COMPASS v2

The COMPASS v2.0 release represents a major advance toward a foundation-model framework for immunotherapy response and survival prediction. In this version, COMPASS has been systematically benchmarked against leading transcriptomic predictors including ENLIGHT, EaSIeR, and IRNet, demonstrating superior generalization across sixteen immunotherapy cohorts. The unified evaluation pipeline integrates consistent preprocessing, model training, and performance metrics, providing a rigorous baseline for transcriptome-level precision oncology research.

Building on the original concept-bottleneck architecture, COMPASS v2.0 introduces a dual-Transformer design that couples the pre-trained transcriptomic encoder with the Clinical Transformer decoder. This configuration enables joint modeling of molecular and clinical representations for both binary drug-response and time-to-event survival outcomes. The resulting framework substantially improves patient-level risk stratification and interpretability, establishing COMPASS as a flexible platform for integrative cancer outcome prediction.

Significant effort has been invested in software quality and usability. The entire codebase has been reformatted to PEP 8 standards using Black and isort, with expanded documentation, clearer module APIs, and complete installation support through PyPI via
pip install immuno-compass. Core components such as PreTrainer, FineTuner, loadcompass, and CRMap have been modularized for easier customization and reproducible experimentation.

To facilitate reproducibility, we provide curated pretraining datasets and standardized preprocessing pipelines. The TCGA pretraining dataset (1,065-gene and 15,672-gene versions) and the ITRP fine-tuning dataset (1,133 immunotherapy patients across sixteen cohorts) are now available on Figshare with persistent DOIs. Full dataset preparation scripts are accessible through the companion repositories [TCGA_dataset_processing](https://github.com/mims-harvard/COMPASS-web/tree/main/TCGA_dataset_processing) and [mRNA_pipeline](https://github.com/mims-harvard/COMPASS-web/tree/main/mRNA_pipeline), ensuring that COMPASS v2.0 remains transparent, reproducible, and extendable for the broader research community.

COMPASS v1

Update README.md

Signed-off-by: Charleshen <[email protected]>