This project predicts the translational efficiency (TE) of genes from RNA sequence features using machine learning. It models log-transformed TE values under eIF3d and eIF4e depletion conditions with Ridge regression, ElasticNet, and LightGBM models. Features are extracted from transcript sequences and merged into a single dataset for training.
- Unzip
TE_eIF_depletion.zip - Run the
.pyfiles in main to extract sequence features from the raw data inTE_eIF_depletion.csv(this takes up hundreds of MBs) - (optional) Run
check.pyas a sanity check on all extracted.csvfiles - Run
merge.pyto get the entire input dataframe - Run your desired model in its directory
We did it this way because we each worked on feature extractions separately:
| Feature | Author |
|---|---|
| Length features | Jannes |
| Nucleotide frequency | Anuschka |
| Codon and aa frequency | Oliver |
| K-mer frequency | Jannes |
| Wobble position | Anuschka |
| One-hot encoding of nt around start codon | Jannes |
| Counts of 20 relevant dicodons according to Gamble et al | Jannes |
| Secondary structure values | Oliver |
Running LGBM.py will create a folder with directories for each of the defined feature subsets along with a general summary of the model results.
Inside the feature subset directories are two folders with the model results for the two depletion conditions (eIF3d and eIF4e). They contain .csv files with the search results across the folds (currently 5), feature importance scores, and predicted vs observed values on the test set. A model summary is also given as a .json file.
Currently the feature subsets are the ones described in the report.