CanDLE:�Illuminating Biases in Transcriptomic Pan-Cancer Diagnosis

1

10/21/2023

Paper ID 8

2

10/21/2023

Cancer Prevalence and Diversity

19.2 M New Cases (2020)

9.9 M Deaths (2020)

• Cancer Diagnosis Problems:
• Many tissues
• Many cancer types
• Expertise level dependent
• Time consuming
• Proposed Solution:
• Use genetic expression vector to predict type of cancer or normal tissue.

�But Wait… �What Public Data Do We Have?

3

10/21/2023

4

10/21/2023

• Healthy TCGA counterpart
• Atlas of healthy tissue expression
• 54 non-diseased tissues
• Nearly 1,000 individuals
• More than 17,000 RNA-seq samples

​

The Cancer Genome Atlas

Genotype Tissue Expression Project

TCGA and GTEx Data Are NOT Comparable!!!

Reference Genome, Alignment and Quantification Algorithm Variability

5

10/21/2023

Vivian et al. Dataset

Wang et al. Dataset

• STAR Alignment and RSEM quantification
• 18,354 samples with 60,498 genes

​

6

10/21/2023

�So…�Problem Solved, Right?

7

10/21/2023

8

10/21/2023

Linear SVM

Both Datasets Present Empirical Translation Biases

Previous Results Lose Clinical Relevance

Z-score Batch Standardization Corrects Translation Biases

1

9

10/21/2023

Which cancer/tissue?

Classification &

all-vs-one Detection

63 Neurons for Multilabel Classification

2 Neurons for All-vs-one Detection

Simplest Gradient Based Approach

Multinomial Logistic Regression

Previous Findings for Representation Learning

[1] Smith, A., et al., 2020: Standard machine learning approaches outperform deep representation learning on phenotype prediction from transcriptomics data. DOI: 10.1186/s12859-020-3427-8

10

10/21/2023

Random 60/20/20% Train/Val/Test Partition

11

10/21/2023

*

*

CanDLE’s Simplicity Can Generalize Better With Removed Biases

*Reimplementation

State-of-the-Art Performance by +7.3% Balanced Accuracy in Test

[2] Hong, J., et al., 2022: A deep learning model to classify neoplastic state and tissue origin from transcriptomic data. DOI: 10.1038/s41598-022-13665-5

Hong’s Method Takes Advantage of Translation Biases

12

10/21/2023

Generally Worst Performing Classes Have Low Training Samples

Comparable Performance With Respect to the State-of-the-art Method by Quinn et al.

[3] Quinn, T., et al., 2019: Cancer as a Tissue Anomaly: Classifying Tumor Transcriptomes Based Only on Healthy Data. DOI: 10.3389/fgene.2019.00599

13

10/21/2023

Select Top 1,000 Genes in Absolut Value For Each Cancer Class

Order by Number of Times that a Gene Was Selected for a Class

1,982 Genes Important for at Least 3 Cancers Were in the Final List

Gene Ontology Biological Processes Enrichment Analysis

Interpretation

Developmental and Morphogenesis Pathways Were Over-Represented

• Batch effects appear even with standard quantification of RNA-seq.
• There is no need for a complex model in RNA-seq based cancer diagnosis.
• Machine learning cancer diagnosis is feasible and can offer valuable biological insights.

14

10/21/2023

Code Availability

https://github.com/g27182818/CanDLE

Thank You�for Your Time!�Questions?

15

10/21/2023