Personalized Neoadjuvant Therapy Recommendations in Breast Cancer from an Explainable Multi-Omics Response Model

We developed and externally validated a multi-omics model integrating pre-NAT clinical data, DCE-MRI images, and medical reports to predict pathologic complete response (pCR) and likelihood of survival after NAT. The prognostic scores provided by the response model can select populations with relatively poor outcomes after treatment according to the factual regimen, which may provide a basis for personalized NAT regimen recommendations, potentially reducing inefficiency or overtreatment by moving beyond selection solely based on cancer stage and subtype.

Environment Setup

Start by installing conda environment, then clone this repository and install the dependencies.

conda create -n morm python=3.11
conda activate morm

pip install torch torchvision torchaudio

git clone https://github.com/fiy2W/MORM.git
cd morm

pip install -r requirements.txt

How to get started?

Prepare data

Multi-omics alignment pretraining

We use contrastive language-image pretraining (CLIP) to align MRI images and medical reports.

python src/train/pretrain_clip.py -c config/config.yaml -d cuda

Train

Train PoE model with five-fold cross validation.

python src/train/train_vae_poe.py -c config/config.yaml -d cuda -f 0
python src/train/train_vae_poe.py -c config/config.yaml -d cuda -f 1
python src/train/train_vae_poe.py -c config/config.yaml -d cuda -f 2
python src/train/train_vae_poe.py -c config/config.yaml -d cuda -f 3
python src/train/train_vae_poe.py -c config/config.yaml -d cuda -f 4

Test

Test model for pCR prediction and survival analysis.

python src/test/test_vae_poe_pcr.py -c config/config.yaml -d cuda
python src/test/test_vae_poe_followup.py -c config/config.yaml -d cuda

Acknowledgements

Contact

For any code-related problems or questions please open an issue or concat us by emails.

• [email protected] (Ritse Mann)
• [email protected] (Tao Tan)
• [email protected] (Luyi Han)