Skip to content

Model Zoo

Johnny Israeli edited this page Jan 22, 2017 · 10 revisions

Researchers develop DragoNN models for a wide range of genomics problems. To help share these models, we introduce the Model Zoo:

  • Standard format for packaging DragoNN model info.
  • Tools to upload and download trained DragoNN models.

To acquire a model:

  1. download the model gist by ./scripts/download_model_from_gist.sh <gist_id> <dirname> to load the model metadata, architecture, solver configuration, and so on. (<dirname> is optional and defaults to ./models).
  2. download the model weights by ./scripts/download_model_binary.py <model_dir> where <model_dir> is the gist directory from the first step.

Run dragonn predict with models in this zoo to obtain predictions (see example [here] (https://github.com/kundajelab/dragonn#15-seconds-to-your-first-dragonn-model)).

Models in the DragoNN manuscript

We provide trained models, data, and code in paper_supplement to reproduce results in the DragoNN manuscript. Data files in paper_supplement/simulation_data are named based on parameters of simulations functions and contain training, validation, and test data for each simulation in the manuscript. We provide architecture and weights files for 139 models discussed in the manuscript in paper_supplement/simulation_models. Architecture and weights files of models based on simulations are prefixed with <simulation_function_name>.<model_architecture_parameters>.

Predicting Regulatory Activities from Massively Parallel Reporter Assays

[MPRA DragoNN Model] (https://gist.github.com/jpaggi/33300cd4bb7e061b50bc25fd85b27d6f). Details can be found in the following [bioarxiv preprint] (http://biorxiv.org/content/early/2017/01/12/099879).

Clone this wiki locally