Prodiff (CANDiT)

Companion code for our Cell Reports Medicine 2025 paper introducing CANDiT, a machine learning framework for AI-guided differentiation therapy in colorectal cancer.

Overview

This repository contains the code and derived data underlying our study introducing CANDiT (Cancer-Associated Nodes for Differentiation Targeting), a machine learning framework that identifies transcriptomic vulnerabilities for differentiation therapy in colorectal cancer (CRC).

Differentiation therapy has transformed the management of some hematologic malignancies (most notably ATRA in acute promyelocytic leukemia) but has never translated to solid tumors, largely because of the intra and inter tumoral heterogeneity that obscures cancer stem cells (CSCs). CANDiT addresses this gap by:

• Centering on CDX2, a master intestinal lineage transcription factor and tumor suppressor that is lost in high risk, poorly differentiated CRCs
• Building a transcriptomic network anchored on CDX2 to nominate therapeutic strategies that reinstate its expression
• Prioritizing PRKAB1, a stress polarity sensor and regulatory subunit of AMPK, as a top druggable node
• Validating a clinical grade PRKAB1 agonist across three complementary models: CRC cell lines, xenografts in mice, and a prospective cohort of patient derived organoids (PDOs)

Across all three platforms, PRKAB1 agonism reactivates lineage commitment, dismantles Wnt / YAP driven stemness programs, and selectively eliminates CDX2 low cancer stem cells while sparing normal cells. In xenografts, this translated to a 68 percent reduction in tumor volume and elimination of mortality (hazard ratio 0.09). A 50 gene response signature derived from integrated modeling across all platforms predicts approximately a 50 percent reduction in recurrence and mortality risk.

Repository contents

Analysis notebooks

Notebook	Purpose
PRODIFF_Fig4_A_G.ipynb	Panels A through G of Figure 4 (in vivo and PDO efficacy)
PRODIFF_Heatmap.ipynb	Heatmap visualizations of the 50 gene response signature
corr_plot_Fig3.ipynb	Correlation analyses supporting Figure 3
multivariate analysis_Fig3E.ipynb	Multivariate model for Figure 3E
univariate analysis_Fig3F_5E.ipynb	Univariate response analyses for Figures 3F and 5E
prodiff_ROC_sig.ipynb	ROC and AUC analyses of the response signature

CANDiT framework

• CANDiT/ — the CANDiT machine learning framework implementation

Data files

File	Content
HCT114.txt, SW480.txt	Transcriptomic profiles from CRC cell lines
organoid.txt	Patient derived organoid (PDO) response data
invivo.txt	Xenograft in vivo response data
IC50_heatmap.txt	Dose response IC50 measurements
ROC_PDO.txt	ROC input for PDO response classification
cell_xeno_deg.txt	Differentially expressed genes across cell line and xenograft models
pgsig-res-1.txt, pgsig-res-2.txt	Signature scoring results
node-2(1).txt, node-3(1).txt	Network node outputs
AM_MV3.txt	Multivariate model inputs
Gene signatures used to predict Rx synergy.xlsx	Curated signatures for drug response prediction
Signatures of PDO and PDX response in CRC.xlsx	Curated PDO and PDX response signatures

Utilities

File	Purpose
bone.py	Boolean Network Explorer (BoNE) helper functions
StepMiner.py	StepMiner tool for identifying step patterns in gene expression
HegemonUtil.py	Python utility for the Hegemon Boolean analysis framework
hegemonutils.pl	Perl companion utilities for Hegemon
explore.conf	Configuration file for the Hegemon exploration workflow

Reproducing the analysis

1. Clone the repository and install dependencies (Python 3.9+ with the scientific stack, plus lifelines for survival analyses, scikit-learn for ROC/AUC, and Perl for the Hegemon utility scripts).
2. Explore the CANDiT framework in CANDiT/ to build or apply the network prioritization model.
3. Run the figure notebooks in numerical order (Fig3 → Fig4 → Fig5) to reproduce panels from the paper.
4. Use prodiff_ROC_sig.ipynb to evaluate signature performance in your own cohorts.

Publication

Sinha S, Alcantara J, Perry K, Castillo V, Ondersma AK, Banerjee S, McLaren E, Espinoza CR, Taheri S, Vidales E, Tindle C, Adel A, Amirfakhri S, Sawires JR, Yang J, Bouvet M, Ghosh P. CANDiT: A machine learning framework for differentiation therapy in colorectal cancer. Cell Reports Medicine, 2025. doi: 10.1016/j.xcrm.2025.102421 PMCID: PMC12711679

If you use CANDiT, the 50 gene response signature, or any derived analyses in your own work, please cite the manuscript.

Press coverage

• Newswise: Precision reprogramming: How AI tricks cancer's toughest cells
• EurekAlert!: Precision reprogramming: How AI tricks cancer's toughest cells
• Mirage News: AI reprograms cancer cells with precision

Contact

Saptarshi Sinha, Ph.D. Assistant Project Scientist, Department of Cellular and Molecular Medicine Interim Director, Center for Precision Computational Systems Network (PreCSN) University of California San Diego Email: sasinha@health.ucsd.edu

License

Released under the MIT License. See LICENSE.