AI Reprograms Bacteriophage Host Range, Creating T7 Phages with Precisely Designed Bacterial Targets

Researchers at the University of Wisconsin–Madison have used machine learning to redesign one of the most important properties of a bacteriophage: deciding which bacteria it can infect. Rather than simply asking an algorithm to make phage T7 more infectious, the team trained models to simultaneously control infectivity and specificity, producing engineered phages that could attack selected Escherichia coli strains while avoiding others. The work, published in Cell Systems, provides a striking demonstration of how artificial intelligence could eventually turn naturally occurring phages into more precisely programmable antibacterial agents.

One promising approach to targeting unwanted bacteria is phage therapy, which uses naturally occurring and engineered viruses known as phages (illustrated above, resting on the surface of a host cell) to infect and destroy specific bacteria. Courtesy of the Raman Lab

The study, Multiobjective learning and design of bacteriophage specificity, was led by Naia Novy, Phil Huss, Sarah Evert, Philip A. Romero and Srivatsan Raman. Rather than relying primarily on generative AI trained on general protein sequences, the researchers built their system around a large experimental dataset generated from phage T7 itself. The goal was to learn how mutations in its receptor-binding protein alter the ability of the virus to recognize and infect different bacterial hosts.

This receptor-binding step is one of the central determinants of bacteriophage host range. Before a phage can inject its genome and reproduce, proteins at the end of the viral particle must recognize compatible molecules on the bacterial surface. Small changes in these receptor-binding proteins can therefore determine whether a phage efficiently infects one strain, several strains or none at all. For therapeutic applications, this creates both an advantage and a problem: phages can be extraordinarily selective, but a phage that works against one patient’s bacterial isolate may fail against a closely related strain.

The Wisconsin team focused on the terminal 82 amino acids of the T7 receptor-binding protein. Previous high-throughput experiments had generated thousands of variants in this region and measured their behavior against five E. coli strains: 10G, BL21, BW25113, BW25113ΔrfaD and BW25113ΔrfaG. These bacteria differ in their surface lipopolysaccharide structures and in their natural susceptibility to T7. Wild-type T7 readily infects several of them, while its efficiency of plating on the ΔrfaD and ΔrfaG strains is approximately four orders of magnitude lower.

Altogether, the machine-learning dataset contained 26,838 T7 receptor-binding protein variants and approximately 125,000 experimental measurements of their performance across the different bacterial hosts. Individual variants contained between one and ten mutations. Rather than teaching a model a simple rule such as “higher infection is better,” the researchers used this multidimensional dataset to teach models how protein sequence relates to host-specific virulence across several bacterial backgrounds simultaneously.

That distinction is crucial. A therapeutic phage may need to be highly active against one pathogen while remaining inactive against closely related bacteria that should be preserved. In another situation, clinicians might want the opposite: a broad phage capable of infecting many variants of the same pathogen. Optimizing only for maximum infectivity cannot solve both problems. The Wisconsin approach instead treats phage engineering as a multiobjective design problem.

The researchers compared four computational strategies. These included convolutional neural networks, models based on the ESM protein language model, and versions of both approaches combined with structural constraints from ProteinMPNN. The models were then used to search protein sequence space for receptor-binding proteins predicted to produce desired combinations of bacterial targeting.

The resulting experiment was unusually large for prospective phage engineering. The team designed, synthesized and tested 7,845 mutant T7 phages. Individual candidates carried three, six, nine or twelve substitutions relative to the wild-type receptor-binding protein, allowing the researchers to test whether the models could continue designing functional viruses even as their sequences moved progressively farther from the natural protein.

Instead of testing one desired phenotype, the researchers defined 26 separate design objectives. Five asked the models to maximize infection of an individual E. coli strain. Twenty required different patterns of specificity in which infection had to be increased on designated target bacteria while being suppressed on designated non-target strains. The final objective was to create a generalist phage capable of infecting all five bacterial strains used during model training.

The system succeeded across all 26 objectives. For the 20 specificity problems, engineered phages satisfying every requested targeting pattern were experimentally identified. Even more strikingly, 19 of the 20 specificity profiles could be obtained with variants carrying only three mutations relative to wild-type T7. Several engineered phages displayed a host preference essentially opposite to that of the natural virus.

This exposes an important property of T7 biology: host specificity is remarkably plastic. A bacteriophage does not necessarily need dozens of genomic changes to acquire a substantially different host preference. In the receptor-binding region studied here, a small number of amino-acid substitutions could move T7 between dramatically different targeting profiles. The authors argue that this reveals an unexpectedly accessible functional landscape in which very different host-range phenotypes may be separated by only a few mutations.

The researchers also examined the simpler problem of maximizing infection of individual bacterial strains. Among phages carrying three AI-selected mutations, approximately 80–95% of designs remained capable of infecting their intended target, depending on the model and host. Success declined as the number of mutations increased, as might be expected when increasingly large changes are introduced into an essential viral protein, but functional designs persisted even at twelve substitutions.

For broader host range, the models generated generalist phages capable of infecting all five training strains. Approximately 80% of three-mutation designs tested for this objective could infect the complete five-strain panel. Even at nine substitutions, some modeling approaches produced generalists at success rates exceeding 50%.

The team then moved beyond pooled high-throughput measurements and reconstructed selected phages individually for conventional plaque-based validation. One three-mutation variant designed to infect the ΔrfaG strain showed an approximately 17,000-fold improvement in efficiency of plating compared with wild-type T7. A nine-mutation generalist maintained approximately wild-type activity on the naturally susceptible strains while increasing activity against the resistant ΔrfaD strain by around 5,000-fold and against ΔrfaG by roughly 110-fold.

Specificity could also be altered dramatically. One six-mutation phage showed an approximately 17,000-fold improvement in the researchers’ specificity metric compared with wild-type T7. Importantly, this did not simply mean increasing infection everywhere. The algorithm was able to trade some activity against one target bacterium for much stronger suppression against a bacterium designated as an anti-target. That ability to optimize competing objectives is precisely what distinguishes the study from approaches designed merely to maximize viral replication.

The University of Wisconsin–Madison highlighted another striking result from the broader experimental screen: some engineered variants displayed improvements in host infection reaching as high as six orders of magnitude compared with their naturally occurring counterpart. Because the study used several different measures of relative fitness and subsequent clonal plaque assays, these large high-throughput effects should not all be interpreted as direct million-fold increases in an absolute clinical killing rate. Nevertheless, the experiments show that relatively small sequence changes can profoundly alter whether T7 successfully infects a bacterial host.

The work also challenges an increasingly common assumption about biological AI: that the largest pretrained protein language model will necessarily provide the best engineering results. When substantial experimental training data were available, relatively conventional convolutional neural networks performed competitively with fine-tuned protein language models. Structural constraints could improve performance for some of the more difficult specificity designs, while ESM-based approaches performed particularly well on several multi-strain objectives.

This matters because the real strength of the approach comes from combining machine learning with large-scale experimental biology. The models were not asked to infer phage behavior from sequence alone. Years of work in the Raman laboratory had already mapped how thousands of receptor-binding protein mutations influence infection. Artificial intelligence provided a way to extract rules from that experimental landscape and then search combinations that would be difficult to identify manually.

The researchers are now trying to move beyond laboratory E. coli strains. Raman told UW–Madison that the group has begun applying related strategies to bacteria responsible for urinary tract infections, organisms associated with microbiome-linked diseases and agricultural pathogens such as Salmonella affecting poultry. The long-term objective is not merely to improve T7, but to establish a general strategy for designing bacteriophages against clinically and economically important bacterial pathogens.

One potential application is personalized phage therapy. Today, matching a patient’s bacterial isolate with an effective phage often requires screening a collection of naturally occurring viruses. That strategy remains highly valuable, but it is constrained by whatever host ranges evolution happened to produce. A validated engineering platform could create another option: start with a well-characterized phage and deliberately alter its receptor-binding machinery until it recognizes the bacterial strain that needs to be treated.

Another possibility is the deliberate construction of broader-host-range therapeutic phages. Resistance can emerge rapidly when bacteria alter or lose the surface receptor used by a virus. A phage engineered to maintain activity across several receptor configurations could theoretically reduce this vulnerability, although this would need to be demonstrated in clinically relevant infection models rather than assumed from laboratory host-range measurements.

The opposite strategy may be equally important. Conventional broad-spectrum antibiotics can eliminate susceptible beneficial organisms together with the pathogen, disrupting microbial communities. The Wisconsin models could explicitly optimize activity against selected bacteria while suppressing infection of others. In principle, such multiobjective design could eventually help produce antimicrobials that eliminate a pathogen while leaving related members of the microbiome relatively untouched. The current study demonstrates this concept across laboratory E. coli strains, not yet within the complexity of a human microbiome.

There are therefore important limits to what has been demonstrated. This is not yet an AI-generated treatment for antibiotic-resistant infection. The experiments centered on bacteriophage T7 and a defined receptor-binding region tested largely across engineered or laboratory E. coli strains. Host range is only one determinant of therapeutic performance. A phage designed to attach efficiently to a pathogen must still replicate effectively under physiological conditions, survive formulation and delivery, overcome bacterial antiviral defenses, interact acceptably with the immune system and demonstrate safety and efficacy in relevant animal models and ultimately clinical trials.

The approach also depends on high-quality experimental training data. The remarkable predictive performance in this study was possible because the researchers had already generated tens of thousands of sequence–function measurements. Comparable datasets do not yet exist for most therapeutically interesting phages and bacterial pathogens. Extending the system will therefore require substantial experimental work alongside advances in machine learning.

Yet the study represents an important conceptual shift. Natural phage discovery traditionally asks a simple question: which virus in nature already infects this bacterium? Directed evolution asks how selection can push an existing virus toward a desired phenotype. Multiobjective machine learning introduces a third possibility: specify the bacterial targeting behavior wanted in advance and computationally search for the mutations predicted to produce it.

The distinction could become increasingly important as phage therapy moves toward more standardized therapeutics. A future engineered phage might not simply be described as “active against E. coli.” Its design specification could instead define exactly which strains it should attack, which bacterial variants it should tolerate, which organisms it should avoid and how broadly it should maintain activity as bacterial receptors change.

That remains a long-term objective. But the Wisconsin team has now demonstrated an essential first step: within the T7 receptor-binding protein, bacterial host preference is sufficiently predictable and sufficiently flexible that machine-learning models can deliberately move it in multiple directions, including toward combinations of targets that natural T7 does not efficiently infect.

For phage therapy, the implications extend beyond making viruses more lethal. The more significant possibility is making them more controllable.

Sources :

University of Wisconsin–Madison — Renata Solan, Enlisting AI in the fight against drug-resistant bacteria, September 24, 2026.
UW–Madison article

Novy N, Huss P, Evert S, Romero PA, Raman S. Multiobjective learning and design of bacteriophage specificity. Cell Systems. 2026;17(9):101712. DOI: 10.1016/j.cels.2026.101712.
Cell Systems paper