Multimodal Integration for Heart Transplant Rejection


Chunqing (Tony) Liang

PhD student in Bioinformatics

Supervisor: Dr. Amrit Singh

Oct 6, 2025

Land acknowledgement

I would like to acknowledge that I work on the traditional, ancestral, and unceded territory of the Coast Salish Peoples, including the territories of the xwməθkwəy̓əm (Musqueam), Skwxwú7mesh (Squamish), Stó:lō and Səl̓ílwətaʔ/Selilwitulh (Tsleil- Waututh) Nations.

Traditional: Traditionally used and/or occupied by Musqueam people

Ancestral: Recognizes land that is handed down from generation to generation

Unceded: Refers to land that was not turned over to the Crown (government) by a treaty or other agreement

Heart Transplant

  • Replace person’sick heart with healthy one from someone who has died

  • Occurs when heart failure is end-stage and no longer treatable in any other way

Heart Transplant Rejection

Types of heart transplant rejection (1)
  • Relies on endomyocardial biopsy (EMB)
  • Assessments from pathologists, subjective and opinion-based
  • Multiple rejection types
    • Acute cellular rejection (ACR)
    • Antibody-mediated rejection
    • Quilty lesion
    • Look similary structurally

Quilty and ACR

Quilty lesion and acute cellular rejection. (A) Quilty lesion. (B) Grade 1R. (C) Grade 2R. (D) Grade 3R. (2)

How to improve diagnosis?

  • What’s better way to assist pathologists to differentiate different rejections
    • What to do if there’s disagreement?
  • Computational pathology comes in handy!

Workflow in computational pathology (3)

In-house data

  • Data from:
    • Prevention of Organ Failure (PROOF) centre
    • Bruce McManus Cardiovascular Biobank (BMCB)
  • Cohort of \(> 3000\) samples from \(500\) patients
  • Composed of different types of no rejection, ACR, AMR, mixed rejection

Available variable in the cohort

Proof Centre

  • Lead by Dr. Scott Tebutt, Mr. Casey Shannon, Ms. Sarah Assadian
  • Majority of biopsies of the cohort is from here
  • Collected data from multiple participating sites like Vancouver, Ottawa and Nebraska
  • Multiple timepoints biopsies from patients to from longitudinal study

Bruce McManus Cardiovascular Biobank

Dr. Ying Wang Dr. Chi Lai


  • Lead by Dr. Ying Wang, Dr. Chi Lai, Dr. Singhera, Ms. Coco Ng, Ms. Tiffany Chang
  • Several hundred more local heart transplant patients from here
  • Currently restoring slides from storage and digitizing them

Objective

  • Improve diagnosis using multimodal data

  • Can we tell long term outcome of rejection?

    • i.e. Cardiac allograft vasculopathy (CAV)
    • CAV is 1 of top 3 causes of death in the first year after heart transplant and increases in prevalence over time (4)

Improve diagnosis using mulitmodal data

Can we tell long term outcome of rejection

  • Rejection have \(\geq 1R\) is more likely to have CAV
  • People that have CAV have similar characteristics with others

Types of available methods

Types of multimodal integration method (5)


Question arises

  • Many integration methods

    —> Which to use, how to choose them?

  • Reproducibility crisis

    —> How to reproduce method and get reliable results?

  • Existing benchmark studies are not 100% complete or all-encompassing

    —> Technical difficulty in implementation?

MESSI pipeline

  • We created a workflow pipeline Multimodal Experiments with SyStematic Interrogation using nextflow
  • Solves reproducibility issue through:
    • Nextflow (6), workflow managment
    • Singularity (7), software environment control
  • Looking for collaborations

MESSI pipeline

Standardized data preprocessing

  • Currently handles both R and Python –> could extend to more
  • Data flows in there , handles N different datasets
  • Important model selection prior to evaluation

Flexible method evaluation

  • Each method is an isolated workflow (dash box)
  • Each method internally runs evaluation in parallel
  • Each result is saved separately …
  • Easily extended to other tasks not just CV
  • On / Off to run interested methods only

Summarized reports of metrics performances

  • “Recursively” collects output from each method including versions
  • Summarizes metrics for downstream analysis
  • Provides rich report of computational resource usages

Datasets evaluated

Computational Resources Usage

  • MOGONET and multiview require longer durations compared to rest
  • MOGONET takes more virtual memory due to its deep learning nature
    • Transferring of data from GPU to CPU
  • Model training and prediction steps demand more memory consumption

Performance on real-world datasets

  • No method perfrom well in all scenarios
  • DIABLO performs at top compared to others
    • The null design variant beats the full design
  • MOGONET performs the weakest

Biological Interpretation on biomarkers identified

MESSI future direction

  • “No method works universally well on all datasets”
  • Classic statistical methods still work, and sometimes even better than deep learning (DL)
  • Pipeline proves way to reproducibly explore, benchmark different aspects of integration methods
    • Resumable
    • Parallel to compute as many resources as allowed at the same time
    • Ease burden of setting up environment
  • Need to add more methods and datasets
    • Especially DL models are more popular now
    • Explore if any dataset could have relation to another despite different disease/condition

Current work on GeoMX

  • TODO

Background of GeoMX

  • TODO

GeoMxTools PCA

GeoMxTools Volcano plot

GeoMxTools Significant genes heatmap

Next steps

  • TODO

Conclusion

  • TODO

Discussion & Future directions

  • “No method works universally well on all datasets”
  • Classic statistical methods still work, and sometimes even better than deep learning (DL)
  • Pipeline proves way to reproducibly explore, benchmark different aspects of integration methods
    • Resumable
    • Parallel to compute as many resources as allowed at the same time
    • Ease burden of setting up environment
  • Need to add more methods and datasets
    • Especially DL models are more popular now
    • Explore if any dataset could have relation to another despite different disease/condition

Thanks!

Acknowledgements

  • Dr. Amrit Singh
  • Dr. Maryam Ahmadzadeh
  • Dr. Young Woong Kim
  • Rishika Daswani
  • Roy He
  • Samuel Leung
  • Raam Sivakumar
  • Jeffrey Tang
  • Michael Yoon
  • Mingming Zhang

Reference

1.
Lipkova J, Chen TY, Lu MY, Chen RJ, Shady M, Williams M, et al. Deep learning-enabled assessment of cardiac allograft rejection from endomyocardial biopsies. Nature medicine. 2022;28(3):575–82.
2.
Cho H, Choi JO, Jeon ES, Kim JS. Quilty lesions in the endomyocardial biopsies after heart transplantation. Journal of pathology and translational medicine. 2019;53(1):50–6.
3.
Wagner SJ, Matek C, Shetab Boushehri S, Boxberg M, Lamm L, Sadafi A, et al. Make deep learning algorithms in computational pathology more reproducible and reusable. Nature Medicine. 2022;28(9):1744–6.
4.
Shetty M, Chowdhury YS. Heart transplantation allograft vasculopathy. 2020;
5.
Subramanian I, Verma S, Kumar S, Jere A, Anamika K. Multi-omics data integration, interpretation, and its application. Bioinformatics and biology insights. 2020;14:1177932219899051.
6.
Di Tommaso P, Chatzou M, Floden EW, Barja PP, Palumbo E, Notredame C. Nextflow enables reproducible computational workflows. Nature biotechnology. 2017;35(4):316–9.
7.
Kurtzer GM, Sochat V, Bauer MW. Singularity: Scientific containers for mobility of compute. PloS one. 2017;12(5):e0177459.