PaperlayerView exact revision on arXiv

Source-linked AI summary

Machine Learning Methods for Cancer Classification Using Gene Expression Data: A Review

Fadi Alharbi, Aleksandar Vakanski

arXiv:2301.12222v1cs.LG

TL;DR

Cancer classification from gene expression data must address datasets with many genes but relatively few samples, alongside interpretability and multimodal-integration challenges. This review surveys conventional and deep learning methods, data technologies and datasets, and feature-engineering strategies, emphasizing newer neural architectures. It concludes that recent work spans RNA-Seq-focused applications and architectures including graph and transformer networks.

  • Problem

    Gene expression cancer-classification datasets typically have small sample sizes and many gene dimensions, while existing approaches also face interpretability and multimodal-integration limitations.

  • Method

    The paper reviews conventional and deep learning methods, gene-expression data collection, supervised datasets, and feature-engineering and preprocessing techniques for cancer classification.

  • Results

    The review covers recent neural architectures, including graph and transformer networks, and emphasizes RNA-Seq as the dominant data format in recent works.

  • Takeaways & Limitations

    The survey provides a comprehensive view of feature engineering, datasets, and traditional and deep learning applications for cancer classification using gene expression data.

  • Takeaways & Limitations

    Microarray measurements have limited accuracy, precision, and specificity and are sensitive to experimental conditions.

Abstract

from arXiv · show

Cancer is a term that denotes a group of diseases caused by abnormal growth of cells that can spread in different parts of the body. According to the World Health Organization (WHO), cancer is the second major cause of death after cardiovascular diseases. Gene expression can play a fundamental role in the early detection of cancer, as it is indicative of the biochemical processes in tissue and cells, as well as the genetic characteristics of an organism. Deoxyribonucleic Acid (DNA) microarrays and Ribonucleic Acid (RNA)- sequencing methods for gene expression data allow quantifying the expression levels of genes and produce valuable data for computational analysis. This study reviews recent progress in gene expression analysis for cancer classification using machine learning methods. Both conventional and deep learning-based approaches are reviewed, with an emphasis on the ap-plication of deep learning models due to their comparative advantages for identifying gene patterns that are distinctive for various types of cancers. Relevant works that employ the most commonly used deep neural network architectures are covered, including multi-layer perceptrons, convolutional, recurrent, graph, and transformer networks. This survey also presents an overview of the data collection methods for gene expression analysis and lists important datasets that are commonly used for supervised machine learning for this task. Furthermore, reviewed are pertinent techniques for feature engineering and data preprocessing that are typically used to handle the high dimensionality of gene expression data, caused by a large number of genes present in data samples. The paper concludes with a discussion of future research directions for machine learning-based gene expression analysis for cancer classification.

1. Introduction

Gene expression analysis supports cancer diagnosis, drug discovery, and investigation of gene roles, while machine learning addresses the high-dimensional, small-sample structure of these data. This review synthesizes conventional and deep learning methods, feature engineering, datasets, and newer neural architectures.

  • Gene expression changes can serve as markers for early cancer detection and provide insights into cancer initiation, progression, and drug-development targets.
  • Gene expression analysis measures transcripts to determine which genes are expressed and at what levels.
  • RNA-Seq offers greater specificity, resolution, sensitivity to differential expression, and dynamic range than DNA microarrays.
  • Machine learning methods have progressed from conventional models such as Decision Trees and Support Vector Machines toward deep learning approaches for gene expression analysis.
  • Feature engineering addresses gene-expression data with many dimensions and relatively few samples by selecting or transforming informative features.
  • The review covers MLP, CNN, RNN, GNN, and transformer architectures, including transfer learning to address small training datasets and high dimensionality.
  • The survey extends prior reviews by combining conventional and deep learning, feature engineering, datasets, and recent graph and transformer approaches.

2. Gene Expression Data

Gene expression data are collected through microarray and RNA-Seq technologies, with RNA-Seq enabling broad transcriptome profiling and analysis of transcript variation. The review also identifies accessible datasets used for cancer-classification research.

  • Gene expression analysis quantitatively examines transcripts in cells or tissues to estimate gene-expression levels.
  • Microarray Data: Microarrays measure expression levels for thousands of genes using DNA probes arranged as microscopic spots on chips or slides.
  • Microarray Data: Microarrays have limitations in accuracy, precision, and specificity, and their measurements are sensitive to hybridization temperature, sample purity, degradation, and amplification.
  • RNA-Seq Data: RNA-Seq can analyze coding and noncoding RNA, transcript variation, post-transcriptional modifications, alternative splicing, and polyadenylation.
  • RNA-Seq Data: RNA-Seq provides important advantages over microarray data across gene range, isoforms, resolution, background noise, cost, rare transcripts, and noncoding RNA.
  • RNA-Seq Data: RNA-Seq converts RNA into cDNA, prepares cDNA libraries, sequences them on an NGS platform, and estimates gene-expression levels from resulting sequences.
  • Gene Expression Datasets: The reviewed datasets include open-source, accessible RNA-Seq and microarray collections based on human tissue that are widely used for cancer classification.

3. Feature Engineering

Feature engineering transforms or selects gene-level variables to emphasize relevant information and reduce the dimensionality of RNA-Seq data. The review organizes these methods into filter, wrapper, and embedded categories, including hybrid strategies.

  • Feature engineering selects relevant or novel features, including marker genes, to improve data transformations and machine-learning analysis.
  • Feature selection in RNA-Seq data identifies genes representing dataset structure in a lower-dimensional space and increases the signal-to-noise ratio.
  • Filter Methods: Filter methods rank genes by relevance and can remove unlikely-to-contribute features before model training.
  • Wrapper Methods: Wrapper methods use a classifier as a black box to search for feature subsets that optimize learning performance.
  • Embedded Methods: Embedded methods incorporate classifier configuration while searching feature subsets, combining characteristics of filter and wrapper approaches.
  • Hybrid Methods: Hybrid approaches can apply filtering first to reduce features, then use a classifier-based wrapper stage for further refinement.
  • Hybrid Methods: Ensemble methods use multiple weak classifiers and have been shown to reduce overfitting and improve predictive performance in gene expression analysis.

4. Methods for Gene Expression Analysis

Machine learning methods have been used in gene expression analysis to identify potential cancers and provide insights into potential treatment options.

  • Machine learning methods identify potential cancers and provide insights into potential treatment options using gene expression analysis.

4.1 Traditional Machine Learning Methods

Traditional machine-learning methods have been applied to cancer classification using selected gene sets, with performance varying across cancer types. Their effectiveness depends strongly on feature quality and selection.

  • Conventional approaches include SVM, kNN, Naïve Bayes, and Random Forest methods for early cancer detection.
  • Feature-selection strategies identified predictive gene subsets for training classifiers, including 256 genes for clear cell sarcoma and 273 relevant genes in another RF-based approach.
  • 99.89% and 99.40% accuracy was achieved for two cancer types across five datasets, but performance decreased for other cancer types.
  • Traditional machine-learning performance depends highly on supplied feature quality and the effectiveness of accompanying feature-selection methods.
  • Table 6 lists prior traditional machine-learning methods for gene-expression analysis.

4.2 Deep Learning Methods

Deep learning methods for gene-expression cancer classification include MLP, RNN, CNN, GNN, and transformer architectures, each exploiting different data representations or dependencies. Reported studies demonstrate applications across cancer types, while model limitations depend on genomic structure, sample size, and computational requirements.

  • Overview: Deep learning models learn hierarchical representations from high-dimensional gene-expression data, reducing reliance on manually selected features.The reviewed architectures include MLP, CNN, RNN, GNN, and transformer neural networks.
  • Multi-Layer Perceptron (MLP) Neural Networks: MLP studies combined gene-expression, clinical, pathway, and miRNA information for cancer classification and survival prediction.Reported examples include 0.8163 AUC and 75.44% accuracy for five-year NSCLC survival, and balanced accuracy greater than 90% for breast and colorectal cancer classification.
  • Recurrent Neural Networks (RNN): RNNs use recurrent connections to retain sequential information, supporting feature selection and cancer prediction across microarray and gene-expression datasets.Reported accuracies include 89.6%, 95%, and a maximum of 95.97% across different studies and datasets.
  • Convolutional Neural Networks (CNN): CNNs learn spatial feature hierarchies and have been applied by transforming high-dimensional RNA-Seq profiles into images for tumor classification.Reported precision values include 98% for an ensemble method and 98.76% for lightweight breast-cancer classification.
  • Graph Neural Networks (GNN): GNNs propagate and aggregate features across graph nodes to model cell-cell and gene-regulatory relationships in gene-expression and multi-omics data.They can incorporate domain knowledge for missing-data imputation and gene-regulatory-network construction, but are sensitive to noise in graph structures.
  • Transformer Neural Networks (TNN): Transformers use multi-head self-attention to capture long-range genomic dependencies and support parallel processing, but may require large datasets.Applications include biomarker recognition across cancer subtypes, multi-omics phenotype discrimination, and pathological-image/genomic-data fusion.

4.3 Transfer Learning

Transfer learning enhances downstream gene-expression models by transferring representations from related source domains. The reviewed work includes CNN-based DEGnext for predicting up- and down-regulated genes in Cancer Genome Atlas data.

  • Transfer Learning: Transfer learning transfers information from related source domains to improve downstream models on untrained cancer datasets.The review presents a table of deep-learning-based transfer-learning methods for gene-expression analysis.
  • Transfer Learning: DEGnext applies transfer learning to a CNN to predict significant up-regulating and down-regulating genes from Cancer Genome Atlas gene-expression data.

4.4 Pathway Analysis

Pathway analysis extracts biological meaning from high-throughput gene expression data by identifying disrupted pathways and, in topology-based methods, incorporating pathway structure.

  • Pathway analysis identifies pathways that may be disrupted by differential gene expression patterns.
  • The adipocytokine signaling pathway distinguished breast cancer from colon and stomach tumors.
  • Pathway analysis methods span Over Representation Analysis, Functional Class Scoring, and Topology Based analysis.The first two generations use pathways as gene sets, while topology-based analysis incorporates pathway topology.
  • These methods support understanding altered genes and pathway routes across cancer types and identifying potential treatments.

5. Future Directions

The review identifies future directions spanning richer molecular features, improved algorithms and benchmarks, interpretable models, multimodal data, biomarkers, and single-cell analyses.

  • Integrating DNA methylations and mutations with RNA-Seq data may produce features that positively impact tumor classification.The paper notes that gene expression alone cannot represent its full impact.
  • Researchers can investigate cancer-specific biomarkers using functional pathway analysis and deep learning features extracted from images.
  • Heterogeneous graph networks may integrate single-cell multi-omics data across experiments, sequencing technologies, and data modalities.
  • Interpretable and explainable models should clarify decision-making and explain cases where models fail.Suggested approaches include counterfactuals and feature attribution for local and global model properties.
  • Future cancer prevention strategies may combine imaging, proteomic, metabolomic, glycemic, and epigenetic modalities to identify and validate surrogate biomarkers.
  • Multimodal and multitask learning with early and late fusion may improve classification performance.
  • Further research may examine circRNA biology and single-cell profiling for cancer gene prediction.
  • Future work should develop computational algorithms and benchmarking approaches for gene expression analysis.

2. Innovation in computational algorithms

A future direction is identifying biomarkers specific to each form of cancer.

  • Researchers should investigate methods for identifying biomarkers specific to each form of cancer.

3. Improved cancer-related biomarkers

Future directions include graph-based integration of single-cell multi-omics, interpretable models, multimodal prevention strategies, and multimodal or multitask learning.

  • Integration of single-cell multi-omics data with graph networks: GNN architectures can integrate single-cell multi-omics data by employing heterogeneous graphs.
  • Design interpretable and explainable approaches: Interpretable ML models should explain computational decision-making and reasons for model failure.
  • Cancer prevention strategies based on multiple data modalities: Cancer prevention strategies may combine imaging, proteomic, metabolomic, glycemic, and epigenetic data to evaluate surrogate biomarkers.
  • Design multi-modal and multi-task learning approaches: Multimodal and multitask ML approaches based on early and late fusion have the potential to improve classification performance.

6. Conclusion

The study reviews machine-learning approaches for cancer classification from gene-expression data, emphasizing deep learning and recent neural-network architectures. It also surveys data collection, datasets, feature engineering, and preprocessing techniques.

  • The review covers traditional and deep-learning methods for cancer classification using RNA-Seq and DNA microarray gene-expression data.
  • It catalogs data-collection methods and commonly used datasets for supervised machine learning in gene-expression analysis.
  • The study provides a taxonomy of feature-engineering and data-preprocessing techniques for gene-expression analysis.
  • It reviews neural-network applications spanning multi-layer perceptrons, convolutional, recurrent, graph, and transformer networks.
  • The review includes graph and transformer architectures that earlier published reviews had not covered.
Loading 2301.12222v1…