mzLearn, a data-driven LC/MS signal detection algorithm, enables pre-trained generative models for untargeted metabolomics

This repository contains the codebase for mzEmbed, a framework for developing pre-trained generative models and fine-tuning them for specific tasks for untargeted metabolomics datasets.

Author:

• Leila Pirhaji

Overview of mzLearn

mzLearn is a data-driven algorithm designed to autonomously detect metabolite signals from raw LC/MS data without requiring input parameters from the user. The algorithm processes raw LC/MS data files in the open-source mzML format, iteratively learning signal characteristics to ensure high-quality signal detection.

Key Features of mzLearn:

• Zero-parameter design: No prior knowledge or QC samples are required.
• Iterative learning: mzLearn autonomously refines signal detection, correcting for retention time (rt) and intensity drifts caused by batch effects and run order.
• Output: A two-dimensional table of detected features defined by median rt and m/z values, with normalized intensities across samples.
• Scalability: Capable of handling large-scale datasets (e.g., 2,075 files in a single run).
• Accessibility: mzLearn’s website for accessing the tool is available at http://mzlearn.com/.

---

Overview of mzEmbed Codebase

mzEmbed extends mzLearn’s capabilities by combining outputs from multiple datasets to develop pre-trained generative models and applying them to a range of metabolomics applications.

Key Components of mzEmbed:

1. Pre-trained Model Development:

   • Combines metabolomics data from multiple studies to create robust pre-trained generative models.
   • Supports Variational Autoencoders (VAEs) for unsupervised learning of metabolite representations.
   • Enables parameter optimization using grid search and Optuna for hyperparameter tuning.
   • Outputs embeddings that capture biological and demographic variability, such as age, disease state.

2. Fine-Tuning Pre-Trained Models:

   • Allows fine-tuning of pre-trained models on independent datasets for improved task-specific performance.
   • Supports fine-tuning for binary classification, multi-class classification, and survival analysis.

3. Task-Specific Model Refinement:

   • Retrains the last layer of fine-tuned models for specific tasks, such as clinical classifcation and surivival analysis.

4. Advanced Architectures:

   • Supports the development of joint learning models for treatment-independent, prognostic stratification of patient.
   • Implements adversarial learning to isolate treatment-specific predictive biomarkers, or predictive stratification of patient.

---

Getting Started

Requirements

• Python 3.9 or higher

Installation

1. Clone the repository:

   git clone [email protected]:ReviveMed/mzEmbed.git
   cd mzEmbed

2. Install dependencies:

   pip install -r requirements.txt

3. building the package:

   cd mz_embed
   python -m build
   pip install -e .
   

---

Usage:

The repository supports six main use cases, including pretraining, fine-tuning, and advanced learning architectures. pretrain and finetune directories includes examples of the Python commands for each use case.

---

License

This project is licensed under the Academic and Non-Profit Use License. See the LICENSE.txt file for details.

---

Citation

If you use mzLearn or mzEmbed in your research, please cite:

mzLearn, a data-driven LC/MS signal detection algorithm, enables pre-trained generative models for untargeted metabolomics Leila Pirhaji, Jonah Eaton, Adarsh K. Jeewajee, Min Zhang, Matthew Morris, Maria Karasarides

---