TSUMUGI: a platform for phenotype-driven gene network identification from comprehensive knockout mouse phenotyping data

TSUMUGI is a comprehensive platform that leverages International Mouse Phenotyping Consortium (IMPC) data to identify and visualize phenotype-driven gene networks, enabling researchers to systematically interpret complex organismal phenotypes and generate hypotheses about coordinated gene functions.

The Gist

Imagine you are a detective trying to solve a massive mystery: How do our bodies work?

For a long time, scientists have been looking at individual suspects (genes) one by one. They knock out a single gene in a mouse and see what goes wrong. This is like checking if a single brick is missing from a wall. But the real mystery is how the whole wall stands up. Complex traits—like aging, diabetes, or how our hearts beat—are rarely caused by just one gene. They are the result of a whole team of genes working together, like a well-oiled machine or a symphony orchestra.

The problem is that we have a mountain of data from the International Mouse Phenotyping Consortium (IMPC). They have tested over 9,000 different mice, each missing a different gene. It's like having a giant library of "what happens if we remove this specific part?" But reading 9,000 separate reports is overwhelming. It's hard to see the big picture or find which genes are actually working together as a team.

Enter TSUMUGI.

What is TSUMUGI?

Think of TSUMUGI as a smart matchmaking app for genes.

Instead of looking at genes in isolation, TSUMUGI looks at the "personality" of the mice. If Mouse A (missing Gene X) and Mouse B (missing Gene Y) both end up with the same weird symptoms—say, they both have trouble walking and their fur falls out—TSUMUGI says, "Hey! These two genes must be part of the same team!"

It connects the dots between genes based on the symptoms they cause, rather than just their chemical structures.

How Does It Work? (The Analogy)

Imagine you are at a huge party with thousands of people (the genes). You don't know who knows whom.

1. The Clue: You notice that three people are all wearing red shoes and holding blue umbrellas (the symptoms/phenotypes).
2. The Match: TSUMUGI realizes that these three people probably know each other because they share these specific traits. It draws a line between them.
3. The Network: Suddenly, you see clusters of people. One group is all wearing red shoes (a "walking problem" team). Another group is holding blue umbrellas (a "fur loss" team).

TSUMUGI takes the massive list of mouse symptoms and turns it into a visual map (a network).

• The Dots (Nodes): These are the genes.
• The Lines (Edges): These show how similar the genes are. A thick, strong line means the genes cause very similar problems, so they are likely close friends working on the same job.
• The Colors: Darker colors might mean the problem is severe; lighter colors mean it's mild.

Two Ways to Use It

The paper describes two ways to use this tool, like having a tourist guide and a professional toolkit:

1. The Web App (The Tourist Guide):

   • This is for anyone who wants to explore. You can type in a symptom you care about (e.g., "diabetes") or a gene you know about.
   • The app instantly draws the map, showing you all the other genes that are "friends" with your gene or share that symptom.
   • You can click around, zoom in, and filter the map. It's interactive and easy to understand, like using Google Maps to find a route.

2. The Command Line (The Professional Toolkit):

   • This is for scientists who want to do deep, custom analysis. They can write code to say, "Show me only the gene teams that affect female mice at old age, but ignore anything related to reproduction."
   • It's like a power-user mode that lets you build your own custom map with very specific rules.

Why Is This a Big Deal?

Before TSUMUGI, scientists had to guess which genes worked together. They often relied on old textbooks or known chemical interactions. But biology is messy, and genes often team up in ways we didn't expect.

TSUMUGI is different because it learns from the data itself. It doesn't need to know the rules beforehand; it just looks at the results.

• It finds hidden teams: It can discover that Gene A and Gene B are working together on a specific disease, even if no one knew they were related before.
• It helps with human health: Since mice and humans are very similar, if TSUMUGI finds a gene team causing a problem in mice, it gives us a huge clue about what might be causing similar problems in humans.

The Bottom Line

TSUMUGI is a powerful new lens that helps us stop looking at genes one by one and start seeing them as teams. By turning thousands of mouse experiments into a clear, interactive map, it helps scientists generate new ideas and hypotheses about how complex diseases work, potentially leading to better treatments for humans in the future.

It's essentially taking a chaotic pile of puzzle pieces and showing us exactly how they fit together to build the picture of life.

Technical Summary

1. Problem Statement

Deciphering complex organismal phenotypes requires understanding the coordinated functions of multiple genes. While existing methods like protein-protein interaction (PPI) networks, biological pathways, and gene co-expression networks are useful, they primarily focus on molecular or intracellular relationships and often fail to capture gene interactions at the organismal level. Similarly, Genome-Wide Association Studies (GWAS) identify risk factors but do not explicitly map the coordinated gene interactions that drive specific phenotypes.

Although the International Mouse Phenotyping Consortium (IMPC) has generated a massive, standardized dataset of phenotypic profiles for over 9,000 single-gene knockout (KO) mouse lines, extracting multi-gene relationships from this data remains challenging. The IMPC data is organized around single-gene knockouts, making it difficult to systematically identify gene modules that share phenotypic signatures without a dedicated analytical framework.

2. Methodology

The authors developed TSUMUGI (Trait-driven Surveillance for Mutation-based Gene module Identification), a platform that constructs phenotype-driven gene networks by quantifying phenotypic similarity between KO mouse lines.

Data Sources:

• IMPC Statistical Data: Extracted from release 23.0, including MGI marker symbols, Mammalian Phenotype Ontology (MP) terms, effect sizes, P-values, sex-specific P-values, zygosity, and pipeline/procedure names.
• Disease Models: Mouse orthologues of human disease-associated genes from the IMPC Disease Models Portal.
• Ontology: MP annotation files from OBO Foundry (release 2025-08-27) for similarity calculations.

Phenotype Similarity Calculation (The Core Algorithm):
TSUMUGI utilizes the Phenodigm algorithm to calculate similarity scores between gene pairs:

1. Abnormal Phenotype Extraction: Records with assigned MP terms (P-value < 0.00001) are treated as abnormal phenotypes.
2. Term Similarity: Calculated as the geometric mean of the Resnik score (based on Information Content of the ontology) and the Jaccard index (overlap of parent term sets). Terms with low Information Content (top 5th percentile) are penalized to avoid overly general matches.
3. Metadata Concordance: A weighted similarity matrix is constructed by multiplying term similarity by a metadata concordance score. This score adjusts for matches in genotype, sex, and life stage (ranging from 1.0 for full match to 0.25 for multiple mismatches).
4. Scaling: The final output is a phenotype similarity score between 0 and 100 for each gene pair.

Network Construction:

• Nodes: Represent genes.
• Edges: Represent phenotype similarity. An edge is established if two genes share at least three abnormal phenotypes.
• Modules: Connected components in the graph represent gene modules with high internal phenotypic similarity.

Implementation:

• Web Application: Built with Cytoscape.js (v3.30.4) for interactive visualization.
• Command-Line Interface (CLI): Distributed via PyPI and Bioconda for reproducible, large-scale local analysis.

3. Key Contributions

• Phenotype-Driven Network Construction: Unlike knowledge-based networks (e.g., PPI), TSUMUGI constructs networks de novo directly from organism-level phenotype data, uncovering functional modules that may be missed by pre-defined interaction maps.
• Flexible Entry Points: Users can initiate analysis via:
  • Phenotype: Find genes sharing a specific MP term.
  • Gene: Find genes with phenotypes similar to a query gene.
  • Gene List: Group and analyze a custom list of genes.
• Advanced Filtering & Highlighting:
  • Filters based on similarity scores, severity, zygosity, sex, and life stage (Embryo, Early, Interval, Late).
  • Highlights genes associated with human diseases.
  • Performs centrality analyses to identify hub genes.
• CLI Capabilities: The command-line tool allows for complex exclusion/inclusion logic (e.g., "extract genes associated with diabetes phenotypes but exclude those associated with reproductive phenotypes"). It also leverages IMPC data to distinguish between "unmeasured" and "measured but non-significant" phenotypes, improving data precision.
• Interoperability: Supports export in PNG, JPG, SVG, CSV, and GraphML formats for downstream integration.

4. Results and Features

• Visualization: The web interface displays a network where node color intensity indicates phenotype severity, and edge thickness reflects the similarity score. Subnetworks (modules) are automatically identified, allowing users to inspect enriched phenotypes within specific clusters.
• Interactive Exploration: Users can iteratively narrow down modules by selecting specific phenotypes or applying metadata filters (e.g., focusing only on female-specific phenotypes or late-life stages).
• Case Study Potential: The authors demonstrate the ability to identify sex-dependent gene modules and aging-related traits by filtering for specific life stages or sexes, serving as a hypothesis-generating framework.

5. Significance

TSUMUGI addresses a critical gap in functional genomics by bridging the divide between large-scale mouse phenotyping data and the discovery of multi-gene interactions at the organismal level.

• Hypothesis Generation: It provides a systematic framework to link shared phenotypic signatures to putative functional gene modules, aiding in the discovery of mechanisms behind complex traits and human diseases.
• Beyond PPI: By relying on phenotypic convergence rather than pre-existing molecular interaction maps, it can reveal novel functional relationships that are not yet annotated in databases.
• Future-Proofing: The platform is designed to integrate additional layers in the future, such as PPI, tissue-specific gene expression, and temporal dynamics, to further refine the mechanistic context of identified gene modules.

Availability:

• Web: https://larc-tsukuba.github.io/tsumugi/
• CLI: PyPI (TSUMUGI) and Bioconda.
• Source Code: GitHub and Zenodo (MIT License).