PredictMod Model Submission

← Go Back to the PredictMod Main Page.

Model submission guidelines

You've recently created a fantastic new AI/ML model. Congratulations! We will be happy to help you share it throughout the broader clinical practice and research communities. There are just a few more steps to follow before we get to that goal.

Required components

There are a few things we require in order to host any new models in our ecosystem.

1. Dependencies.

We will host your model in it's very own docker container, isolated from all of its noisy neighbors. In order to facilitate that, we need to know everything about the software your model runs, including major software and any package specifications that come along with it. For example, we need to know not just python and scikit-learn, but specifically major and minor revisions, e.g. python 3.10 or scikit-learn==1.2.3.

2. Inputs, the model, an example of the model running, and expected outputs

When we implement your model, we all need to know if it is working as expected or if there has been an unexpected "update" that prevents it from working. So, we need to have:

1. One or more example input(s): Examples where the expected outcome is known a priori to validate that your model is working as expected

2. A pickle, or other model-persistence object, that allows models to be loaded directly from disk to memory.

2.1 - In absence of a pickle or similar, we require the appropriate training data and training script(s) for your model so that we can train & persist the model in situ.

3. An example: An end-to-end, concrete instance of working code that loads an input, loads the model, performs the computation, and generates the prediction.

4. Outputs; the list of expected outcomes from the provided input data point(s).

3. A complete BioCompute Object (BCO) describing the model provenance, data, and other pertinent information

BCOs are an IEEE standardized means of tracking the provenance of Bioinformatics Analyses (cf the paper introducing BCOs, doi: 10.5731/pdajpst.2016.006734). Though BCOs can be written manually, we provide a tool to help simplify creation at the BCO Portal. Broadly, PredictMod uses BCOs to ensure fidelity of the models to what users expect the correct output ought to be.

Correctly constructed BCOs for model submission to PredictMod minimally contain the identifying information about the data used to train a model, the process(es) used in training and/or optimizing the model, and expected outputs from the model. An ideal BCO contains the all provenance data, or links to it, and is sufficient to fully recreate the model from the data to reproduce the expected outputs.

Next steps

Once all you have all of the required components in hand, reach out to us at mazumder_lab@gwu.edu and we will be happy to begin onboarding your new model into our platform!

Questions?

Check out our User Guide or FAQs for reference material. Still have questions? Please email us at mazumder_lab@gwu.edu with questions you may have about the process.