GeezerVison

There is a recent explosion of deep learning algorithms that to tackle the computational problem of predicting drug treatment outcome from baseline molecular measurements. To support this,we have built a benchmark dataset that harmonizes diverse datasets to better assess algorithm performance.

This package collects diverse sets of paired molecular datasets with corresponding drug sensitivity data. All data here is reprocessed and standardized so it can be easily used as a benchmark dataset for the This repository leverages existing datasets to collect the data required for deep learning model development. Since each deep learning model requires distinct data capabilities, the goal of this repository is to collect and format all data into a schema that can be leveraged for existing models.

The goal of this repository is two-fold: First, it aims to collate and standardize the data for the broader community. This requires running a series of scripts to build and append to a standardized data model. Second, it has a series of scripts that pull from the data model to create model-specific data files that can be run by the data infrastructure.

For the access to the latest version of CoderData, please visit our documentation site which provides access to Figshare and instructions for using the Python package to download the data.

All coderdata files are in text format - either comma delimited or tab delimited (depending on data type). Each dataset can be evaluated individually according to the CoderData schema that is maintained in LinkML and can be udpated via a commit to the repository. For more details, please see the schema description.

The build process can be found in our coderbuild directory. Here you can follow the instructions to build your own local copy of the data on your machine.

We have standardized the build (coderbuild) process so an additional dataset can be built locally or as part of the next version of coder. Here are the steps to follow:

1. First visit the coderbuild directory and ensure you can build a local copy of CoderData.

2. Checkout this repository and create a subdirectory of the coderbuild directory with your own build files.

3. Develop your scripts to build the data files according to our LinkML Schema. This will require collecting the following metadata:

• entrez gene identifiers (or you can use the genes.csv file
• sample information such as species and model system type
• drug name that can be searched on PubChem

You can validate each file by using the linkML validator together with our schema file.

You can use the following scripts as part of your build process:

• coderbuild/utils/fit_curve.py: This script takes dose-response data and generates the dose-response statistics required by CoderData/
• coderbuild/utils/pubchem_retrieval.py: This script retreives structure and drug synonym information required to populate the Drug table.

4. Wrap your scripts in standard shell scripts with the following names and arguments:

5. Put the Docker container file inside the Docker directory with the name Dockerfile.[datasetname].

6. Run build_all.py from the root directory, which should now add in your Dockerfile in the mix and call the scripts in your Docker container to build the files.