CALM discovers interpretable ROI–pathway associations from completely unpaired data. Linear transforms map imaging and genetics from disjoint cohorts into a shared latent space, aligned by matching their class-conditional distributions.

pip install torch numpy pandas scikit-learn monai nibabel nilearn matplotlib seaborn scipy

Dataset paths are cluster-specific placeholders. Point them to your own data first:

• Dataset / checkpoint roots → code/utils/const.py
• Stage-1 output and checkpoint paths → STAGE1_OUT / *_CKPT_TMPL in job_scripts/

Then run the two-stage procedure in order:

# Stage 1 — pretrain the encoders
bash job_scripts/stage1_imaging.sh
bash job_scripts/stage1_genetics.sh

# Stage 2 — train the linear projections (alignment)
bash job_scripts/stage2_alignment.sh

Alignment is driven by two losses. A class-conditional MMD (L_cmmd) matches the imaging and genetics latent distributions within each diagnostic group, aligning the modalities without paired samples. A supervised contrastive loss (L_con) then pulls same-class samples together across modalities and pushes different classes apart, keeping the diagnostic groups separable — with an orthogonality regularizer (L_orth) preventing degenerate projections.

If any of the results in this paper or code are useful for your research, please cite the corresponding paper:

@inproceedings{wang2026calm,
  title={CALM: Interpretable Cross-Modal Alignment for Biomarker Discovery from Unpaired Data},
  author={Wang, Jueqi and Jacokes, Zachary and Van Horn, John Darrell and Pelphrey, Kevin A. and Schatz, Michael C. and Venkataraman, Archana},
  booktitle={Medical Image Computing and Computer-Assisted Intervention -- MICCAI 2026},
  year={2026},
  publisher={Springer}
}