This is the official repository for the preprint Improving deep models of protein-coding potential with a Fourier-transform architecture and machine translation task.
This repository began as a fork of OpenNMT-py, a library for neural machine translation built by Harvard NLP. We have modified and extended it for modeling biological translation.
Install bioseq2seq from source:
git clone https://github.com/josephvalencia/bioseq2seq.git
cd bioseq2seqThis codebase requires Python 3.10. To set up and activate virtual environment
python3 -m venv venv
source venv/bin/activate
All dependencies for single-model training and inference are included in requirements.txt. Use pip to install them then install the project locally.
python -m pip install -r requirements.txt
./install.sh
We provide pretrained PyTorch weights for our best model best_bioseq2seq-wt_LFN_mammalian_200-1200.pt and our testing set
mammalian_200-1200_test_RNA_nonredundant_80.fasta, as described in the manuscript. To obtain a prediction
python bioseq2seq/bin/translate.py --checkpoint best_bioseq2seq-wt_LFN_mammalian_200-1200.pt --input mammalian_200-1200_test_RNA_nonredundant_80.fasta --num_gpus 1
The number after coding_prob: is the predicted probability of being protein-coding. The next four lines are the top beam hypotheses in descending order, i.e. the top line is the official prediction, <PC> for protein-coding and <NC> for long noncoding. The beam scores are not directly interpretable as coding probabilities.
If reliable peptide predictions are desired for mRNAs, use best_bioseq2seq_CNN_mammalian_200-1200.pt or best_bioseq2seq_LFN_mammalian_200-1200.pt, which were trained on the unweighted bioseq2seq task. The usage of --max_decode_len shown below will allow for protein outputs of up to 400 aa. The output file format will be the same as above, but multiple <PC> beams are possible.
python bioseq2seq/bin/translate.py --checkpoint best_bioseq2seq_LFN_mammalian_200-1200.pt --input mammalian_200-1200_test_RNA_nonredundant_80.fasta --mode bioseq2seq --num_gpus 1 --beam_size 4 --n_best 4 --max_tokens 1200 --max_decode_len 400
For further usage options, see
python bioseq2seq/bin/translate.py --help
Detailed instructions and code for reproducing the paper results are given in experiments/. This includes:
- data preprocessing
- hyperparameter tuning
- model training
- inference and evaluation
- comparisons with alternative software for protein coding potential
- filter visualization
- attribution calculation and motif finding
- plots for paper figures