OperonHunter

• The directory images_ecoli/ contains images used to train a neural network and test its performance on the E. coli genome.
• The directory images_bsubtilis/ contains images used to train a neural network and test its performance on the B. subtilis genome.
• The directory genomes/ contains the genome files used in our method.
• The directory strings/ contains string scores used in our method.

Neural Network Training and Testing

We use the Fastai library to train our model. The script train.py demonstrates an example training a ResNet18 on the images_ecoli/ dataset. It saves the best model observed during training.

The script test.py loads one of the previously saved models and tests it on images found in the test/ directory under the images_ecoli/ directory. It outputs two files: operon_pegs.txt and noperon_pegs.txt representing operon pairs and non-operon pairs respectively. Each line in the output can be thought of as a gene pair comprising the listed gene peg and its immediate successor in that genome.

One pre-trained model is provided in the models/ directory under the images_ecoli/ directory.

Hyper-parameter fine tuning was performed manually.

Image Generation

The following steps were performed to generate the images:

1. Call the Compare Region Viewer service provided by PATRIC. An example command is: curl --max-time 300 --data-binary '{"method": "SEED.compare_regions_for_peg", "params": ["$peg", 5000, 20, "pgfam", "representative+reference"], "id": 1}' https://p3.theseed.org/services/compare\_region" Where $peg is the query gene of interest. Repeat this call for all genes/pegs of interest and place all the output jsons in one folder (let's call it input_jsons).
2. Run the program JsonToCoordinates.py with input_jsons as input to parse the JSON files into a different format to be used by the image generating software. The resulting file will be oc.txt, which is the input to CoordsToJpg.java. This script needs to be in the same directory as genomes/ and strings/
3. Compile and run the Java program CoordsToJpg.java which will convert the coordinate file into images.
4. Split your images into the appropriate classes. An example folder has been provided (images_ecoli/) which contains a training and testing directory and the images that can be used to train a model and test it on the E. coli genome.

Results Reproducibility

In an attempt to make our results reproducible, we updated this repository with a script (reproducible.py) that eliminates the randomness from the experiments (we had not set the random seet of the get_transforms() function in FastAI earlier, and so we noticed some variance in our experiments, leading to results sometimes better and sometimes worse than those reported in our paper). The posted scripts produce results that are very close to those reported in Tables 1-5 in our paper. Specifically, the Aggregate Precision, Aggregate Specificity, Accuracy, F1 Score are still the same. The E. coli. Specificity is lower (0.944 vs 0.95, they differ by one example and the effects of rounding). The Aggregate Sensitivity is higher (0.93 vs 0.92), the MCC is higher (0.85 vs 0.84), while the percentage of exact operon matches is lower (0.83 vs 0.85). Note that in the posted datasets, the folder extra/ include images of examples belonging to the Operon class. Also note that when training the B. subtilis model we unfroze the learner so that earlier layers can update their weights as well. We found this to be more effective given the larger training datast available for this genome.

---

To cite this work, please use:

Assaf, R., Xia, F. & Stevens, R. Detecting operons in bacterial genomes via visual representation learning. Sci Rep 11, 2124 (2021). https://doi.org/10.1038/s41598-021-81169-9