to read the evidence

This is a preprint combining computational prediction, in vitro validation (91% hit rate), and in vivo mouse model efficacy. The evidence is strong for the discovery methodology but has not yet been peer-reviewed. The in vivo data is limited to a single pathogen model.

When this study was published

Posted as a preprint in 2024, this represents cutting-edge AI-driven drug discovery. The de la Fuente-Nunez lab has been a pioneer in machine learning-guided antimicrobial peptide discovery.

The bigger picture

This study from Cesar de la Fuente-Nunez's lab at UPenn represents the convergence of two major scientific trends: AI-driven drug discovery and venom-derived therapeutics. The approach is scalable — as more venom genomes are sequenced, the peptide library grows. The success rate suggests that AI can identify antimicrobial peptides far more efficiently than traditional screening. If these venom-encrypted peptides can be developed into clinical antibiotics, they could help address one of the world's most pressing health crises.

Questions still open

  • Can these venom-encrypted peptides be optimized for oral bioavailability, or will they require injection like most peptide drugs?
  • How do bacteria develop resistance to these membrane-disrupting VEPs, and how quickly does resistance emerge compared to conventional antibiotics?
  • Could this AI venomics approach be extended beyond antibiotics to discover venom peptides for pain, cancer, or other therapeutic areas?

Common questions

How did AI find antibiotics hidden in venom?
The researchers fed a deep learning model (APEX) the sequences of over 16,000 venom proteins from animals worldwide. The AI broke these proteins into over 40 million small peptide fragments and predicted which ones could kill bacteria based on patterns learned from known antimicrobial peptides. It identified 386 candidates that were structurally unlike any existing antibiotics, and 91% of those tested actually worked.
Why look for antibiotics in animal venom?
Venom has been refined by hundreds of millions of years of evolution to be biologically potent. It contains thousands of peptides designed to rapidly incapacitate or kill other organisms. These same membrane-disrupting properties that make venom effective against prey can also be directed against bacteria. Unlike traditional antibiotics that target specific proteins (which bacteria can mutate around), membrane-disrupting peptides are harder for bacteria to resist.

Read the original research

Venomics AI: a computational exploration of global venoms for antibiotic discovery.

bioRxiv : the preprint server for biology

Citation

Guan, Changge; Torres, Marcelo D T; Li, Sufen; de la Fuente-Nunez, Cesar. (2024). Venomics AI: a computational exploration of global venoms for antibiotic discovery.. bioRxiv : the preprint server for biology. https://doi.org/10.1101/2024.12.17.628923