Machine Learning Consensus Tool Identifies Reproducible Biomarkers Across Heterogeneous Transcriptomic Cohorts

High-throughput transcriptomic profiling has long promised to decode the molecular architecture of complex human pathologies, yet translating high-dimensional RNA sequencing data into stable, clinically actionable biomarker signatures remains hampered by pervasive cohort heterogeneity and model-specific algorithmic biases. Conventional analytical pipelines typically rely on a rigid two-step framework: first filtering candidate genes based on differential expression thresholds, and subsequently training a single classification algorithm on the preselected subset. This established paradigm exhibits several well-documented vulnerabilities. Differential expression thresholds evaluate transcripts in isolation, discarding complex multivariate dependencies, non-linear interactions, and subtle coordinated shifts across biologically coherent pathways. Furthermore, training isolated classification models leaves feature selection susceptible to idiosyncratic mathematical assumptions, yielding signatures that achieve high accuracy on localized training data but experience steep performance degradation when applied across external cohorts. When supervised models are fitted to features preselected using class labels from the same dataset, the process also introduces risks of data leakage, wherein apparent predictive utility reflects label rediscovery rather than generalizable biological signal.

To overcome these structural limitations, an international research team developed TENTACLES, an acronym for Transcriptomic Exploration Tool through Aggregation of Classifiers, released as an open-source R package engineered to systematically prioritize stable biomarkers through algorithmic consensus. Designed as a disease-agnostic, modular computational architecture, TENTACLES integrates up to 15 supervised machine learning classifiers and six unsupervised clustering methods to identify gene panels that demonstrate cross-cohort generalizability. Rather than staking discovery on the variable importance metrics of any single algorithmic architecture, the framework automates library normalization, batch correction, multi-model hyperparameter tuning, cross-algorithm feature importance aggregation, and rigorous downstream validation across multiple independent datasets. By aggregating variable importance rankings across diverse model families—ranging from linear models sensitive to additive effects to non-parametric tree ensembles capturing complex feature interactions—the platform extracts biological signals that resist technical noise and platform-dependent variance.

The end-to-end analytical pipeline of TENTACLES is organized into five sequential, core computational modules that guide high-dimensional count matrices through progressive refinement. Data processing initiates within the preProcess function, which applies library-size normalization via edgeR and executes batch-effect correction via the sva package when integrating disparate datasets, followed by exploratory variance verification using Principal Variance Component Analysis. The runClassifiers module subsequently optimizes and trains supervised classifiers simultaneously across a uniform cross-validation scheme, pairing specific feature selection strategies with designated classifiers and computing model-native variable importance metrics. The getConsensus function aggregates these importance outputs across fitted models, retaining only those candidate transcripts designated as important by a user-defined threshold of participating algorithms. Downstream properties of the consensus panel are evaluated through testConsensus, which executes dimensionality reduction, fold-change calculations, receiver operating characteristic analyses, and within-cohort neural network testing. Finally, valConsensus systematically computes and screens all non-empty gene subsets across six unsupervised clustering algorithms to isolate the most compact gene combination that maximizes class separability.

To rigorously test the capabilities of TENTACLES under demanding real-world conditions, the investigators deployed the framework against Crohn’s disease as a high-heterogeneity case study. Crohn’s disease represents a chronic, relapsing inflammatory bowel disorder characterized by transmural, segmental inflammation that can manifest discontinuously throughout any segment of the gastrointestinal tract. Marked clinical variability, ranging from intermittent abdominal discomfort to stricturing and penetrating complications requiring surgical resection in up to seventy percent of patients, creates substantial transcriptomic noise across independent studies. The validation strategy marshaled 689 intestinal biopsy profiles from four publicly available bulk RNA sequencing cohorts archived within the NCBI Sequence Read Archive. Analysis was focused on approximately 10,000 biologically interpretable transcripts annotated within Gene Ontology or Kyoto Encyclopedia of Genes and Genomes databases, filtering out low-abundance features by enforcing a threshold of counts per million exceeding one in at least ten percent of analyzed samples.

The initial ensemble-based discovery phase examined Dataset 1, comprising 218 Crohn’s disease cases and 42 healthy controls under BioProject accession PRJNA248469. Ten supervised classifiers spanning distinct mathematical families were trained using five-fold cross-validation: Random Forest, linear Support Vector Machines, Multi-Layer Perceptrons, Bagged Multi-Layer Perceptrons, C5-rules, LightGBM, Decision Trees, Multivariate Adaptive Regression Splines, Extreme Gradient Boosting, and Partial Least Squares. Hyperparameters were independently tuned across five candidate configurations per fold, selecting optimal models via held-out F1-score. Nine of the ten classifiers demonstrated consistent predictive performance, recording a median cross-validated accuracy of 0.80 and an F1-score of 0.60, which advanced to apparent values of 0.87 and 0.70 upon full-dataset refitting. Linear Support Vector Machines achieved the highest cross-validated metrics, with an accuracy of 0.85 and an F1-score of 0.67. Conversely, Extreme Gradient Boosting exhibited markedly poor stability under the standardized tuning grid, recording an accuracy of 0.55 and an F1-score of 0.30, and was consequently excluded from consensus formation to preserve methodological symmetry across the ensemble.

Feature importance scores from the nine high-performing classifiers revealed structured convergence across algorithmic paradigms. Random Forest prioritized the largest individual gene set with 415 features, while Multi-Layer Perceptrons shared an exclusive intersection of 265 features. To isolate a core predictive panel, the getConsensus module extracted genes recognized as important by at least six of the nine algorithms, establishing a strict majority threshold that yielded a refined 28-gene consensus signature. Crucially, CXCL8 and OSM emerged as the most recurrently selected transcripts across the ensemble, prioritized by eight and nine classifiers, respectively. These two inflammatory drivers exhibited strong coordinated expression within Crohn’s disease biopsy samples, recording a determination coefficient of 0.87 compared to 0.55 in healthy controls. When the 28-gene panel was evaluated on a held-out thirty percent test split of Dataset 1 using a within-cohort Multi-Layer Perceptron, the model recorded an accuracy of 0.962, an area under the receiver operating characteristic curve of 0.990, and a Brier score of 0.111.

The consensus panel was subsequently subjected to cross-cohort evaluation on Dataset 2, an external merged cohort comprising 134 Crohn’s disease patients and 45 healthy controls from BioProjects PRJNA565216 and PRJNA985602. Batch-effect correction preserved biological variance, and a local Multi-Layer Perceptron trained on Dataset 2 achieved an accuracy of 0.80, an area under the curve of 0.821, and a Brier score of 0.18. Systematic benchmarking against single-classifier feature sets and conventional differential expression panels demonstrated the clear superiority of the consensus approach. While a Decision Tree signature achieved an isolated area under the curve of 0.87, its accuracy dropped to 0.74; conversely, the 28-gene TENTACLES signature achieved the highest accuracy across all comparators and the highest composite aggregate score of 2.44 across accuracy, discrimination, and calibrated probability metrics. Concurrently, standard differential expression analysis exposed acute cohort instability: of 3,687 differentially expressed genes identified in Dataset 2, only 368 overlapped with the 582 differentially expressed genes from Dataset 1, leaving approximately ninety percent unshared. In contrast, 17 of the 28 consensus genes, representing roughly 61 percent, proved to be reproducible differential markers across both cohorts while reducing feature dimensionality by ninety-five percent relative to Dataset 1 differential gene lists.

To condense the biomarker panel for clinical translation, the investigators examined the directional concordance of feature importances derived via the Olden algorithm between the Dataset 1 and Dataset 2 Multi-Layer Perceptron models. Exactly 14 of the 28 consensus genes displayed identical directional contributions across both independent patient cohorts: BACE2, HK2, HCAR2, OSM, FCGR1A, IL15RA, HK3, ELL2, BATF2, RPA1, C6, FCER1A, ADRA1B, and CSF3R. This refined 14-gene panel was carried forward into Dataset 3, an independent cohort of 34 Crohn’s disease patients and 42 healthy controls under accession PRJNA702434. The valConsensus module exhaustively assessed every non-empty combinatorial subset of the 14 genes across six unsupervised clustering algorithms: Gaussian Mixture Models, Hierarchical Clustering, standard K-Means, and K-Means applied across Principal Component Analysis, Uniform Manifold Approximation and Projection, and t-distributed Stochastic Neighbor Embedding. Combinations were ranked according to mean unsupervised F1-score, which isolated a definitive minimal core signature of five genes—CSF3R, OSM, BACE2, HK2, and IL15RA—achieving an accuracy of 0.77, an F1-score of 0.72, precision of 0.70, and recall of 0.79.

The ultimate test of generalizability examined whether this pre-specified 5-gene core could stratify patient samples in a fourth, completely independent cohort without any model refitting or supervised optimization. Dataset 4 merged accessions PRJNA728646 and PRJNA961704, encompassing 109 Crohn’s disease samples and 65 healthy controls. Restricting the count matrix solely to CSF3R, OSM, BACE2, HK2, and IL15RA, the researchers conducted two-dimensional t-distributed Stochastic Neighbor Embedding followed by unsupervised two-cluster K-Means partitioning. With class labels applied exclusively a posteriori for objective performance verification, the 5-gene panel achieved an unsupervised classification accuracy of 0.72, precision of 0.82, recall of 0.71, and an F1-score of 0.76. Biologically, each transcript in the minimal core maps directly to established pathophysiological drivers: OSM encodes a cytokine driving intestinal epithelial barrier disruption; HK2 governs critical glycolytic reprogramming during tissue inflammation; IL15RA regulates mucosal lymphocyte survival; BACE2 acts as an emerging mucosal biomarker; and CSF3R orchestrates neutrophil homeostasis and recruitment.

Computational profiling conducted on a conventional desktop workstation equipped with an Intel Core i5 processor and 32 gigabytes of random access memory confirmed that TENTACLES executes efficiently without requiring high-performance supercomputing infrastructure, though combinatorial subset searches within valConsensus scale with feature count. The authors acknowledged notable study constraints, including the primary validation being restricted to Crohn’s disease bulk intestinal biopsies, reliance on pre-annotated pathway genes that excluded unannotated non-coding RNAs, and the retrospective nature of public data repositories. Residual confounding from patient demographic differences, variable tissue depths, and medication histories cannot be entirely ruled out. Nonetheless, because the underlying software architecture functions in a completely disease-agnostic manner, TENTACLES establishes a reproducible, scalable framework capable of extracting compact, transferable biomarker panels across high-dimensional transcriptomic datasets in autoimmune, inflammatory, and oncological research.

Subject of Research: Consensus machine learning framework for transcriptomic biomarker discovery across heterogeneous cohorts

Article Title: TENTACLES: a consensus machine learning tool for robust biomarker discovery in heterogeneous data

Article References: Montesi, G., Mouta, G. D. S., Novedrati, M., Cunha, A. F., Fuschi, A., Lucchesi, S., Sonnati, C., Ciabattini, A., Santoro, F., Medaglini, D., & Nakaya, H. I. (2026). TENTACLES: a consensus machine learning tool for robust biomarker discovery in heterogeneous data. BioData Mining, 19(1), Article 64. https://doi.org/10.1186/s13040-026-00568-8

Image Credits: AI Generated

DOI: 10.1186/s13040-026-00568-8

Keywords: TENTACLES, machine learning, transcriptomics, biomarker discovery, Crohn's disease, ensemble learning, RNA sequencing, feature prioritization, gene expression, supervised classification, unsupervised clustering, computational biology, cross-cohort validation, reproducibility, intestinal biopsies, variable importance