Madrigal: A Unified Multimodal Model for Predicting Drug Combination Effects

👀 Overview of Madrigal

Madrigal is a unified multimodal AI model for predicting drug combinations across clinical outcomes and compounds, including approved drugs and novel compounds in development.

🚀 Installation

1. First, clone this Github repository and install following the section below
2. Set up data directories and create a .env file (see below).
3. Download datasets from our data repo in Harvard Dataverse and reorganize according to your .env setup.
4. [Optional] Download pretrained checkpoints from our checkpoint repo in Huggingface and reorganize according to your .env setup.

🛠️ Training and Testing

We provide sample model pretraining (second-stage modality alignment) and training scripts in scripts/. Specifically, the second-stage pretraining scripts are provided in ./scripts/cl_pretrain/, and the DDI fine-tuning scripts are provided in ./scripts/ddi_finetune/. The scripts will need to be adapted to your machine.

The first-stage modality adaptation training scripts or notebooks can be found in modality_pretraining/. You can also run inference with model checkpoints using sample Jupyter notebook (to be uploaded).

🌟 Personalize based on your own dataset

TODO

More details

Installing madrigal

Before installing madrigal, please set up a new conda environment through mamba env create -f env_no_build.yaml (we recommend mamba instead of conda). Then activate this environment with mamba activate primekg.

To install a global reference to madrigal package in your interpreter (e.g. from madrigal.X import Y), run the following:

cd /path/to/Madrigal # Path to main repo directory
python -m pip install -e .

Note that the current code and notebooks still use novelddi as reference (from novelddi.X import Y). To be able to use the scripts and notebooks directly, please make sure to edit setup.py and change the line name='madrigal' to name='novelddi'. Then, test the install by trying to import madrigal (or novelddi) in the interpreter:

python -c "import madrigal; print('Imported')" 

Now you should be able to use import madrigal (or import novelddi) from anywhere on your system, as long as you use the same python interpreter (recommend keeping to one conda env).

Setting up data and checkpoint directories

We use dotenv to ensure transferrability of code between platforms. Please add a file .env to the project directory (root of this project) and specify the following paths:

PROJECT_DIR=<project_dir>
BASE_DIR=<root_data_dir>
DATA_DIR=<processed_data_dir>
ENCODER_CKPT_DIR=<encoder_checkpoints_dir>
CL_CKPT_DIR=<contrastive_pretraining_checkpoints_dir>

For example, in our system, the paths look like:

PROJECT_DIR="/home/<your_hms_id>/workspace/DDI/NovelDDI/"
BASE_DIR="/n/data1/hms/dbmi/zitnik/lab/users/<user>/DDI/"
DATA_DIR="/n/data1/hms/dbmi/zitnik/lab/users/<user>/DDI/processed_data/"
ENCODER_CKPT_DIR="/n/data1/hms/dbmi/zitnik/lab/users/<user>/NovelDDI/pretraining/"
CL_CKPT_DIR="/n/data1/hms/dbmi/zitnik/lab/users/<user>/DDI/model_output/pretrain/"

Currently, hard-coded paths to embedding checkpoints exist in the get_str_encoder, get_kg_encoder, get_cv_encoder, and get_tx_encoder functions in ./novelddi/model/models.py. Corresponding modality pretrained checkpoints are provided in ./pretrained_checkpoints/first_stage/. Please edit the paths to load checkpoints from your desired storage locations.

⚖️ License

The code in this package is licensed under the MIT License.

Known issues

1. The torchdrug module needs to be imported after importing torch_geometric modules.
2. Newer version than torchdrug=0.2.0.post1 is required, as earlier versions cause an issue in LR scheduler.