A Shared Flame for the Community: Adding Torch to the DLOmix framework for Deep Learning Proteomics

[ayla-s]

Title

A Shared Flame for the Community: Adding Torch to the DLOmix framework for Deep Learning Proteomics

Abstract

DLOmix is a Python framework designed to simplify the development, training, and comparison of deep learning models in MS-based proteomics. It provides researchers with powerful, user-friendly tools to work with complex proteomics data. By integrating existing models and their common building blocks as readily reusable, executable, but also customizable components, DLOmix will facilitate easier comparisons between these models, enhancing researchers' ability to evaluate their performance and applicability. However, the current implementation of DLOmix relies primarily on Keras/TensorFlow, which may limit accessibility for users who prefer alternative frameworks, such as PyTorch.

DLOmix coverage of areas and workflows in a typical applied deep learning research setup.

We propose to extend the DLOmix framework by integrating support for PyTorch. This will involve adapting key components - such as dataset management, model architectures, and evaluation tools - to create PyTorch equivalents. This enhancement will improve the framework's usability and offer greater flexibility for researchers.

This update will make deep learning techniques more accessible to a wider range of researchers in the proteomics community, particularly those who find PyTorch more intuitive. As already done for the existing DLOmix framework, we will also develop tutorials and example notebooks to help users navigate the new features and transition from TensorFlow to PyTorch seamlessly.

Following the hackathon, we are dedicated to maintaining and documenting the DLOmix repository to ensure it remains a valuable resource for researchers interested in applying deep learning to their work. By fostering collaboration and knowledge sharing, we aim to drive innovation and progress in the field of MS-based proteomics.

Project Plan

The DLOmix hackathon will be split into the following steps, subtasks in 2., 3. and 4. can easily be parallelized:

1. Introduction of the DLOmics repo and its components, currently supporting the TensorFlow framework.
2. Implement/integrate existing PyTorch equivalents for:
   2.1 extend dataset classes for PyTorch based on HuggingFace datasets, which is integrated already for TensorFlow Datasets.
   2.2 model components and common building blocks
   2.3 existing models integrated in DLOmix, e.g., Prosit RT or Fragment Ion Intensity predictor
   2.4 evaluation utilities, e.g. custom losses and metrics
   2.5 common model architectures
3. Create tutorials/example notebooks for the usage of the new functionalities, e.g. training the PyTorch models.
4. optional, if there is time, integrate into codebase and documentation:
   4.1 additional models
   4.2 reporting utilities
5. Draft a manuscript to report the PyTorch updates.

The DLOmix repository is actively maintained and the integration as well as documentation will be continued after the hackathon.

Technical Details

Programming Language: Python, TensorFlow, PyTorch

Existing implementation: Keras/TensorFlow https://github.com/wilhelm-lab/dlomix

Datasets we typically use are based on:

• ProteomeTools (https://www.proteometools.org/)
• Its processed form PROSPECT (https://github.com/wilhelm-lab/PROSPECT)
• Ready-to-use HuggingFace Collection for three formulated machine learning tasks (https://huggingface.co/Wilhelmlab)

Contact information

Wilhelmlab, Computational Mass Spectrometry, TU Munich
Ayla Schröder [email protected]
Omar Shouman [email protected]

[theGreatHerrLebert]

Really good idea! Maybe also checkout Keras3, which would allow to define the deep learning models in an abstract fashion and then let the user decide which backend to use for inference.