MOBIUS

MOBIUS – Multi-Omics Biomarker Integration User-friendly Suite

MOBIUS is a modular, browser-based framework for the analysis and integration of heterogeneous multi-omics tabular data, with a particular focus on biomarker discovery, feature selection, machine learning, and network-based analysis.

Rather than enforcing a predefined end-to-end analytical pipeline, MOBIUS provides interoperable components that can be flexibly combined to build and compare customized multi-omics workflows.

It supports early, late, and hybrid integration strategies, enabling researchers to systematically investigate how alternative data-integration and feature-selection approaches affect predictive performance, signature complexity, and biological interpretability.

---

Key Features

MOBIUS provides an integrated environment for:

• Multi-omics data preparation and harmonization
• Exploratory Data Analysis (EDA)
• Correlation network construction
• Complex network analysis and visualization
• Network pruning and network comparison
• Community detection
• Traditional feature selection
• Ensemble feature selection
• Network-aware feature selection
• Machine-learning model training and evaluation
• Early, late, and hybrid multi-omics integration
• Cross-validation and independent test-set evaluation
• Interactive visualization of results
• Reproducible analysis through Docker containerization

---

MOBIUS Architecture

MOBIUS is organized into three main modules:

1. Data Preparation

The Data Preparation module handles the import, harmonization, and preprocessing of metadata and omics datasets.

It supports:

• metadata and omics table import;
• sample matching across multiple omics layers;
• configurable data transformations;
• supervised task definition;
• training/test splitting;
• correlation network construction;
• optional external validation datasets.

Correlation networks represent omics features as nodes and their statistical relationships as weighted edges.

---

2. Complex Network Analysis

The Complex Network module provides tools for exploring and manipulating correlation networks.

Available operations include:

• network loading and subnetwork extraction;
• threshold-based edge pruning;
• centrality analysis;
• community detection;
• interactive network visualization;
• comparison of networks across experimental conditions;
• identification of common and condition-specific relationships.

Networks can therefore be analyzed, pruned, and compared to refine their topology, reduce redundant signals, and identify biologically interpretable feature communities.

---

3. Machine Learning Analysis

The Machine Learning module provides multiple strategies for biomarker identification and predictive-model evaluation.

MOBIUS includes:

Exploratory Data Analysis

• PCA
• Kernel PCA
• t-SNE
• feature distribution visualization
• statistical testing

Traditional Feature Selection

Filter, embedded, and wrapper approaches implemented using standard machine-learning libraries.

Ensemble Feature Selection

Multiple feature selectors are combined across bootstrap replicas to identify robust and recurrent features.

Network-based Feature Selection

MOBIUS implements a network-aware minimum Redundancy Maximum Relevance (mRMR) strategy formulated as a multi-objective optimization problem.

Feature relevance and redundancy are optimized simultaneously using the NSGA-II evolutionary algorithm.

Feature relevance can be estimated using:

• univariate importance scores;
• model-driven SHAP importance scores.

Network communities are exploited to account for redundancy among correlated molecular features.

Model Evaluation

Feature signatures can be evaluated using multiple classification algorithms, including:

• Logistic Regression
• Support Vector Machines
• k-Nearest Neighbors
• Naive Bayes
• Random Forest
• Gradient Boosting

Evaluation can be performed using configurable cross-validation strategies, including k-fold CV and leave-one-out cross-validation (LOOCV), as well as independent external test datasets.

---

Multi-Omics Integration

MOBIUS enables the construction and comparison of different multi-omics integration paradigms.

Early Integration

Omics layers are combined before feature selection and model training.

Omics 1 ─┐
Omics 2 ─┼─> Joint Feature Selection ─> Model
Omics 3 ─┘

Late Integration

Feature selection and/or predictive modeling are performed independently for each omics layer before combining the resulting signatures or predictions.

Omics 1 ─> Feature Selection ─┐
Omics 2 ─> Feature Selection ─┼─> Integration ─> Model
Omics 3 ─> Feature Selection ─┘

Hybrid Integration

MOBIUS components can be combined to construct customized strategies mixing early- and late-integration steps.

This modular design allows systematic and reproducible comparison of alternative multi-omics workflows.

---

Installation

The recommended way to run MOBIUS is through Docker Compose.

Requirements

You need:

• Docker
• Docker Compose

No manual installation of Python packages or system dependencies is required when using the provided Docker image.

---

Quick Start with Docker Compose

Create a file named:

docker-compose.yml

with the following content:

services:

  mongo:
    image: mongo:7.0
    container_name: mongo_db
    restart: unless-stopped

    ports:
      - "27018:27017"

    volumes:
      - mongo_data:/data/db

    networks:
      - mobius_network


  mobius:
    image: cursecatcher/mobius:mongo
    pull_policy: always
    container_name: mobius
    restart: unless-stopped

    ports:
      - "8501:8501"

    environment:
      - MONGO_URI=mongodb://mongo:27017

    depends_on:
      - mongo

    networks:
      - mobius_network


networks:
  mobius_network:


volumes:
  mongo_data:

Start MOBIUS with:

docker compose up

or run it in the background with:

docker compose up -d

Then open:

http://localhost:8501

in your web browser.

---

Stop MOBIUS

To stop the containers:

docker compose down

Persistent MongoDB data are retained in the Docker volume.

To also remove stored data:

docker compose down -v

Warning: the -v option permanently removes the MongoDB volume and all data stored by MOBIUS.

---

Build from Source

Clone the repository:

git clone https://github.com/qBioTurin/mobius.git
cd mobius

Build the MOBIUS Docker image:

docker build -t mobius .

The recommended production deployment uses Docker Compose because MOBIUS relies on MongoDB for persistence and GridFS-based storage.

---

Technology Stack

MOBIUS is primarily developed in Python 3 and uses:

• Streamlit — browser-based graphical interface
• NumPy
• pandas
• scikit-learn
• graph-tool
• pymoo — multi-objective optimization
• Plotly — interactive visualization
• SHAP — model interpretation
• MongoDB — data persistence and caching
• MongoDB GridFS — storage of large datasets
• Docker / Docker Compose — reproducible deployment

---

Typical Workflow

A typical MOBIUS analysis consists of:

Multi-Omics Data
       │
       ▼
Data Preparation
       │
       ├── Data harmonization
       ├── Transformations
       ├── Train/Test definition
       └── Network construction
       │
       ▼
Exploratory Analysis
       │
       ▼
Complex Network Analysis
       │
       ├── Network analysis
       ├── Community detection
       ├── Network pruning
       └── Network comparison
       │
       ▼
Feature Selection
       │
       ├── Traditional
       ├── Ensemble
       └── Network-based mRMR / NSGA-II
       │
       ▼
Model Evaluation
       │
       ├── Cross-validation
       ├── Independent validation
       └── Performance comparison
       │
       ▼
Multi-Omics Biomarker Signature

Because individual components are interoperable, users are not restricted to this sequence and can construct alternative workflows according to the biological question under investigation.

---

Applications

MOBIUS has been evaluated on real multi-omics datasets addressing different classification problems.

The framework enables researchers to compare alternative integration strategies and identify compact, predictive, and biologically interpretable multi-omic signatures.

In particular, MOBIUS facilitates the investigation of trade-offs among:

• predictive performance;
• number of selected biomarkers;
• redundancy among selected features;
• robustness of the signature;
• biological interpretability.

---

Reproducibility

Reproducibility is a core design principle of MOBIUS.

The combination of:

• modular analytical components;
• explicit workflow construction;
• parameter tracking;
• persistent result storage;
• Docker containerization;

makes it possible to reproduce and compare alternative multi-omics analyses within a controlled computational environment.

---

Citation

If you use MOBIUS in your research, please cite:

Licheri N., Sirovich R., Ferrero G., Pardini B., Naccarati A., Aucello R., Beccuti M., Cordero F.
MOBIUS: a Multi-Omics Biomarker Integration User-friendly Suite via machine learning and network representations.

Citation information will be updated upon publication.

---

Authors

MOBIUS was developed by researchers from:

• Department of Computer Science, University of Turin, Italy
• Italian Institute for Genomic Medicine (IIGM), c/o IRCCS Candiolo, Turin, Italy
• Department of Clinical and Biological Sciences, University of Turin, Italy
• Department of Mathematics “Giuseppe Peano”, University of Turin, Italy

Authors:

• Nicola Licheri
• Roberta Sirovich
• Giulio Ferrero
• Barbara Pardini
• Alessio Naccarati
• Riccardo Aucello
• Francesca Cordero
• Marco Beccuti

---

License

Please refer to the LICENSE file for information about the terms of use and redistribution of MOBIUS.

---

Repository

Source code:

https://github.com/qBioTurin/mobius

Docker image:

https://hub.docker.com/r/cursecatcher/mobius

---

Contact

For questions, bug reports, or feature requests, please use the GitHub issue tracker:

https://github.com/qBioTurin/mobius/issues