The scientist using AI to hunt for antibiotics just about everywhere

AI predicts a post-antibiotic era

Bioengineer César de la Fuente and synthetic biologist James Collins warned in a July 2025 essay that the world is entering a "post-antibiotic" era where common infections become fatal. Writing in Physical Review Letters, the researchers noted that the discovery pipeline for new drugs remains thin due to high costs and low returns on investment.

Drug-resistant bacteria, fungi, and viruses currently contribute to more than 4 million deaths every year. A recent analysis in The Lancet predicts this figure will surge past 8 million by 2050. Common strains of Escherichia coli and Staphylococcus aureus are increasingly evading the current arsenal of medications.

De la Fuente is now using artificial intelligence to reverse this trend at the University of Pennsylvania. His team trains AI tools to search biological genomes for peptides with antibiotic properties. These molecules, which consist of up to 50 amino acids, could defend the body against microbes that withstand traditional treatments.

The high cost of drug failure

Antimicrobial resistance is an expensive problem that many pharmaceutical companies have abandoned. De la Fuente explains that conventional ways to find and test drugs are prohibitively expensive and often lead to dead ends. Many firms have folded because there is no significant return on investment for new antibiotics.

Antibiotic discovery has historically relied on brute-force mechanical methods. Scientists typically dig into soil or water to extract antimicrobial molecules from organic matter. This process is slow, noisy, and relies heavily on luck to find a viable candidate.

The scale of the search space makes traditional methods nearly impossible. Researchers estimate there are roughly 10^60 possible organic combinations that could form a drug. For context, the entire Earth contains only about 10^18 grains of sand.

Mining the code of life

De la Fuente views biology as a massive source of information similar to computer code. While DNA uses four letters, proteins and peptides use 20 letters, with each letter representing a specific amino acid. AI models can be trained to recognize the specific sequences that encode antimicrobial peptides, or AMPs.

AMPs are a critical part of the human immune system and serve as a first line of defense against infection. Unlike conventional antibiotics that usually have one trick for killing bacteria, AMPs often use a multimodal approach. They can simultaneously disrupt cell walls and damage the genetic material inside a pathogen.

This multipronged attack makes it harder for bacteria to evolve resistance. A pathogen might develop a defense against a single mode of action, but it struggles to survive a simultaneous assault on its entire cellular structure. De la Fuente believes these molecules can be engineered into configurations never seen in nature.

Searching extinct genomes for medicine

The Penn Machine Biology Group is searching for these molecules in unexpected places, including the genetic code of extinct species. This "molecular de-extinction" project scans published sequences of organisms that have been gone for millennia. The team believes some ancient organisms may have evolved defenses that are useful against modern threats.

The project has already identified several resurrected compounds from the history of life on Earth. These include:

• Mammuthusin-2: A peptide derived from woolly mammoth DNA.
• Mylodonin-2: A compound recovered from the genetic code of the giant sloth.
• Hydrodamin-1: An antimicrobial molecule found in the ancient sea cow.
• Neanderthal peptides: Sequences excavated from the genomes of extinct hominids.

In August 2025, the team also described peptides hidden in the genetic code of archaea, which are ancient single-celled organisms. They have also excavated potential drug candidates from the venom of snakes, wasps, and spiders. This digital binge has allowed de la Fuente to amass a library of more than 1 million genetic recipes.

Moving from prediction to generation

The field of AI drug discovery is shifting from simply screening existing libraries to designing new molecules from scratch. James Zou, a computer scientist at Stanford, notes that researchers are moving from predictive models to generative approaches. This allows scientists to create synthetic peptides that have never existed in the natural world.

Last year, de la Fuente’s team used a generative AI model to design a suite of synthetic peptides. They tested two of these compounds on mice infected with a drug-resistant strain of Acinetobacter baumannii. The World Health Organization classifies this germ as a "critical priority" for research.

The synthetic peptides successfully and safely treated the infection in the animal models. This success follows the 2020 discovery of halicin, a broad-spectrum antibiotic identified by James Collins and Jonathan Stokes. Halicin is currently in preclinical development and represents one of the first major successes of AI-led discovery.

The future of automated discovery

De la Fuente is now developing a multimodal model called ApexOracle to move candidates closer to clinical testing. This system is designed to analyze a new pathogen and pinpoint its specific genetic weaknesses. It then matches the pathogen to antimicrobial peptides that are most likely to work against it.

ApexOracle also predicts how an antibiotic built from those peptides would perform in laboratory tests. The system converges chemistry, genomics, and natural language processing into a single tool. De la Fuente says the goal is to resist resistance by staying ahead of bacterial evolution.

Using AI has already saved decades of human research time by automating the screening of billions of molecules. While these peptides have not yet been transformed into usable human drugs, the speed of discovery is increasing. De la Fuente believes this technology gives researchers a fighting chance to catch up to the growing threat of superbugs.

The current challenge remains the transition from the lab to human patients. Details regarding dosage, delivery methods, and specific cellular targets still require extensive sorting. However, the ability to mine the history of life for new medicines has fundamentally changed the economics of the antibiotic pipeline.