Master Thesis Code - Felix Krause

Main Repository for Data Science Master Thesis at University of Vienna 2024-25:

"Benchmarking and Optimizing Deep Learning Architectures for Protein-to-mRNA Ratio Prediction"

Link selection

• Reference paper: Hernandez-Alias et al (2023), Using protein-per-mRNA differences among human tissues in codon optimization
• GENCODE data format glossary
• PTR ratios data source

Set up environment

Create virtual environment with mamba/conda

mamba env create -f environment_files/environment_linux_py3.10.yml
mamba activate master-env

Data Structure

Create a project folder for training (data) and model data (runs), and specify project path in src/utils/utils.py in the function set_project_path().

root
├── project folder
│   ├── data
│   ├── runs
│   │   ├── dev
│   │   │   ├── logs
│   │   │   ├── weights
│   │   │   ├── weights_best
│   │   ├── lstm
│   │   │   ├── ...
│   │   ├── xlstm
│   │   │   ├── ...
│   │   ├── gru
│   │   │   ├── ...
│   │   ├── transformer
│   │   │   ├── ...
│   │   ├── best_model (PTRNet)
│   │   │   ├── ...
└── master-thesis  (this repo)

📦 Repository Overview

This repository supports a deep learning benchmark study for predicting protein-to-mRNA (PTR) ratios using mRNA sequence and structure features.

🔧 Configuration

• src/config/: YAML configs for model architecture, training, and hyperparameter tuning (e.g., for Mamba, LSTM, Transformer, etc.).

📊 Data Handling

• src/data_handling/: Scripts for preprocessing, structure prediction, codon/nucleotide dataset creation, and stratified splitting.

🧠 Models

• src/models/: Implementation of deep learning models (MLP, CNN, RNNs, Transformer, xLSTM, Mamba, LegNet, PTRnet).
• Modularized by model type with shared predictor logic.

🎓 Pretraining

• src/pretraining/: Tools for masked language model (MLM) pretraining and motif discovery.

🏋️ Training

• src/training/: Training logic, early stopping, learning rate scheduling, and Optuna-based tuning.

📈 Evaluation

• src/evaluation/: Model evaluation, metrics, predictions, and plotting utilities.

🛠️ Utilities

• src/utils/, src/log/: Helper functions, logging setup, and device management.

🚀 Entry Point

• src/main.py: Main script for running training or tuning, configurable via CLI flags.
• src/multi_run*.sh: Example scripts to train multiple models sequentially.

Set up structure prediction

Install folding algorithms for secondary structure predictions. Follow arnie tutorial.

The bpRNA code for loop type predictions is already in the repo.

Command collection

Environment

Export environment dependencies

mamba env export -n master-env > environment_files/environment_linux_py3.10.yml

Update environment dependencies

mamba env update -n master-env -f environment_files/environment_linux_py3.10.yml

Use AIM logging

aim up

On how to search and filter runs in AIM: https://aimstack.readthedocs.io/en/latest/using/search.html

Use Optuna dashboard

From within the data folder for optuna, run:

optuna-dashboard sqlite:////path/to/optuna/model_name.db

start jupyter server manually

jupyter notebook --no-browser --port=8888

Other

Count files in a directory

ls -1 | wc -l