How Artificial Intelligence and Robotics Are Transforming the Discovery of Therapeutic Phage Cocktails

For decades, one of the greatest limitations of phage therapy has not been finding bacteriophages, but finding the right bacteriophages. Nature has produced an almost unimaginable diversity of phages, with an estimated 10³¹ viral particles present on Earth. Yet identifying which of these viruses can effectively eliminate a specific bacterial pathogen remains a complex and time-consuming process. As antimicrobial resistance continues to spread worldwide, researchers are increasingly searching for ways to accelerate phage discovery and transform what has traditionally been an artisanal process into a scalable and predictive science.

©The Phage Therapy

A new study published in Nature Communications demonstrates how robotics, computer vision and artificial intelligence may help overcome this challenge. Researchers from Locus Biosciences, Lawrence Berkeley National Laboratory and several collaborating institutions have developed one of the most advanced automated platforms yet reported for therapeutic phage development. Their work resulted in the creation of a highly effective phage cocktail targeting multidrug-resistant uropathogenic Escherichia coli while simultaneously establishing a technological framework that could reshape the future of precision phage therapy.

The project reflects a broader shift occurring across the life sciences. Aeron Tynes Hammack, one of the scientists involved in the development of the platform, originally worked in fields far removed from microbiology. With a background spanning chemistry, nanotechnology and high-throughput experimental systems, he became interested in applying the principles of large-scale automation to phage discovery. The central idea was simple but powerful: if phages are essentially biological nanoparticles, why not study them using the same industrial-scale technologies that transformed materials science and genomics?

The challenge becomes apparent when considering the mathematics of phage selection. Unlike conventional antibiotics, phages display remarkable specificity. A virus capable of killing one strain of Escherichia coli may be completely ineffective against another strain carrying only minor genetic differences. To maximize efficacy and reduce resistance emergence, therapeutic formulations often combine several phages into cocktails. However, the number of possible combinations rapidly becomes overwhelming. Selecting just six phages from a collection of 500 candidates generates more than 21 trillion potential cocktails. Testing each combination experimentally would be impossible using traditional laboratory workflows.

To address this problem, the researchers built a fully automated discovery pipeline integrating robotics, cloud computing and machine learning. The effort began with an extensive environmental sampling campaign. More than 1,084 wastewater and environmental samples collected across 48 U.S. states yielded 1,143 bacteriophages capable of infecting E. coli. Following genome sequencing and safety assessment, 421 phages were selected for detailed characterization.

The scale of experimentation achieved by the platform is unprecedented in phage research. Robotic systems continuously prepared bacterial cultures, distributed phages into hundreds of thousands of microplate wells and monitored bacterial growth in real time. More than 3.8 million phage-bacteria interactions were evaluated during the project. Automated imaging systems analyzed over 2.2 million bacterial colonies and phage plaques, while more than 1.5 million bacterial growth curves were generated and processed computationally.

At the core of the platform lies a simple physical principle. Growing bacterial populations scatter light, increasing culture turbidity. Successful phage infection reduces bacterial density through cell lysis, causing measurable decreases in optical density. By continuously recording these optical signals, researchers could quantify bacterial killing with extraordinary precision. Every experiment generated a digital signature describing the interaction between a particular phage and a specific bacterial strain.

Artificial intelligence transformed these measurements into predictive models. Instead of blindly screening billions of possible combinations, machine-learning algorithms continuously analyzed previous results to identify promising candidates. Each new experiment refined the model, allowing the system to focus resources on the most effective phages and cocktail designs. This iterative feedback loop created a self-improving discovery platform capable of navigating biological complexity far more efficiently than conventional approaches.

The researchers focused their efforts on urinary tract infections caused by uropathogenic Escherichia coli, one of the world's most common bacterial infections. More than 400 million urinary tract infections occur annually, and resistance to frontline antibiotics continues to rise globally. To ensure clinical relevance, the team assembled a validation panel consisting of 356 clinical isolates collected from female patients across 39 U.S. states. Nearly one-third of these isolates were classified as multidrug-resistant.

The final product of the optimization process was a six-phage cocktail designated LBP-EC01. Laboratory testing revealed remarkable performance. The cocktail demonstrated activity against 96.4 percent of all clinical isolates tested and achieved reductions exceeding 99.999 percent of bacterial populations in 94 percent of strains within only four hours. Such reductions correspond to a decrease of more than five logarithmic units, meaning that a population of one billion bacterial cells can be reduced to fewer than ten thousand surviving organisms.

Importantly, the platform was not merely optimized for laboratory strains. Researchers also evaluated the cocktail against bacterial isolates obtained from participants enrolled in the ongoing Phase 2 ELIMINATE clinical trial. Among 38 genetically distinct E. coli isolates recovered from trial participants, 97.4 percent remained susceptible to the phage cocktail. No significant emergence of resistance was observed during the monitored treatment period, supporting the robustness of the cocktail design strategy.

Beyond the immediate success of LBP-EC01, the broader significance of the study lies in its methodology. Historically, phage therapy has often relied on empirical approaches, where candidate phages are tested sequentially until an effective match is identified. This process can be labor-intensive and difficult to standardize. By contrast, the new platform treats phage discovery as a large-scale optimization problem driven by quantitative data, automation and predictive algorithms.

The implications extend well beyond urinary tract infections. Similar workflows could be adapted for pathogens such as Pseudomonas aeruginosa, Klebsiella pneumoniae, Acinetobacter baumannii and other members of the ESKAPE group that increasingly threaten modern healthcare. As phage libraries continue to expand and machine-learning models become more sophisticated, future systems may eventually identify optimal therapeutic phages in hours rather than weeks.

Perhaps most importantly, this work offers a glimpse into the future of precision antimicrobial medicine. Researchers increasingly envision a clinical workflow in which rapid bacterial diagnostics are combined with AI-driven phage selection and automated manufacturing platforms capable of generating personalized treatments on demand. While significant regulatory and logistical challenges remain, the integration of robotics, artificial intelligence and synthetic biology is rapidly moving phage therapy toward a new era.

What once depended on painstaking manual screening may soon become a highly automated process guided by millions of experimental measurements and predictive computational models. In the fight against antimicrobial resistance, this convergence of biology and machine intelligence could prove as important as the phages themselves.

Source : Penke T.J.R., Tynes Hammack A., McMillan L.J. et al. High-throughput methods leveraging robotics and computer vision for the development of therapeutic phage cocktails. Nature Communications (2026), https://doi.org/10.1038/s41467-026-68684-x