Potential of Artificial Intelligence Algorithms for Identification of Relevant Diagnostic and Prognostic Biomarkers of Early-Stage Liver Cancer
Ali Bou Nassif, Darko Castven, Manar Abu Talib, Jibran Sualeh Muhammad, Ahmed Ammar Kubba, Jens Marquardt, Abdalla Sayed Ali
cs.AI
Quick verdict
Claims 90.74% five-stage HCC classification & in-vitro DNAJB14 validation; real translational value, but leakage and no per-class metrics weaken ML claims.
Summary
The paper trains a Keras feedforward neural network (64- and 32-neuron ReLU layers, 20% dropout) to classify 770 liver tissue samples into five hepatocarcinogenesis stages — surrounding liver, low- and high-grade dysplastic nodules, early and progressed HCC — from 11,150 gene-expression features. The dataset merges a private Lübeck cohort with TCGA and GSE89377 via a semi-supervised procedure that is not described in detail. Gene-set ablations reach 59.26% (Lübeck, 5 genes), 61.48% (TCGA, 4 genes) and 86.67% (constructed set, 5 genes including DNAJB14); SelectKBest applied on the full dataset then yields 90.74% accuracy at K=15 and lowest loss 0.3187 at K=20, with inverse-frequency class weights offsetting severe imbalance (smallest class: 26 samples). SHAP explanations of three tree ensembles (random forest with ant-colony and genetic selection, XGBoost with particle swarm) consistently rank DNAJB14 first. Huh7 experiments show siRNA knockdown reduces viability, colony and sphere formation (up to roughly 52% and 48%), migration and invasion, and GEPIA2/TCGA analysis links high expression to worse overall survival. Key limitations: no per-class metrics or confusion matrix, probable feature-selection leakage, no external validation, undocumented dataset construction, and unreported accuracy for the explained models.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A feedforward neural network using 15 genes selected by SelectKBest achieves 90.74% test accuracy for five-stage hepatocarcinogenesis classification, with lowest loss 0.3187 at 20 genes.
- SHAP analysis of three model/optimizer ensembles consistently ranks DNAJB14 as the most influential gene, primarily for distinguishing early and progressed HCC.
- siRNA-mediated knockdown of DNAJB14 in Huh7 cells impairs viability, colony and sphere formation, migration and invasion, presented as the first functional validation of DNAJB14 in HCC.
- High DNAJB14 expression is associated with worse overall survival in TCGA HCC patients via GEPIA2 analysis.
Strengths
In-vitro functional validation of the top XAI hit
Rare in CS-led ML papers: siRNA knockdown of DNAJB14 in STR-authenticated Huh7 cells (Sections 2.4.2-2.4.11, Fig. 11), confirmed by qRT-PCR and Western blot, reduces viability, colony/sphere formation (~52% and ~48% reduction), migration (Ibidi with cycloheximide proliferation control) and invasion (xCELLigence), with ANOVA and n=3.
Clinically finer classification target than most reviewed baselines
Five-stage hepatocarcinogenesis (sl/lgdn/hgdn/ehcc/phcc) on 770 samples with 11,150 features, versus the largely binary cancer/normal setups in the reviewed literature (Table 1); the dysplastic-nodule classes matter for early detection.
XAI triangulation across model families and optimizers
Imbalance is disclosed and mitigated, and model choices are ablated
Weaknesses
No per-class evaluation despite extreme imbalance
Probable feature-selection leakage and contradictory split description
The decisive dataset artifact is undocumented and partly private
Performance of the explained models is never reported
Thin clinical statistics and domain inaccuracies
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.15638, cs.AI) combines deep learning, explainable AI and wet-lab validation for five-stage hepatocarcinogenesis classification: surrounding liver (sl), low/high-grade dysplastic nodules (lgdn/hgdn), early and progressed HCC (ehcc/phcc). Four central claims: (1) a Keras feedforward network on 15 SelectKBest-selected genes reaches 90.74% test accuracy, with lowest loss 0.3187 at 20 genes (Table 4); (2) SHAP across three model/optimizer ensembles consistently ranks DNAJB14 first (Section 3.2); (3) siRNA knockdown of DNAJB14 in Huh7 cells impairs viability, colony/sphere formation, migration and invasion — presented as the first functional validation of DNAJB14 in HCC; (4) GEPIA2/TCGA analysis links high DNAJB14 expression to worse overall survival. The translational arc from SHAP hit to bench confirmation is the paper's most distinctive element and its strongest evidence.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What accuracy did the AI model achieve for liver cancer staging?
The best model used 15 genes chosen by SelectKBest and reached 90.74% overall test accuracy on a 770-sample, five-class gene-expression dataset; the lowest loss (0.3187) came from a 20-gene variant. The paper reports only overall accuracy — no per-class results or confusion matrix — on a dataset where the two largest classes cover about 83% of samples.
What is DNAJB14 and why does it matter in this paper?
DNAJB14 is an endoplasmic-reticulum-localized member of the DNAJ (Hsp40-family) co-chaperone proteins. In this study, SHAP explainability analysis ranked it the most influential gene across three machine-learning models, and lab experiments showed that switching it off in Huh7 liver-cancer cells reduces their survival, colony and sphere growth, migration and invasion. TCGA survival data also associate high expression with worse outcomes.
Can I reproduce this study or download its dataset?
Not fully. The five-class dataset merges a private Lübeck cohort with public TCGA and GSE89377 data via a semi-supervised procedure the paper does not specify in detail, and no code or processed data are released. Reimplementing the neural network itself is straightforward, but the exact training data cannot be reconstructed.
How was the class imbalance in the dataset handled?
The authors used inverse-frequency class weights during training — for example, a weight of 5.92 for the 26-sample surrounding-liver class versus 0.47 for progressed HCC — so rarer classes count more in the loss. However, because only overall accuracy is reported, the actual effect on minority-class detection cannot be verified from the paper.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures (SHAP plots, training curves, Kaplan-Meier, in-vitro panels) cannot be inspected from extracted text; judgments rely on the paper's textual descriptions.
- Table/figure cross-references are garbled in extraction (e.g., 'Figure 111', Table 1 vs 2 mixups); a few flagged inconsistencies may partly reflect extraction noise rather than authorial error.
- Per-class performance could theoretically exist inside figures that were not analyzable, though the text never mentions any.
- Reference list appears duplicated in two numbering series, suggesting a journal-formatted preprint; some sections may be truncated.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.