REVIEW 4 major objections 5 minor 1 references
An Evolutional Neural Network Framework for Classification of Microarray Data
T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash
Pith's one-line read The paper's GA-MLP classifier reports 99.87% accuracy on colon microarray data using two selected genes.
desk verdict The reported 99.87% accuracy is arithmetically impossible under the paper's own 90/10 split, and the evaluation has too many unsupported details to trust any of the numbers. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing machinery is the GA-MLP pipeline. The genetic algorithm encodes candidate gene subsets as binary chromosomes (bit 1 lets a gene participate, bit 0 excludes it) and searches the 2,000-dimensional space down to two genes; the MLP then takes the two-element concentration feature vector, which the paper calls the declined concentration feature vector, as input, tries hidden-layer sizes from 3 to 15, and uses backpropagation for up to 60 epochs to an error goal of 0.01. The GA is what makes the classification tractable, and the MLP is what converts the reduced representation into a cancer/normal decision.
What would settle it
Re-run the experiment with the MLP hidden-layer size selected inside the training folds only and evaluate a fixed two-gene signature on a held-out test set. The paper's own numbers point to a check: with 62 samples split 90/10, each test set has about six samples, so per-run accuracy is a multiple of roughly one-sixth, and no average of 20 such runs can equal 99.87%.
Extended reading notes
Core claim
The paper's central discovery is that a binary-encoded genetic algorithm can shrink a 2,000-gene colon cancer expression set to exactly two genes, and a three-layer MLP trained on the resulting declined concentration feature vector classifies the 62 samples into tumor and normal classes at 99.87% accuracy over 20 runs of a 90/10 split. The paper presents this result as beating SVM and Naive Bayes, both reported at 93.55%, with the same or fewer selected features.
Load-bearing premise
The load-bearing premise is that the reported 99.87% accuracy is a valid out-of-sample estimate, which requires that the test data never influenced the choice of hidden-layer size or any other model decision.
Editorial extensions
If this is right
- A two-gene expression signature would be enough to separate colon cancer from normal tissue, shrinking a diagnostic assay from whole-chip profiling to two measurements.
- The same GA+MLP pipeline could be applied to other cancer microarray datasets as a template for dimensionality reduction and classification.
- On the colon dataset, the GA-selected two-feature MLP would outperform the compared SVM and Naive Bayes classifiers by about 6 percentage points.
- Using two genes instead of 2,000 reduces computational time and complexity during both training and testing.
Reading between the lines
- Beyond the paper: the stated protocol cannot produce the reported number, since a 90/10 split of 62 samples yields test sets of about six samples and per-run accuracy is a multiple of one-sixth; an independent evaluation would be needed to see whether a two-gene signature really generalizes.
- Beyond the paper: using the same test set to choose the MLP hidden-layer size (3 to 15) and then to report accuracy risks an optimistically biased estimate; nested cross-validation would give a more trustworthy figure.
- Beyond the paper: the gene identities selected by the GA should be reported and checked for stability across runs; without that, it is unclear whether the result is a fixed diagnostic signature or a run-specific artifact.
- Beyond the paper: if a fixed pair of genes were re-tested on an independent colon cancer cohort, the actual accuracy would settle whether the claim extends beyond this 62-sample dataset.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a hybrid genetic-algorithm (GA) feature selector combined with a multi-layer perceptron (MLP) classifier for binary classification of microarray gene-expression data. The method is applied to the public colon cancer dataset (62 samples, 2000 genes), with the claim that it reduces the feature set to two genes and achieves 99.87% accuracy, compared to 93.55% for SVM and Naive Bayes. The manuscript contains no theoretical derivations; its contribution is entirely empirical. The evaluation, however, is described in an internally inconsistent way, and the headline accuracy figure is arithmetically impossible under the stated 90/10 split on 62 samples.
Significance. If the reported result were valid, a two-gene signature with near-perfect classification accuracy on a well-known colon cancer microarray dataset would be a practically interesting contribution to the feature-selection and classification literature. The paper also has the merit of attempting to combine evolutionary search with neural-network classification for high-dimensional, small-sample data. However, the central empirical claim is not supported by the described protocol: the arithmetic of the reported accuracy is incompatible with the stated train/test split, the evaluation is not independent because the hidden-layer size is chosen on the test set, and the baseline comparisons are not reproducible. As presented, the contribution is not established.
major comments (4)
- [Section IV] The reported accuracy of 99.87% (final results table in Section V) is arithmetically incompatible with the evaluation protocol described in Section IV. The colon dataset contains 62 samples (Section III.B), and Section IV states that 20 runs are performed with a 90/10 train/test split, so each test set has either 6 or 7 samples. If accuracy is pooled over runs, the total number of test samples is T = 6a + 7(20-a) for some integer a between 0 and 20, and the total number of correct classifications C is an integer; thus accuracy is C/T. No such fraction equals 0.9987 (the achievable values near 0.9987 are 0.9917 and 1.0000). If instead the reported accuracy is the unweighted mean of per-run accuracies, each per-run accuracy is a multiple of 1/6 or 1/7, so the mean has a denominator dividing 840, which 0.9987 does not. The central numeric claim therefore cannot arise from the stated experiment and must be re-derived or corrected.
- [Section IV] The experimental protocol is internally inconsistent. The text first says '20 runs are done' with 90% training and 10% testing, but later in the same section it states 'All these results are obtained by using 10-fold validation and 50 to 100 features'. Table 1 is said to show performances on the 'Ling-Spam dataset', whereas the paper is about the colon cancer dataset. These statements cannot all describe the experiments that produced the final comparison table, and the manuscript does not clarify which protocol, dataset, and feature counts were actually used.
- [Section IV] The evaluation procedure is not independent. Section IV says 'The nodes of hidden layer are tested from 3 to 15'; selecting the hidden-layer size that maximizes test accuracy on the same 10% test sets used to report the final accuracy means the reported accuracy is a selected maximum rather than an unbiased estimate. No nested cross-validation, separate validation set, or correction for multiple comparisons is described. Additionally, no variance, confidence intervals, or per-run accuracies are reported for the 20 runs, despite the tiny test sets (6 or 7 samples each), which make the per-run estimates extremely noisy.
- [Section V] The comparison with SVM and Naive Bayes is not reproducible from the information given. The final results table reports SVM at 93.55% with 2 features and Naive Bayes at 93.55% with 3 features, but the manuscript does not state the feature-selection method, hyperparameters, or train/test protocol for these baselines, nor whether they used the same splits as the proposed method. The column labels 'SVM_KEGG', 'NB_KEGG', and 'MLP_GA' and the phrase 'declined concentration feature vector' in the conclusion refer to concepts not defined in the paper. Without these details, the claimed superiority is unsupported.
minor comments (5)
- [Sections I and II] There are multiple incomplete or malformed citations (e.g., '[]', '[..]', '[...]') and references cited as [5-9] or [6] that cannot be resolved from the reference list; the bibliography needs to be completed.
- [Section III.E] In the definitions of the confusion-matrix terms, TN and FP are both described as 'normal tissues which are predicted as Cancerous'; TN should be normal tissues correctly predicted as normal. This typo makes the accuracy formula ambiguous.
- [Section III.C] The GA parameters (population size, crossover rate, mutation rate, and number of generations) are mentioned as important but their values are never reported; the feature-selection procedure cannot be reproduced without them.
- [Section VI] The phrase 'declined concentration feature vector' appears in the conclusion but is never defined in the methods; either define the term or remove it.
- [General] Table numbering is inconsistent: the dataset description is 'Table I', then 'Table 1' appears in Section IV, and the final comparison table is unlabeled; the text should refer to tables consistently.
Circularity Check
No circular derivation: GA+MLP pipeline is empirical; reported accuracy inconsistencies are correctness issues, not circularity.
full rationale
The paper contains no formal derivation or first-principles result; it is an empirical pipeline in which a GA selects two genes and an MLP classifies them. GA feature selection and MLP training are independent of the reported test statistic (accuracy = (TP+TN)/(FN+FP+TP+TN)), which is a standard evaluation formula, not an input to the method. No cited theorem, uniqueness claim, or ansatz is load-bearing; prior-work citations are background. The main concerns are not circularity: (1) Section IV states a 90/10 split over 20 runs on 62 samples, which yields 6-7 test samples per run, making the reported 99.87% accuracy arithmetically implausible (a data-integrity/correctness issue); (2) Section IV says hidden-layer nodes were 'tested from 3 to 15', which risks model selection on the test set, but the paper does not define the fitness function or selection rule, so there is no exhibited reduction from output to input. These flaws affect validity, not circular self-derivation, so the circularity score is 0.
Assumptions & free parameters
free parameters (4)
- MLP hidden layer node count =
unknown (tested from 3 to 15)
- GA population size, crossover and mutation rates, number of generations =
not reported
- Train/test split ratio =
90/10
- MLP training epochs and error goal =
60 epochs, 0.01 error goal
assumptions (3)
- domain assumption The GA fitness function based on classification accuracy identifies a generalizable two-gene subset.
- ad hoc to paper The colon cancer dataset (62 samples) with a 10% test split provides a statistically reliable accuracy estimate.
- standard math A three-layer MLP with a binary input vector can represent the cancer/normal decision boundary.
Cite this review
Pith. "Pith review of An Evolutional Neural Network Framework for Classification of Microarray Data." pith.science (2026). https://pith.science/paper/ADN2WSJK
@misc{pith2026241113326,
author = {Pith},
title = {Pith review of: An Evolutional Neural Network Framework for Classification of Microarray Data},
year = {2026},
howpublished = {\url{https://pith.science/paper/ADN2WSJK}},
note = {Machine review of arXiv:2411.13326}
}
read the original abstract
DNA microarray gene-expression data has been widely used to identify cancerous gene signatures. Microarray can increase the accuracy of cancer diagnosis and prognosis. However, analyzing the large amount of gene expression data from microarray chips pose a challenge for current machine learning researches. One of the challenges lie within classification of healthy and cancerous tissues is high dimensionality of gene expressions. High dimensionality decreases the accuracy of the classification. This research aims to apply a hybrid model of Genetic Algorithm and Neural Network to overcome the problem during subset selection of informative genes. Whereby, a Genetic Algorithm (GA) reduced dimensionality during feature selection and then a Multi-Layer perceptron Neural Network (MLP) is applied to classify selected genes. The performance evaluated by considering to the accuracy and the number of selected genes. Experimental results show the proposed method suggested high accuracy and minimum number of selected genes in comparison with other machine learning algorithms.
Reference graph
Works this paper leans on
-
[1]
An Evolutional Neural Network framework for Classification of Microarray Data Maryam Eshraghi Evari, Md Nasir Sulaiman, Amir Rajabi Behjat ABSTRACT DNA microarray gene-expression data has been widely used to identify cancerous gene signatures. Microarray can increase the accuracy of cancer diagnosis and prognosis. However, analyzing the large amount of ge...
work page 2000
Reviewed August 12, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.