The Pulse
Microsoft Maps an Inspectable AI Workflow for Phage Therapy
Microsoft says its Discovery app helped researchers design a reviewable bioinformatics pipeline to find phages that may target drug-resistant bacteria. The proposed workflow ranks candidates for further study but still requires laboratory v
AI.info Team · Written with AI ·
Microsoft says its Discovery app helped a research team turn a phage-therapy question into a proposed, inspectable bioinformatics pipeline—but the demonstration did not establish that any phage treats an infection. The company published the account on September 30, 2026, describing a workflow intended to help scientists find and prioritize viruses that may attack drug-resistant bacteria.
The distinction matters: the team used the app to organize research tasks, build a prototype and review its outputs. The proposed pipeline points toward candidates for laboratory testing; it is not a clinical result or evidence that a treatment works.
From a pathogen genome to candidate phages
The workflow starts with the genome of a target bacterium. It checks for antibiotic-resistance markers and potential antiphage defenses, then searches environmental metagenomic information for phages that might infect the target. Microsoft says the process queries NCBI and Europe PMC for phage genome records, related literature and selected metadata.
Candidate selection combines several kinds of evidence rather than relying on a single prediction: DNA similarity, known phage-bacterium relationships, predicted interaction with the bacterial surface and the bacterium’s defenses. The pipeline is designed to remove incomplete or potentially unsafe options and rank the remaining candidates by predicted effectiveness, safety and production practicality. It excludes temperate or lysogenic phages, which can integrate into bacterial genomes.
The system can also suggest limited engineering if no natural phage meets the requirements, but Microsoft says an engineered candidate would need checks for safety, stability and unintended effects. The workflow recommends follow-up experiments, including tests of infection, bacterial resistance, antibiotic combinations, safety and production.
Scientists can inspect the agent’s work
Microsoft presents inspectability as a design feature: the Discovery app turns a scientific request into a task tree, with methods, evidence, inputs and outputs that researchers can review and revise. The team also organized scientific literature and data sources, then expanded the agents’ access to tools and databases through Model Context Protocol adapters.
The post says the project began with a curated collection of 170 articles. Six local MCP servers connected the agents to resources including PubMed, Europe PMC, NCBI BLAST, Semantic Scholar, ClinicalTrials.gov and AlphaFold. Microsoft says the agents revised their research approach as new evidence and expert feedback came in, while scientists approved changes before affected tasks proceeded.
This expert review was essential for distinguishing scientifically grounded results from plausible-sounding but unsupported AI output and for identifying areas that required refinement before the pipeline could be trusted and reused.
Christine Caggiano, Microsoft Discovery Blog
Four hours produced a draft, not a therapy
Microsoft says the work moved from generating a task tree to a draft report in less than four hours, including time spent sourcing genomic references and running bioinformatics workflows. The app proposed a six-module architecture with defined inputs and outputs, quality and safety checks, alternative approaches, control datasets and a laboratory-validation plan.
The team also used an API prototype to check that the computational pipeline operated as intended. Bioinformatics experts reviewed its stages and whether information passed coherently between them. That verification concerns the software workflow; it does not show that the predicted phages infect or kill a target bacterium.
Wet-lab testing remains the next threshold
Microsoft describes the result as an implementation-ready research pipeline and says expert judgment remained involved in consequential decisions. The company’s post calls for laboratory experiments to confirm whether a candidate can infect and kill the target, assess resistance and test safety, stability and production. It also proposes using experimental findings to improve future predictions.
The demonstration therefore offers a plan for narrowing a search, not a shortcut around biological testing. Microsoft reports no experimentally confirmed phage candidates, infection outcomes or patient treatments from this exercise.