tRNA-based classification model

The dataset contains:

1. Generic files used for training the dataset
2. Supplementary data used for labeling
3. An HTML file with a step-by-step description of the research
4. Python scripts used to train the models
5. The two best models were selected based on the lowest number of false negatives (FNs) on a third, independent test dataset.

Setup

Download Miniconda and use:

conda env create -f environment.yml

to replicate the working environment.

If any packages are missing during python code execution, install them manually using pip, based on import error messages.

Steps for replication:

1. Download supplementary data from https://doi.org/10.7554/eLife.71402
2. ftp_urls.txt contains a list of genome download addresses (most of them are available).
3. Run full.sh to download genomes and extract features for model training from full dataset, saved as FEATURES_ALL.ndjson (genomes are removed to preserve memory)
4. Run 80_20_split_fixed.py on FEATURES_ALL.ndjson together with both supplementary files to perform an automatic stratified 80/20 split, with archaeal and contaminated genomes filtered out.
5. Run Mass_models.py on FEATURES_ALL.ndjson, Supp1.csv, Supp2.xlsx
6. Run predict_dir.py to generate predictions for all trained models on FASTA genomes. If files provided, annotate predictions with ground truth from the TSV file, and report metrics separately for Isolate and MAG genomes.

Example run settings (All resutls were obtained using seed=42):

python3 80_20_split_fixed.py
  --ndjson FEATURE_ALL.ndjson
  --supp1 Supp1.csv
  --supp2 Supp2.xlsx
  --outdir split_dataset

python3 Mass_models.py   
  --ndjson split_dataset/subset01/   
  --supp2 Supp2.xlsx   
  --supp1 Supp1.csv   
  --outdir .   
  --train_mode both   
  --weight_mode both   
  --model all   
  --metric all   
  --n_trials 30   
  --timeout 5400

python3 predict_models_dir.py \
  --genomes_dir /path/to/fasta_dir \
  --models_dir results_models \
  --outdir predictions

Code and files will be modified and further developed in a packaged container after all required tests and training are completed.