Synonyme Codon Prediction

Project Medical Data Science - SS 2024

Insa Belter, Maximilian Kühn, Felix Mucha, Nils Rekus, Floris Wittner

Installation

Requirements

• Python 3.8 or higher (Python 3.12 is recommended). More information on how to install Python can be found here.
• pip Python package installer (usually included in Python installations)

Setup

1. Clone the repository
2. Create a virtual environment

   python -m venv .venv

3. Activate the virtual environment

   source .venv/bin/activate

4. Windows only: to use Cuda for NVIDIA GPU acceleration, install pytorch with the following command first (for more information see Pytorch Installation Guide)

   pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

5. Install all other requirements

   pip install -r requirements.txt

Project Content

Folder Structure

.
├── data
│   └── organism
│       └── cleanedData.pkl
├── ml_models
│   └── organism
│       ├── best_rnn_model.pt
│       ├── best_tcn_model.pt
│       └── best_encoder_model.pt
├── notebooks
│   ├── archive
│   └── *.ipynb
├── scripts
│   ├── archive
│   └── *.py
├── unit_tests
├── README.md
└── requirements.txt

Notebooks

• 00_Using_Classifiers
  • Notebook for applying the best classifiers from each model architecture to new data
• 01_Data_Aggregation_and_Preparation
  • Data aggregation: How was the data gathered?
  • Data preparation: How was the data cleaned and split for training and testing purposes
  • Secondary Protein Structure as possible additional feature
• 02_Data_Exploration
  • How many sequences do we have per organism?
  • How are the amino acids distributed?
  • What is the Codon Usage Bias?
• 03_Baseline_Classifiers
  • Classifiers for comparing the trained machine learning classifiers to the possible baseline resulting from the Codon Usage Bias (CUB)
• 04_Index_based_Analysis_and_Classifier
  • Index based analysis of amino acid sequences
  • Classifier based on results of this analysis (Index based CUB)
• 05_Other_Statistical_Analysis_Approaches
  • Correlation between neighbouring amino acids
  • Amino acid analysis based of chemical properties of codons
• 06: RNN
  • 06_1_RNN_Training
  • 06_2_RNN_Testing
• 07: TCNN
  • 07_1_TCN_Training
  • 07_2_TCN_Testing
• 08: Encoder-only Transformer
  • 08_1_Encoder_Training
  • 08_2_Encoder_Testing
• 09_Accuracy_Results_Overview
  • Training validation accuracies per model (RNN, Encoder, TCNN)
  • Accuracy per (best) Model per Organism in comparison to baseline
    • Index-based CUB vs Baseline Max CUB
    • RNN vs Baseline Max CUB
    • TCNN vs Baseline Max CUB
    • Encoder vs Baseline Max CUB
  • All accuracies in one diagram
    • Max CUB, Index-based Max CUB, RNN, TCNN, Transformer

Scripts

• Data Aggregation
  • data loading: Load Fasta files, check for corrupted sequences and save cleaned data
  • data splitting: Data splitting for training, testing and validation
  • codon usage bias: Calculation of Codon Usage Bias (CUB)
• ML helper: Helper functions for training the machine learning models
• ML evaluation: Evaluation functions for the machine learning models
• Files with classifier implementations for each model architecture
  • Classifier Class
  • Baseline
  • RNN
  • TCN
  • Encoder & Modified Pytorch Encoder (can output Attention weights)
  • Index Classifier
• Chemical Property: Chemical property analysis of codons
• change Table
• [WIP] secondary structure: Exploration of secondary structure as additional feature

Data

Contains the cleaned sequence data for 3 organisms (E.Coli, D.Melanogaster, H.Sapiens). The data is split in training, testing and validation data and then saved as a pickle file. Also all the splits are saved in a shuffled version.

The data folder also contains various files, which track the model training progress and the model performance.