• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Sunday, November 22, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Health

Synergistic Machine Learning Framework Identifies Biologically Relevant Cancer Genes

by
November 22, 2026
in Health
Reading Time: 6 mins read
0
Synergistic Machine Learning Framework Identifies Biologically Relevant Cancer Genes

Synergistic Machine Learning Framework Identifies Biologically Relevant Cancer Genes

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

A computational framework combining eight nature-inspired metaheuristic search algorithms with five machine learning classifiers and pathway enrichment analysis has demonstrated improved capability in identifying biologically relevant cancer-associated genes from high-dimensional microarray data. Developing effective diagnostic and prognostic biomarkers remains a difficult challenge in oncology because transcriptomic datasets typically involve tens of thousands of gene expression measurements collected across a limited number of patient samples. To address this discrepancy, researchers Shamima Naznin and M. Saifur Rahman established an integrated computational pipeline designated as CGI_MMLP. The methodology bridges algorithmic feature selection with functional biological assessment to isolate compact gene subsets from high-dimensional, low-sample-size microarray profiles. When evaluated on benchmark datasets representing diffuse large B-cell lymphoma, acute leukemia, and prostate cancer, the integrated strategy prioritized candidate genes that aligned strongly with established oncogenic pathways. In biological validation tests, the framework achieved a pathway relevance ratio of 68.33 percent in lymphoma, 81.6 percent in leukemia, and 58.0 percent in prostate cancer, substantially exceeding the biological enrichment proportions attained by single-model feature selection baselines.

Microarray technology simultaneously monitors the expression levels of thousands of genes, providing valuable insights into the genetic and epigenetic alterations that disrupt normal cellular regulation and accelerate malignant proliferation. However, analyzing high-dimensional, low-sample-size gene expression datasets presents severe statistical hurdles, including non-linear data distributions, measurement noise, and an elevated vulnerability to model overfitting. The researchers tackled distribution irregularities during the data preprocessing stage by applying the Yeo-Johnson power transformation to the expression matrices. Unlike conventional logarithmic transformations that require strictly positive values, the Yeo-Johnson transformation accommodates both positive and negative expression values, effectively stabilizing variance and diminishing distributional skewness before downstream computation. Following data transformation, the pipeline implements an initial dimensionality reduction filter utilizing the absolute Pearson correlation coefficient. By ranking all profiled genes according to their linear relationship with the diagnostic target variable, the framework extracts the top one hundred most correlated candidates. This pre-filtering procedure reduces the search space for subsequent iterative algorithms, decreasing overall computational run time while retaining highly informative features.

The core computational engine of CGI_MMLP relies on an ensemble wrapper architecture that pairs eight distinct metaheuristic search algorithms with five supervised machine learning classifiers. The metaheuristic suite incorporates single-objective swarm intelligence approaches, including Bee Colony Optimization, Ant Colony Optimization, and the Wolf Search Algorithm, alongside single-objective evolutionary techniques such as the standard Genetic Algorithm, general Evolutionary Algorithms, and the plant-inspired Flower Pollination Algorithm. To account for trade-offs between predictive performance and gene signature compactness, the framework also integrates multi-objective evolutionary algorithms, specifically the Non-dominated Sorting Genetic Algorithm II and the Evolutionary Non-dominated Radial Slots-based Algorithm. These multi-objective formulations concurrently optimize classification accuracy while minimizing the total number of retained genes. Each optimization technique navigates the candidate space in tandem with five distinct classifiers: Support Vector Machines, Naive Bayes, AdaBoost, Decision Trees, and Random Forests. By evaluating diverse combinations across different algorithmic paradigms, the wrapper systematically balances global exploration with local search exploitation across complex feature landscapes.

To synthesize the variable gene subsets generated by different classifier and metaheuristic pairings, the researchers introduced a weighted F-measure consensus ranking formulation. Single-objective algorithms select the gene subset maximizing cross-validated classification accuracy, whereas multi-objective algorithms generate a non-dominated Pareto front, from which the highest-performing subset with the fewest constituent genes is extracted. The framework then assigns each gene an accumulated ranking score determined by summing the subset-level cross-validated F-measure values from every algorithm-classifier combination that selected that gene. Under this formulation, genes chosen repeatedly across diverse optimization routines receive higher weights, particularly when identified by combinations exhibiting superior classification performance. Following the scoring process, the top thirty ranked genes are isolated for each cancer type. Sensitivity analyses examining candidate subset sizes of ten, twenty, thirty, fifty, and seventy genes demonstrated that retaining thirty genes maintained robust biological coverage without generating the redundant pathway associations observed in larger sets.

The experimental validation utilized three public microarray datasets obtained from the ICO2S repository, encompassing diffuse large B-cell lymphoma with 4,026 genes across 62 samples, acute leukemia with 7,129 genes across 72 samples, and prostate cancer comprising 2,136 genes across 105 samples. To prevent data leakage and ensure rigorous generalization estimates, the researchers implemented a strict evaluation protocol. Each dataset was partitioned through stratified sampling into an 80 percent training set and an independent 20 percent held-out test set, preserving original diagnostic class distributions. All data processing steps, including missing-value imputation, Yeo-Johnson transformation, Pearson correlation ranking, five-fold cross-validation wrapper tuning, and classifier parameter optimization, were performed entirely within the training partition. Independent test sets remained completely unobserved until final model evaluation. Parameter grid searches were conducted using a Support Vector Machine classifier on the lymphoma data to establish optimal settings, such as population sizes and operator probabilities across metaheuristics, before cross-validation benchmarking.

Across the independent test evaluations, CGI_MMLP demonstrated competitive diagnostic predictive accuracy when benchmarked against two state-of-the-art gene selection frameworks, LVTree and FF_SVM. On the lymphoma test set, CGI_MMLP attained the highest performance with an accuracy of 0.956 and an F-measure of 0.961, outperforming LVTree at 0.941 and FF_SVM at 0.933. On the prostate cancer test partition, CGI_MMLP similarly achieved superior results, recording an accuracy of 0.9167 and an F-measure of 0.9245 compared to 0.867 for both baseline techniques. On the acute leukemia test dataset, CGI_MMLP achieved an accuracy of 0.883 and an F-measure of 0.882125, slightly below the test accuracy of 0.928 recorded by LVTree and 0.887 by FF_SVM. During training cross-validation assessments, all three methods achieved perfect 1.0 accuracy and F-measure scores on the leukemia dataset, while CGI_MMLP matched FF_SVM with a 1.0 accuracy on prostate cancer, reflecting the high separability of reduced feature spaces.

Beyond classification metrics, the research prioritized verifying whether computationally selected gene signatures exhibited authentic biological relevance. Gene sets identified by each method were submitted to over-representation analysis within the Enrichr platform to quantify their Pathway Relevance Ratio, defined as the proportion of statistically significant pathways relative to total identified cancer-associated terms. In the lymphoma dataset, CGI_MMLP produced a Pathway Relevance Ratio of 68.33 percent, whereas LVTree and FF_SVM achieved only 8.33 percent and 6.67 percent, respectively. For acute leukemia, CGI_MMLP achieved an 81.6 percent ratio, compared to 78.3 percent for LVTree and 0.0 percent for FF_SVM. In prostate cancer, CGI_MMLP reached 58.0 percent, surpassing LVTree at 48.0 percent and FF_SVM at 44.0 percent. Jaccard similarity index comparisons revealed minimal gene overlap between CGI_MMLP and FF_SVM, registering just 1.72 percent in lymphoma and leukemia, which indicates that conventional methods often prioritize statistically predictive genes lacking functional disease associations.

Ablation analyses conducted on the lymphoma dataset confirmed the functional necessity of integrating diverse metaheuristic and classifier components. The full pipeline combining all classifiers with all eight metaheuristic algorithms achieved the highest Pathway Relevance Ratio at 68.33 percent. Restricting the system to Support Vector Machines while retaining all metaheuristics reduced the ratio to 30.0 percent. Conversely, pairing the best single metaheuristic, the Flower Pollination Algorithm, with all five classifiers yielded the lowest relevance ratio at 23.33 percent, while combining only the Flower Pollination Algorithm with Support Vector Machines produced a ratio of 36.66 percent. These comparative findings underscore that diverse metaheuristics explore complementary trajectories through high-dimensional feature spaces. Additionally, assessing the impact of initial Pearson correlation filtering revealed substantial efficiency and diagnostic gains; for example, integrating Pearson pre-filtering with Support Vector Machines elevated classification accuracy across ant colony, genetic, and multi-objective optimization methods while significantly reducing execution times across all evaluated models.

Functional enrichment and network analysis visualized through NetworkX confirmed that prioritized gene sets converged onto coherent biological mechanisms. In diffuse large B-cell lymphoma, the top thirty genes mapped recurrently to hematopoietic and lymphoid tissues, highlighting cell cycle regulators such as Cyclin-Dependent Kinase 1 and Minichromosome Maintenance complex members MCM2, MCM3, MCM6, and MCM7, which control DNA replication licensing. In acute leukemia, the identified genes included B-cell developmental markers VPREB1 and CD19, cell cycle regulator CCND3, hematopoietic transcription factor MYB, and chemokine receptor CXCR4. When compared directly against a recently published hybrid ensemble framework known as WCLFJHEF on leukemia data, CGI_MMLP achieved an 88 percent pathway relevance ratio across top-twenty gene sets versus 20 percent for the baseline. Furthermore, tissue-specific enrichment confirmed that CGI_MMLP targets associated strongly with bone marrow and blood cells, whereas the comparative baseline highlighted non-hematological tissues such as the pharynx and small intestine.

For prostate cancer, biological process evaluations identified significant enrichment in epithelial cell proliferation, vasculature development, angiogenesis regulation, and prostaglandin metabolic pathways, highlighting established biomarkers such as KLK3, TP63, ERG, hepsin, and dermatopontin. Despite these positive outcomes, the authors noted several methodological constraints inherent to the study design. High-dimensional, low-sample-size datasets remain susceptible to subtle model overfitting, and the independent test sets comprised relatively small sample cohorts ranging from twelve to sixteen patient samples, making reported test accuracy metrics sensitive to single misclassifications. The investigators emphasized that no experiments were performed on human subjects or live animals, relying entirely on anonymized public datasets. Future research initiatives will aim to expand the CGI_MMLP architecture to multi-class diagnostic problems, evaluate cross-dataset transferability across heterogeneous cohorts, incorporate single-cell RNA sequencing data, and integrate drug-gene interaction networks to assist in identifying actionable therapeutic targets.

Subject of Research: Computational cancer gene identification combining metaheuristic feature selection and pathway enrichment validation

Article Title: CGI_MMLP: A synergistic metaheuristics-machine learning framework for cancer gene identification with pathway-enrichment-based validation

Article References: Naznin, S., & Rahman, M. S. (2026). CGI_MMLP: A synergistic metaheuristics-machine learning framework for cancer gene identification with pathway-enrichment-based validation. Informatics in Medicine Unlocked, 67, Article 101835. https://doi.org/10.1016/j.imu.2026.101835

Image Credits: AI Generated

DOI: 10.1016/j.imu.2026.101835

Keywords: cancer gene identification, metaheuristic algorithms, machine learning classifiers, pathway enrichment analysis, microarray gene expression, high dimensional low sample size, CGI_MMLP, feature selection, Yeo Johnson transformation, Pearson correlation ranking, diffuse large B cell lymphoma, acute leukemia, prostate cancer, pathway relevance ratio, Enrichr, support vector machine, multi objective optimization

News Source: Nathaniel Bowman. (November 22, 2026). Synergistic Machine Learning Framework Identifies Biologically Relevant Cancer Genes. Scienmag.

Tags: acute leukemiacancer gene identificationCGI_MMLPdiffuse large B-cell lymphomaEnrichRFeature Selectionhigh dimensional low sample sizemachine learning classifiersmetaheuristic algorithmsmicroarray gene expressionMulti-objective optimizationpathway enrichment analysispathway relevance ratioPearson correlation rankingProstate Cancersupport vector machineYeo Johnson transformation
Share12Tweet7Share2ShareShareShare1

Related Posts

Visceral Fat Mitochondria Alter Substrate Preferences in Prediabetes Before Baseline Respiration Declines

Visceral Fat Mitochondria Alter Substrate Preferences in Prediabetes Before Baseline Respiration Declines

November 22, 2026
Liquid Biopsy Detects TP53 Mutations in Bangladeshi Liver Cancer Patients

Liquid Biopsy Detects TP53 Mutations in Bangladeshi Liver Cancer Patients

November 22, 2026

Inhaled Corticosteroid Initiation Linked to Fewer Severe Childhood Asthma Attacks

November 22, 2026

Indian Symposium Reports Progress and Structural Gaps in Dance Movement Therapy

November 22, 2026

POPULAR NEWS

  • Alexithymia and Childhood Abuse Link Autistic Traits to Psychotic Experiences in Young Adults

    Alexithymia and Childhood Abuse Link Autistic Traits to Psychotic Experiences in Young Adults

    29 shares
    Share 12 Tweet 7
  • Multi-Layer Framework Protects Patient Privacy in Psychiatric Audio Recordings

    29 shares
    Share 12 Tweet 7
  • Breast Tomosynthesis Image Quality Drops Sharply When Compression Reduction Exceeds Fifty Percent

    29 shares
    Share 12 Tweet 7
  • Small-Molecule Cocktail Preserves Primitive Human Hematopoietic Stem Cells During Expansion

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Custom Genomic Probe Set Unveils Widespread Unknown Hybridization in Pondweeds

Real-World Data Linkage Reveals Return to Work Patterns After Breast Cancer

Visceral Fat Mitochondria Alter Substrate Preferences in Prediabetes Before Baseline Respiration Declines

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 89 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.