U-Probe: universal agentic probe design for imaging-based spatial-omics

U-Probe is a universal, agentic platform that leverages a DAG-based assembly engine and LLM-driven conversational workflows to overcome the limitations of expert-dependent, rigid probe design tools, enabling the rapid generation of synthesis-ready sequences for diverse and novel spatial-omics architectures.

The Gist

Imagine you are a master chef trying to cook a complex dish (studying how genes work inside a cell). To do this, you need specific, custom-made utensils (probes) that can grab onto specific ingredients (genes) without touching anything else.

For a long time, making these utensils was like trying to build a custom wrench using only a hammer and a screwdriver. You needed to be a highly trained engineer (a bioinformatics expert) to understand the blueprints, calculate the exact angles, and ensure the wrench wouldn't break the ingredients. If you wanted to invent a new type of wrench, you had to start from scratch and write your own code.

Enter U-Probe: The "AI Chef's Assistant" for Gene Hunting.

This new tool, called U-Probe, changes the game completely. Here is how it works, explained simply:

1. The Problem: The "One-Size-Fits-None" Dilemma

Scientists have been discovering new ways to look inside cells (like taking 3D photos of genes). Each new method requires a slightly different "utensil" design.

• Old Tools: Were like rigid cookie cutters. They only worked for specific shapes (like MERFISH or seqFISH). If you wanted a star-shaped cookie, you had to buy a whole new machine or build one yourself.
• The Barrier: You needed a PhD in computer science just to tweak the settings. If you didn't know the "secret language" of the machine, you couldn't make your probes.

2. The Solution: U-Probe's "Lego" System

U-Probe is built like a digital Lego set.

• The Blocks: Instead of hard-coding one specific shape, U-Probe lets you define probes as modular "blocks" (target sequences, barcodes, amplifiers).
• The Assembly Line: It uses a smart assembly engine (called a DAG) that snaps these blocks together in the right order. Whether you want a simple stick, a complex loop, or a brand-new shape never seen before, U-Probe can assemble it without needing to rewrite the factory code.
• The Analogy: Imagine you are ordering a custom pizza. Old tools only let you choose from a menu of 5 pre-set pizzas. U-Probe lets you say, "I want pepperoni on the left, pineapple on the right, and a crust made of cheese," and it builds it instantly.

3. The Magic Feature: Talking to the Machine

This is the coolest part. You don't need to know the "Lego" instructions or the computer code. You can just chat with it.

• The AI Agents: U-Probe has a team of AI assistants (a "Leader," a "Panel Designer," and a "Probe Builder").
• How it works: You can type (or speak) in plain English: "I need to find which genes are active in a mouse lung infected with the flu, and I want to see where the immune cells are."
• The Result: The AI team figures out the rest. They pick the right genes, design the custom "Lego" structure, check that the probes won't stick to the wrong things, and give you the final recipe ready to be synthesized in a lab.

4. Real-World Proof: Did it work?

The scientists tested this "AI Chef" in three different kitchens:

1. The Flu Lab: They designed probes to map out exactly where different immune cells were hiding in a mouse lung infected with bird flu. The AI did the design, and the results showed a clear picture of the infection battle.
2. The Virus Hunter: They designed probes to hunt down three different types of herpes viruses in cells. The tool created thousands of tiny probes that covered the entire virus genome, allowing for super-fast detection.
3. The Tiny Difference: They designed a probe to spot a single letter change in a virus's DNA (like telling the difference between a 'G' and an 'A'). This is incredibly hard to do, but U-Probe's flexible design system made it possible.

Why This Matters

Before U-Probe, if you wanted to study a new virus or a new way of looking at cells, you had to hire a team of computer experts to build a custom tool. Now, a biologist can just describe what they want in English, and the AI builds the tool for them.

In short: U-Probe turns the complex, code-heavy world of genetic engineering into a simple conversation, allowing anyone to build custom tools to explore the microscopic world of life.

Technical Summary

1. Problem Statement

The field of imaging-based spatial-omics (e.g., spatial transcriptomics, 3D genome studies, and clinical diagnostics) relies heavily on Fluorescence In Situ Hybridization (FISH). However, the design of FISH probes faces two critical bottlenecks:

• Expertise Barrier: Existing design tools (e.g., OligoMiner, PaintSHOP) require deep domain knowledge. Users must manually understand probe architectures, select thermodynamic parameters (melting temperature, secondary structure), and evaluate off-target specificity. This creates a high barrier for non-computational biologists.
• Architectural Fragmentation: Rapidly emerging spatial-omics methods (e.g., PRISM, TDDN-FISH, MiP-seq) introduce novel probe architectures that existing tools cannot support. Current tools often "hard-code" support for specific protocols (like MERFISH or seqFISH), forcing labs developing new methods to build custom pipelines from scratch.

2. Methodology: The U-Probe Platform

The authors present U-Probe, a universal, agentic probe design platform that addresses these limitations through two core technical innovations:

A. Declarative Configuration & DAG-Based Assembly Engine

• Declarative System: Instead of hard-coding probe structures, U-Probe uses a YAML-based configuration system. Users define probes as modular templates with named parts (e.g., target arms, barcodes, linkers).
• Directed Acyclic Graph (DAG) Engine: The system represents probe assembly as a DAG.
  • Expression Nodes: Perform operations on sequences (slicing, reverse complement, barcode lookup).
  • Template Nodes: Concatenate child outputs into final probe sequences.
  • Topological Sort: The engine resolves dependencies to assemble arbitrarily complex, multi-part probes without code modifications.
• Quality Control: The engine computes standard metrics (GC content, $T_m$, secondary structure stability via ViennaRNA, off-target mapping via Bowtie2, k-mer frequency via Jellyfish) and applies user-defined filtering and sorting logic.

B. LLM-Based Multi-Agent System

• Framework: Built on the PantheonOS framework, U-Probe integrates Large Language Model (LLM) agents to enable conversational design.
• Agent Hierarchy:
  • Leader Agent: Parses natural language requests and coordinates the workflow.
  • Panel Agent: Handles high-level tasks like analyzing scRNA-seq data, identifying marker genes, and optimizing gene panels.
  • Probe Agent: Translates design specifications into YAML configuration files and invokes the core engine.
• Workflow: Users can describe experimental goals (e.g., "Design a MiP-seq panel for influenza-infected mouse lung") in plain English. The agents autonomously generate configuration files, select parameters, and output synthesis-ready sequences.

3. Key Contributions

1. Universal Architecture: U-Probe is the first tool capable of designing arbitrary probe structures (from established protocols like MERFISH to entirely novel architectures) via a declarative configuration system, eliminating the need for code rewrites when new methods emerge.
2. Democratization via AI: By integrating LLM agents, U-Probe lowers the expertise barrier, allowing non-experts to design complex spatial-omics experiments through natural language interaction.
3. Open-Source Ecosystem: The tool is released as an open-source package with CLI, Web, and Agent interfaces, ensuring accessibility and reproducibility.

4. Experimental Results & Validation

The authors validated U-Probe across three distinct scenarios:

• Scenario 1: Agent-Driven MiP-seq Panel Design (Influenza)
  • Task: Designed a MiP-seq probe panel for H5N6 and H1N1 influenza-infected mouse lung tissue using scRNA-seq data.
  • Outcome: The agent autonomously selected marker genes and designed probes. Imaging successfully resolved spatial cell type distributions and differential immune infiltration patterns, validating the agent's ability to bridge single-cell and spatial data.
• Scenario 2: Genome-Tiling DNA-FISH (Herpesviruses)
  • Task: Designed genome-tiling probe pools for three herpesviruses (FHV, PRV, HSV).
  • Outcome: Achieved high genome coverage (77–98%) and enabled rapid, sensitive detection of viral infections in cultured cells.
• Scenario 3: Novel Architecture (Single-Nucleotide Mutation)
  • Task: Designed a novel RCA-based ligation probe to discriminate the H9N2 avian influenza PB2 627E/K polymorphism (a single-nucleotide mutation).
  • Outcome: This architecture (involving circle probes and ligation) could not be expressed by existing tools. U-Probe successfully generated the design, achieving allele-specific discrimination and multiplexed co-infection detection in a single experiment.

5. Significance

• Future-Proofing: U-Probe solves the "architectural fragmentation" problem. As new spatial-omics technologies emerge, researchers can define new probe structures via configuration files rather than waiting for software updates.
• Accessibility: The conversational AI interface bridges the gap between computational complexity and biological application, enabling researchers in non-model organisms or clinical settings to perform high-level spatial analysis without dedicated bioinformatics support.
• Scalability: The system is designed to interface with genomes of any species, including polyploid organisms, and can evolve alongside the PantheonOS framework to incorporate new design strategies automatically.

In summary, U-Probe represents a paradigm shift in probe design, moving from rigid, expert-dependent pipelines to a flexible, AI-driven, and universally adaptable platform for the next generation of spatial-omics research.