Week 1 HW: Final Project Idea

Generative AI for Designing Programmable Bacteriophages

Project summary

My central question is whether generative biological language models can design new receptor-binding proteins (RBPs) that reprogram the host specificity of an existing bacteriophage. A bacteriophage, or phage, is a virus that infects bacteria. And, as mentioned, the term receptor-binding protein (RBP) is not unique to phages; it covers any protein that mediates attachment to a target receptor, including attachment proteins on other viruses. In a phage, the RBP is the part that helps determine which bacterium it recognizes. We treat the phage as a prototype platform: keep its replication machinery mostly fixed, and use it as the living test system that reveals whether a generative model’s custom-designed RBP actually functions, in a real bacterium recognizing a real target, rather than only on paper.

We would use non-pathogenic bacteria and an obligately lytic phage approved by the host laboratory. It would be a controlled proof of concept. The central question is:

Can generative biological models design receptor-binding proteins that reprogram the host specificity of an existing bacteriophage?

As a demonstration, we will test whether the resulting engineered phages can selectively target one chosen bacterial strain within a mixed, non-pathogenic biofilm while leaving the other strains largely unchanged. The biofilm acts as the environment to test whether the AI-designed targeting function actually works. The long-term vision is a generalizable platform for designing programmable biological targeting systems, of which biofilm control is one application.

The project has three central components:

1. Generative AI — the design engine. Generate and rank novel RBP sequences using protein- and DNA-level foundation models.
2. Phage — the prototype test platform. Manufacture each AI-designed RBP into a working phage, so its function can be validated experimentally rather than only predicted computationally.
3. Mixed biofilm — the test bed. Determine whether a validated designed RBP remains selective in a more realistic microbial environment. The proposed workflow is a design-build-test-learn loop:

Generative AI design → automated construction → host-range screening → biofilm test → experimental feedback → improved designs

1. What I want to biomanufacture and why

The deliverable is a preparation of infectious phage particles carrying a Generative AI designed RBP, with a defined targeting behavior. The phage is a prototype platform: it lets us test a computational design against real bacteria instead of only predicting how it would behave. In a model community of several non-pathogenic bacterial strains, phage carrying a working design should infect the selected strain, reduce its abundance, and leave the others largely unaffected. The experiment can be carried out on biofilms, which are communities of microbes attached to a surface and held together by a protective matrix.

A generative model explores RBP sequences that would be impractical to propose one mutation at a time, while a living bacterial system manufactures and amplifies the resulting phage. Previous work has shown that changing or mutating phage tail fibers can alter host range, so the biological premise is experimentally grounded (Yehl et al., 2019; Ando et al., 2015). The new part we want to explore is a repeatable, AI-guided design-build-test-learn workflow for selecting the target, generating RBP variants, making phages, measuring their specificity, and using those measurements to guide the next round.

Success would require three results:

1. The engineered phage remains able to form infectious particles.
2. It acts on the intended strain more strongly than on the non-target strains.
3. It changes the composition of a simple mixed biofilm in the intended direction.

2. Living system

We would begin with a well-characterized, obligately lytic T7-like phage and non-pathogenic Escherichia coli laboratory strains, subject to the host laboratory’s biosafety rules. T7-like phages are useful starting points because their genetics and tail-fiber biology have been studied extensively. In T7, the gp17 tail fiber contacts the bacterial surface, and high-throughput mutation of its receptor-binding region has produced functional variants with changed host range (Huss et al., 2021).

The system would work at a high level as follows:

Inputs: known phage RBP sequences and structures, information about the bacterial target and its surface receptor when available, a starting phage genome, non-pathogenic bacterial host cells, growth medium, and buffer.

Design and biological process: generative models would propose RBP sequence variants and computational filters would reduce that large set to a practical experimental library. Selected DNA variants would then be incorporated into the phage platform. A compatible bacterial production host would grow, become infected, copy the phage genome, assemble new phage particles, and lyse. The released phages would be clarified, concentrated, and tested against a panel containing the target and non-target strains.

Outputs: a purified phage preparation, its concentration in plaque-forming units per millilitre (PFU/mL), a measured host-specificity profile, and a dataset connecting each designed RBP sequence to its experimental behavior.

Generative AI as the design engine

We would use two complementary kinds of biological language model:

1. A protein language model (PLM), such as ESM-3 or a comparable accessible model, for RBP-level design. This model works on amino acid sequences. It generates and scores candidate RBP proteins by learning patterns from many known proteins (Hayes et al., 2025). We keep the part of the RBP that anchors it to the phage particle fixed, and let the model vary the part that actually contacts the bacterial surface. The model also scores how likely a candidate sequence is to fold correctly, which helps filter out sequences that would misfold or fail to assemble. The output of this step is a ranked list of candidate RBP protein sequences.

2. A DNA language model, such as Evo 2, for nucleotide- and genome-context analysis. This model operates on raw DNA sequences at single-nucleotide resolution across a long stretch of the genome, i.e., not just the gene of interest but its surrounding context.

The genetic code is degenerate, meaning that most amino acids are encoded by more than one codon (e.g., leucine has six different three-letter DNA codons that all mean “leucine”). So once the PLM picks a target protein sequence, there isn’t one DNA sequence that encodes it; there are potentially millions of DNA sequences that would all translate to the exact same protein.

The model would be used to check the same change at the DNA level, inside the whole phage genome. This matters because a protein-level model only sees the isolated protein; it has no idea that the DNA encoding that protein might overlap a promoter, a ribosome-binding site, a gene-packaging signal, or another gene reading in a different frame, all things that exist in real phage genomes and that a DNA change could silently break even if the resulting protein sequence looks fine. Evo 2 scores whether a given DNA-level encoding of the desired protein change is itself plausible in that genomic context (Brixi et al., 2026).

The experimental results would close the loop. Each candidate would receive measurements for phage assembly, titre, killing of the target, and activity against non-target strains. Those labels could be used to re-rank the next model-generated library or to train a small task-specific predictor on top of model embeddings. Generative AI therefore reduces the sequence search space, while automation supplies the data needed to improve the next design round.

3. Automation

Prototype and screening

We would use an liquid-handling robot (e.g., Opentrons) for the repetitive plate-based parts of the AI-guided design-build-test-learn cycle. The robot would set up DNA assembly reactions, distribute bacterial cultures and phage candidates, prepare serial dilutions, and assemble 96-well host-range assays. Opentrons supports standard 96-well plates and single- or eight-channel pipettes (Opentrons documentation). A plate reader would record bacterial growth over time. Promising variants would then be confirmed with plaque assays and sequencing.

A simple first screen could place 80 AI-generated RBP variants in one 96-well plate and reserve 16 wells for parental-phage, no-phage, and media controls. The same candidate plate could be tested separately against each member of the bacterial panel. Automation matters here because specificity is comparative: every candidate must be measured at the same starting cell density, phage dose, timing, and liquid volume. Consistent, machine-readable results also make it possible to connect sequence with phenotype and feed the data back into our generation loop.

Manufacturing

The Opentrons would help find a useful design, but it would not manufacture the pilot quantity. Scale-up would use a stirred-tank bioreactor with automated control of temperature, agitation, aeration, pH, optical density, and infection time. After lysis, filtration and tangential-flow filtration could clarify and concentrate the product. In-process samples would be taken automatically, while infective titre would still be checked with plaque assays.

4. How much I would make, and how long it would take

4a. Amount

We define a meaningful first scale as enough phage for one treatment of 100 m² of biofilm-covered surface in a contained pilot study. Because an effective surface dose has not yet been established for this exact phage-biofilm pair, we use 10⁷ PFU/cm² as an explicit design assumption. This value is a starting point to test, not a claimed universal dose.

Convert the area:

100 m² × 10,000 cm²/m² = 1.0 × 10⁶ cm²

Calculate the infectious particles placed on the surface:

1.0 × 10⁶ cm² × 1.0 × 10⁷ PFU/cm² = 1.0 × 10¹³ PFU

We would manufacture ten times this amount to cover losses during storage, dilution, and uneven delivery:

1.0 × 10¹³ PFU × 10 = 1.0 × 10¹⁴ PFU

The target for the pilot is therefore 10¹⁴ usable PFU.

For production, we assume:

• harvested culture titre: 1.0 × 10¹⁰ PFU/mL;
• recovery of infective phage after clarification and concentration: 50%;
• usable yield: 5.0 × 10⁹ PFU/mL of culture.

The titre assumption is reasonable for an early estimate: pilot-scale production of T4-like coliphages has reported peak titres of 10⁹ to 10¹⁰ PFU/mL in flasks and bioreactors (Bourdin et al., 2014). The 50% recovery is a conservative assumption and would need to be measured for the chosen phage.

Required culture volume:

1.0 × 10¹⁴ PFU ÷ 5.0 × 10⁹ PFU/mL = 2.0 × 10⁴ mL = 20 L

We therefore plan for one 20 L working-volume run, or two 10 L runs operated in parallel. The recovered phage could then be diluted into whatever carrier volume the surface-application method requires.

4b. Time

For a design that has already passed small-scale tests, one manufacturing run would take approximately:

Two 10 L reactors running in parallel would therefore make the estimated pilot amount in about two days. With only one 10 L reactor, two sequential runs would take about four days. These times exclude initial RBP design and screening. A first design-build-test-learn round would more likely take one to two weeks, and several rounds may be needed before selecting a production candidate.

The calculation is linear and deliberately simple. If the engineered phage reaches only 10⁹ PFU/mL rather than 10¹⁰, the required culture volume rises from 20 L to 200 L. Measuring real titre and recovery at small scale is therefore essential before choosing production equipment.

5. Main unknowns

1. Can the models generate useful RBP designs? A high model score or plausible predicted structure may not correspond to correct receptor binding, phage assembly, or host specificity. It is also unclear how much training data directly connects RBP sequence to bacterial receptor and host range.
2. Does recognition lead to productive infection in a biofilm? An RBP may bind the target cell but still fail at irreversible attachment, genome entry, replication, or escape from bacterial defenses. The biofilm matrix may also block access to receptors.
3. Will the design manufacture well and remain selective? RBP changes may reduce assembly, stability, burst size, or final titre, and the engineered phage may affect unintended strains. The assumed 10¹⁰ PFU/mL titre, 50% recovery, and 10⁷ PFU/cm² dose all need experimental validation.

The first milestone is therefore modest and testable: use generative models (ESM-3, Evo-2) to propose and filter a small panel of RBP variants, identify at least one that remains functional, and measure whether it changes recognition within a safe, non-pathogenic model community. Even a negative result would be useful because it would reveal where model scores fail to predict the full phage-host interaction and would provide experimental data for a better second design round.

Fallback: soluble RBP binding assay

If phage assembly fails, we have a fallback: express the designed RBP on its own as a soluble protein and measure its binding to target and non-target cells directly.

This isolates the recognition question from the rest of phage biology. A full phage-based test bundles many steps together: folding, assembly onto the capsid, genome entry, evasion of bacterial defenses, replication, and lysis. If the phage fails, it is not clear which step broke. The soluble-protein test removes all of that and checks exactly one thing: whether the designed protein binds the receptor it was designed for, and not the others.

This cannot show a working phage. But it still tests the core hypothesis of the project, that a generative model can design a protein with programmable binding specificity, in a simpler setting with fewer ways to fail. A phage-assembly failure would not prevent me from getting a clear answer on whether the AI-driven design approach itself works.

AI use disclosure

I used OpenAI ChatGPT (GPT-5.6 Luna) to brainstorm the concept, identify and compare relevant recent literature, and check the back-of-the-envelope estimates. I reviewed and selected the final scientific assumptions, citations, and framing.

Representative prompt: “Help me develop an bacteriophage platform that bridges generative AI to biomanufacturing.” Later finetuned the proposal to include more details about the biological chassis, manufacturing workflow, meaningful application, scale and timing, technical challenges, and fallback.

Representative literature prompt: “Find recent literature on AI-guided protein design, phage receptor-binding proteins, phage host-range engineering, and programmable phage platforms. Distinguish what has already been demonstrated from what remains novel in the proposed biomanufacturing experiment.”