Deep Learning-Based, Multiclass Approach to Cancer Classification on Liquid Biopsy Data

Maksym A. Jopek, Krzysztof Pastuszak, Sebastian Cygert, Myron G. Best, Thomas Würdinger, Jacek Jassem, Anna J. Żaczek, Anna SupernatIEEE Journal of Translational Engineering in Health and Medicine 2024 · DOI

TL;DR

Researchers have developed a deep learning framework that classifies cancer types by analyzing RNA from blood platelets (liquid biopsies). By converting gene expression data into pathway-based "heatmaps" and processing them with optimized Convolutional Neural Networks (CNNs), the system achieved up to 90.5% accuracy in identifying cancer locations, offering a non-invasive path toward early, multi-organ cancer screening.

What Problem Does This Paper Address?

Early cancer detection is the "holy grail" of oncology, but traditional tissue biopsies are invasive, painful, and often impossible for certain tumor locations. While "liquid biopsies"—tests that look for cancer markers in blood—are a promising alternative, most current methods focus on a binary "cancer vs. no cancer" result.

The challenge lies in multiclass classification: identifying exactly where the cancer originated. Blood contains a wealth of information, specifically within "tumor-educated platelets" (TEPs). These platelets "pick up" RNA signatures from tumors, but the resulting data is high-dimensional and noisy. This paper aims to bridge the gap between complex genomic data and clinical utility using deep learning.

Methods

The authors propose a novel pipeline that treats genomic data like an image processing task:

  1. Gene-to-Image Transformation: Instead of feeding raw gene counts into a model, they organized gene expression data according to biological pathways (using the KEGG database). This created 345 x 243 "heatmap" images where each pixel represents a gene's expression level.
  2. CNN Architectures: They tested several heavy-hitters in computer vision, including ResNet, DenseNet, and EfficientNet.
  3. The "RN9-implat" Model: Surprisingly, they found that shallower architectures outperformed deeper ones. Their final model, RN9-implat, is a modified ResNet9 that discards "MaxPooling" layers. The authors hypothesized that standard pooling (which downsamples images) might discard vital, fine-grained gene information.
  4. Explainable AI (XAI): To move away from the "black box" nature of AI, they used SHAP (SHapley Additive exPlanations) to identify which specific genes were driving the model's decisions, ensuring the AI was picking up on biologically relevant markers.

Datasets

The study utilized several publicly available RNA-seq datasets (GSE156902, GSE158508, etc.) covering:

  • Asymptomatic Donors (AD): Healthy controls.
  • Cancer Types: Ovarian (OC), Endometrial (EC), Non-small cell lung cancer (NSCLC), Glioblastoma (GBM), Sarcoma (SARC), and Brain Metastasis (BM).
  • Grouping Strategy: They evaluated both a 6-class "cancer-type-specific" dataset and a "location-specific" dataset where similar cancers (like OC and EC) were grouped together.

Key Insights and Ideas

  • Less is More in Architecture: In many AI fields, "deeper is better." Here, the authors found that complex backbones like DenseNet-169 did not improve performance. The shallower ResNet9 was more efficient and accurate for this specific biological data.
  • Preserving Spatial Information: By removing pooling layers, the model retained the exact "coordinates" of gene expression, suggesting that in genomic heatmaps, every pixel matters.
  • Biological Validation: The XAI analysis identified 60 high-impact genes (such as DIAPH1, ACER2, and MLLT4). Many of these are already recognized in the Human Protein Atlas as prognostic markers, which validates that the model is learning real biology rather than just statistical noise.
  • The "Location" Hurdle: The model occasionally confused brain metastasis with glioblastoma. This makes sense biologically, as both involve the brain environment, highlighting that while TEPs are powerful, organ-specific signatures can overlap.

Key Results

  • Cancer vs. Healthy: The model was nearly perfect at distinguishing healthy donors from gynecological cancers (99% accuracy).
  • Multiclass Accuracy:
    • 66.51% balanced accuracy when trying to distinguish between all 6 specific cancer sites.
  • 90.50% balanced accuracy on the location-specific dataset (grouping similar origins).
  • Top-2 Accuracy: The model reached 89.76%, meaning that in nearly 9 out of 10 cases, the correct cancer type was among the model's top two guesses.

Limitations and Future Work

The primary bottleneck is data availability. While liquid biopsies are growing in popularity, public datasets for rare cancers are scarce. Deep learning models are "data-hungry," and small sample sizes for certain cancers can lead to overfitting.

Future work will likely focus on:

  • Integrating Multi-omics: Combining platelet RNA with other markers like cell-free DNA (ctDNA) to increase sensitivity.
  • Prospective Clinical Trials: Moving from retrospective public data to real-time patient monitoring.
  • Refining the "Black Box": Further developing XAI tools to help clinicians trust and act on AI-driven diagnoses.

Summary updated

Abstract

<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Objective</i> : The field of cancer diagnostics has been revolutionized by liquid biopsies, which offer a bridge between laboratory research and clinical settings. These tests are less invasive than traditional biopsies and more convenient than routine imaging methods. Liquid biopsies allow studying of tumor-derived markers in bodily fluids, enabling the development of more precise cancer diagnostic tests for screening, disease monitoring, and therapy personalization. This study presents a multiclass approach based on deep learning to analyze and classify diseases based on blood platelet RNA. Its primary objective is to enhance cancer-type diagnosis in clinical settings by leveraging the power of deep learning combined with high-throughput sequencing of liquid biopsy. Ultimately, the study demonstrates the potential of this approach to accurately identify the patient’s type of cancer. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Methods</i> : The developed method classifies patients using heatmap images, generated based on gene expression arranged according to the Kyoto Encyclopedia of Genes and Genomes pathways. The images represent samples of patients with ovarian cancer, endometrial cancer, glioblastoma, non-small cell lung cancer, and sarcoma, as well as cancer patients with brain metastasis. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Results</i> : Our deep learning-based models reached 66.51% balanced accuracy when distinguishing between those 6 sites of cancer origin and 90.5% balanced accuracy on a location-specific dataset where cancer types from close locations were grouped. The developed models were further investigated with an explainable artificial intelligence-based approach (XAI) -SHAP. They returned a set of 60 genes with the highest impact on the models’ decision-making process. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Conclusions</i> : Our results show that deep-learning methods are a promising opportunity for cancer detection and could support clinicians’ decision-making process in finding the solution for the black-box problem.
BibTeX
@article{jopek2024deep,
  author = {Maksym A. Jopek and Krzysztof Pastuszak and Sebastian Cygert and Myron G. Best and Thomas Würdinger and Jacek Jassem and Anna J. Żaczek and Anna Supernat},
  title = {{Deep Learning-Based, Multiclass Approach to Cancer Classification on Liquid Biopsy Data}},
  journal = {IEEE Journal of Translational Engineering in Health and Medicine},
  year = {2024},
  doi = {10.1109/jtehm.2024.3360865},
}
Figure