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

In plain words

Doctors and researchers can measure how active each gene is in a tissue sample. That activity profile is a long list of numbers, with many more genes than samples, and it can hint at whether a tumour is present and what kind it is. This paper is a review: it surveys how machine learning has been used on such profiles to classify cancers.

The main obstacle is the shape of the data. With many more genes than samples, a model easily latches onto noise, so most methods first pick a smaller set of useful genes. The review sorts these selection methods into three families: cheap score-based filters, search-based wrappers that retrain a classifier many times, and selection built into the training itself. It then walks through classic models such as support vector machines and random forests, and the neural networks that have taken over more recently: multi-layer perceptrons, convolutional, recurrent, graph and transformer networks. It also covers borrowing knowledge from a model trained on other data (transfer learning), the datasets people use, and open research directions. The authors focus on RNA sequencing data, the more common measurement type today.

TL;DR

• A survey, not a method: no new model, training procedure or code is proposed §1.
• Gene expression data have far more genes than samples, so feature selection (filter, wrapper, embedded) is a central step §1 §3 fig. 1.
• The review reports many accuracies, several above 90%, from different datasets and protocols that are not directly comparable tab. 4 tab. 6–tab. 10.
• Open directions: more data types such as DNA methylation and mutations, interpretable models, graph methods for single-cell multi-omics, and multimodal or multitask learning §5 tab. 11.

Problem

Gene expression profiles have a large number of genes per sample and relatively few samples, and many genes are redundant or irrelevant for a given classification task §1 §3. The authors state that conventional machine learning depends strongly on the quality of its features, which makes feature selection important §4.1. They also note that transformer models need large amounts of data and that recurrent models are more prone to overfitting when data are scarce §4.2.2 §4.2.5; transfer learning is used to offset small training sets §1 §4.3. Reported limitations of current methods include sensitivity to parameter settings, weak interpretability and limited integration with other data types §1.

Method

This is a survey. The "method" is the taxonomy the review uses, and no experiments are run by the authors.

1. Data sources §2 tab. 2 tab. 3. Microarrays measure expression with hybridisation to a fixed probe array. RNA sequencing (RNA-seq) converts RNA to cDNA, sequences it and counts reads per gene. The review compares the two tab. 2 and treats RNA-seq as the dominant format. It lists 14 public or open-source cancer datasets, from 62 to 14,237 samples tab. 3 §2.4.
2. Feature engineering §3 fig. 1 tab. 5. Three families:
   • Filter methods score each gene independently of any classifier and keep the top-ranked ones with a threshold §3.1.
   • Wrapper methods wrap a classifier in a search over gene subsets; the classifier's performance is the objective §3.2.
   • Embedded methods perform selection while a given learner is trained §3.3.
   • Hybrid methods chain a filter in front of a wrapper or use ensembles §3.4. Table 5 lists pros and cons per family, with univariate versus multivariate and deterministic versus randomised variants for filters and wrappers.
3. Traditional ML §4.1 tab. 6. SVM, k-nearest neighbours, naive Bayes and random forest, often combined with a selection step.
4. Deep networks §4.2. The review describes, for each architecture, the input encoding and the reported results:
   • MLP on expression vectors, sometimes with clinical data or pathway-derived features §4.2.1 tab. 7.
   • RNN (LSTM and related) applied to expression data for feature selection or classification §4.2.2 tab. 8.
   • CNN on 1D gene vectors, or on 2D "images" made by placing genes in a fixed layout; the review notes that colour-palette encodings lose information §4.2.3 tab. 9.
   • GNN on graphs whose nodes are genes or cells and whose edges come from correlation thresholds or interaction databases §4.2.4.
   • Transformer models with multi-head self-attention, including multi-omics variants §4.2.5.
5. Transfer learning §4.3 tab. 10. A model trained on a large source dataset is reused for a smaller target cancer dataset.
6. Pathway analysis §4.4. Methods that relate differential expression to curated pathway databases (KEGG, Reactome) in three generations.

Equations and algorithm boxes: none in the paper.

Overall pipeline pattern the review implies (a generic summary, not a procedure given by the authors; see §3 for the families):

# Typical gene-expression classification pipeline  [§3] [§4]
# Input:  expression matrix X (samples, genes); labels y
# Output: trained classifier and selected genes

def train_pipeline(X, y, selector, model):
    # Reduce the many genes to a smaller subset  [§3]
    genes = selector.select(X, y)
    # Restrict the data to those genes  [§3] [fig. 1]
    X_reduced = X[:, genes]
    # Fit a classical or deep classifier  [§4.1] [§4.2]
    model.fit(X_reduced, y)
    return model, genes

Filter, wrapper and embedded selectors differ only in how select is implemented, as described in the three steps above §3.1–§3.3 fig. 1.

Hyperparameters

Not applicable: the review proposes no model of its own and states no training hyperparameters for it. Settings of the reviewed works are not tabulated in the survey; only performance figures are, in tab. 4 and tab. 6–tab. 10. The text gives no learning rates, epochs or batch sizes for any model §4.

Results

Author-reported, as tabulated by the review: accuracies and other metrics come from the cited original works. The review does not rerun them, and the datasets, sample sizes, class counts and validation protocols differ across rows, so the numbers are not comparable with each other. Headline values:

Not reported by the review: any sample split, repeat count, or variance for these numbers.

Paper vs code

No official code. The paper has no code link, and the arXiv metadata records none. Search done: the authors' GitHub repository list (the avakanski account, 32 repositories) has no repository for this review; two of its multi-omics graph-network repositories mention cancer classification but are not linked from the paper and were not used. A GitHub repository search on the review's topic returned one unrelated student project. Because this is a survey, there is no method implementation to compare against the text.

Pitfalls

• No method, no implementation target. An agent cannot implement "the paper"; the content is a map of published methods, each with its own original implementation. Use the cited original works for implementations §4.
• Metric mismatch for Xiao et al. The text says the three RNA-seq cancers reached "a precision of 98%" §4.2.3; Table 9 lists accuracies of 95.60%, 94.63% and 94.62% for LUAD, STAD and BRCA tab. 9. Treat both as unreconciled.
• Mostavi et al. figure not in the table. The text gives 88.42% for breast cancer subtype prediction with a 1D-CNN §4.2.3; Table 9 lists 95.7%, 92.5% and 95.7% for the same model family, with no subtype row. Which task each number refers to is not clear from the survey.
• Segal et al. number. The text says 75 of 76 instances were classified correctly in leave-one-out validation §4.1, which is about 98.7%; Table 6 lists 98.5% tab. 6. The table value cannot be derived from the text.
• Laplante et al. class counts. The text describes classifying 20 anatomical areas and also identifying 27 cancer types, with a 27-neuron output layer §4.2.1. The survey does not say which count applies to the reported 96.9%.
• Dataset labels. Table 7 lists the Breast Cancer Wisconsin (Original) dataset as "Gene Expression Data" tab. 7. The dataset's name suggests a non-genomic source; the survey does not say what features it used. Check the original work before relying on that row as gene expression evidence.
• Units. Table 7 gives an ROC-AUC of "0.714%" for Zhang et al. tab. 7; AUC is not a percentage. Read it as 0.714.
• Unmapped values. Zhang & Liu 2021 lists six accuracies (97–99%) with no mapping to methods tab. 4. Wu & Hicks (four classifiers) and Gakii et al. (three methods) are mapped only by order tab. 4.
• Naming. The text calls the Zhang et al. method "SVM-RFE-PO" (parameter optimisation) §4.1, while Table 6 writes "SVM-RFE-PSO" tab. 6.
• Comparability. The accuracy spread from 62-sample to 14,237-sample datasets tab. 3 means the top rows of tables 4 and 6–10 cannot be ranked against each other. The review gives no protocol column (split, cross-validation, repeats).
• Literature search protocol not stated. Sections 1–2 compare prior reviews tab. 1 and state a focus, but give no databases, date range or inclusion criteria for the works surveyed.
• Unclear sentence on determinism. §3.5 links deterministic versus randomised wrapper methods to the influence of random variation on the model in a way that is hard to parse; the table tab. 5 is clearer than the prose.
• Unsupported general claims. Statements such as embedded methods "generally produce an improved performance" §3.3 and that most future models will be based on deep networks §1 are the authors' outlook, without a cited experiment in the survey text.
• Figure 1 is credited to another work. The caption attributes the figure to reference [62] fig. 1. This card describes it in words only and does not reproduce it.
• Table 2 is qualitative. The microarray versus RNA-seq comparison lists yes/no features with no citation per row tab. 2.
• Future directions are suggestions. Table 11 and §5 list open problems; they are not results §5 tab. 11.

Lineage

• Builds on: none in the sense of a method; it reviews prior reviews, listed in tab. 1 (Sathe et al. 2019; Koumakis et al. 2020; Zhu et al. 2020; Gunavathi et al. 2020; Tabares et al. 2020; Bhonde et al. 2021; Mazlan et al. 2021; Karim et al. 2021; Thakur et al. 2021; Montesinos-López et al. 2021; Bhandari et al. 2022; Khalsan et al. 2022; Alhenawi et al. 2022).
• Compared against: the authors state that no previous review offers a comprehensive discussion of GNN and transformer architectures for this task §1. Their comparison of prior reviews by coverage (conventional ML, feature engineering, DL types, microarray and RNA-seq) is in tab. 1.
• Superseded by: not known from content in this wiki.