This package provides an interface to all of the Orion capabilities discussed in the manuscript entitled Deep generative AI models analyzing circulating orphan non-coding RNAs enable accurate detection of early-stage non-small cell lung cancer
This package and notebooks were run on Intel(R) Xeon(R) Platinum 8259CL CPU @ 2.50GHz machine with Ubuntu 22.04.4 LTS and the following python v3.10.12 packages. For installation of scvi-tools on M1-chip Mac machines, please visit: scvi-tools
Ubuntu 22.04.4 LTS
python 3.10.12
absl-py 1.0.0
numpy 1.26.4
numpyro 0.15.0
pandas 1.5.3
pytorch-lightning 2.2.5
scipy 1.10.0
scvi-tools 1.1.2
seaborn 0.12.2
shap 0.43.0
torch 2.0.1
torchmetrics 1.4.0.post0
tqdm 4.64.1
Instructions for installation with conda (runtime: 5 minutes):
conda create -n orion python=3.10.12
conda activate orion
conda install pytorch=2.0.1 numpy=1.26.4 pandas=1.5.3
pip install scvi-tools==1.1.2Please see the following notebook which shows a basic usage of Orion with simulated datasets (runtime: 4 minutes).
Essentially, Orion requires a dictionary of the data with the following keys:
"oncrna_ar": A numpy array of [samples, oncRNA] features count data
"oncrna_names": Name of `oncrna_ar` columns
"patient_names": Names of `oncrna_ar` rows
"onehot_ar": One-hot encoded class labels to train/predict
"smrnamat": A numpy array of [samples, small RNAs] features counts data for learning RNA content
"smrna_names": Name of `smarnamat` columns
"batch_list": A list of integers indicating batches to consider for triplet margin loss anchors
In addition, Orion requires a dictionary of model hyperparameters with some most relevant ones shown below:
"n_input": Number of oncRNAs
"n_input_lib": Number of small RNAs
"dp": Dropout
"loss_scalers": A list of scalers corresponding to NLL, KLD_Z, CE, and TML losses
"lr": Learning rate
"n_hidden": Number of hidden units
"num_lvs": Number of latent variables
"n_layers": Number of layers
"num_epochs": Number of epochs
"mini_batch": Number of samples in each mini batch
"tm_rounds": Number of rounds to sample anchors and compute TMLoss per mini batch
"num_classes": Number of classes
"weight_sample_loss": Weight to assign for each sample for classification task
"use_generative_sampling": Whether to use generative sampling for training the classifier
With these two dictionaries, Orion can be trained as:
trained_model_dict = train_orion_model(
data_dictionary,
train_idx, # Array of sample indices to be used for training
tune_idx, # Array of sample indices to be used for reporting metrics
select_features, # a subset or all of features in `onc_mat` and `oncrna_names` keys of data dictionary
dict_params=dict_params,
)The datasets required for running the orion-generate-predictions.ipynb are available on Zenodo.
Please use the GitHub issues for requesting assistance with the Orion package.