Explainable Multi-View Expert System for Alignment-Free ncRNA Classification

Overview & Abstract MV-MGB (Multi-View Meta-Gradient Boosting) is an explainable, CPU-native expert decision support system engineered for high-throughput, alignment-free functional classification and open-set diagnostic triage of non-coding ribonucleic acid (ncRNA) transcripts. Non-coding RNA functional annotation represents a core computational pillar of modern transcriptomics, molecular diagnostics, and precision oncology. However, contemporary deep foundation models and convolutional neural networks remain operationally hindered by massive multi-GPU hardware dependency, vulnerability to training-testing sequence contamination (data leakage), and uncalibrated posterior confidence that conceals misclassification risks in clinical settings. MV-MGB resolves these foundational bottlenecks by mapping primary nucleotide transcripts across four complementary biophysical and geometric spaces in strict linear computational time O(L): View 1: Sublinear-Scaled Oligonucleotide TF-IDF Spectra (35,000D): Sublinear frequency weighting of variable-length k-mers (k∈[3,8]) with L2-regularized logistic regression, capturing conserved catalytic cores and primary sequence motifs without alignment. View 2: Topological Natural Vector Moments (21D): Statistical coordinate distributions (zeroth, first, and normalized second central moments across nucleotide distributions), capturing global base-composition geometry and spatial variance via Histogram Gradient Boosted Decision Trees. View 3: Biophysical Structural Fingerprints (22D): Alignment-free heuristic structural signatures including G-quadruplex propensity scores, minimum free energy (MFE) approximations, dinucleotide purine-pyrimidine base-stacking enthalpies, and loop-entropy metrics evaluated with a 200-estimator Random Forest. View 4: Turner Stacking Thermodynamic B-Spline Density Profiles (60D): Continuous B-spline representations (B=15 knots per nucleotide) derived from empirical nearest-neighbour Turner free energies (ΔG∘), modeled by a two-layer Multi-Layer Perceptron (MLP 128×64). To prevent sparse 35,000-dimensional sequence k-mers from overwhelming compact biophysical descriptors, specialized Tier-1 based estimators project each distinct view into out-of-fold probability simplices (Δ12) via 5-fold stratified cross-validation. A Tier-2 histogram gradient-boosted decision tree (HistGBDT) integrates these 52-dimensional stacked meta-features. Finally, a Chow-type Bayesian diagnostic triage rule routes ambiguous predictions to secondary laboratory review and rejects novel or out-of-distribution (OOD) genomic decoys. Key Empirical Highlights & Benchmark Performance State-of-the-Art Accuracy on Canonical Benchmark (Dataset 1, N=2,253): Achieves 95.65% Accuracy (95% BCa CI [94.81%, 96.45%]), 0.9564 Weighted F1, 0.9577 Macro F1, and 0.9528 Matthews Correlation Coefficient (MCC), statistically outperforming retrained CNNs (nRC, ncrna-deep), GCNs, Transformers (RNA-FM), and deep foundation language models (RiNALMo). Strict Linear Time Complexity: In-memory batch inference throughput exceeds 1,162 to 1,402 sequences/second on standard consumer multi-core CPUs (0.71 to 0.86 ms/sequence), achieving 99.2% reduction in energy consumption compared to GPU-based language models. Zero Sequence Leakage: Validated via 64-bit MD5 cryptographic hash verification (MD5(Strain)∩MD5(Stest)=∅), guaranteeing complete partition independence. Biophysical Grounding: View 4 Turner thermodynamic stacking energies strongly correlate with full ViennaRNA partition function base-pairing probabilities (r=+0.3089,p<10-300). Autonomous Biological Motif Discovery: Linear weight decomposition and Fisher exact testing autonomously recover canonical sequence motifs without structural inputs: tRNA TΨC-loop (GUUCGA, log2FC = +4.31), hammerhead ribozyme core (CUGAUGAG, log2FC = +4.49), and snoRNA Box C/D motifs (UGAUG, log2FC = +2.10). Open-Set Genomic Screening: Evaluated across 249,144 transcripts spanning 177 Rfam functional families, attaining an open-set uncharacterized family rejection AUROC of 0.8427. Flow-Type Bayesian Diagnostic Decision Triage To prevent catastrophic false-positive calls in clinical diagnostics and downstream wet-lab validation, MV-MGB implements a Chow-type Bayes-optimal 3-way triage model governed by tunable confidence thresholds: Triage Action Mathematical Condition Operational Meaning in Workflow ACCEPT pmax≥τreview (default ≥ 0.85) High-confidence automated call; directly routed to automated laboratory reporting. HUMAN_REVIEW τreject≤pmax<τreview (0.50 to 0.85) Borderline ambiguity or functional overlap: routed to secondary expert verification. REJECT_NOVEL pmax<τreject (default < 0.50) Uncharacterized novel ncRNA or out-of-distribution decoy; rejected from closed taxonomy. Bundled Datasets in this Archive All benchmark and validation corpora evaluated in the manuscript are curated, verified against sequence contamination, and bundled inside the datasets/ folder: Dataset Directory Format Train Sequences Test Sequences Classes / Families Description datasets/dataset1/ FASTA, DOT, PKL 6,320 2,253 13 Classes Creux Canonical benchmark (80% CD-HIT identity ceiling, certified leak-free). datasets/dataset1-nd/ FASTA, DOT, PKL 5,966 2,127 13 Classes Non-degenerate variant strictly filtered for IUPAC ambiguity characters. datasets/dataset2/ FASTA, DOT, PKL 31,801 13,646 13 Classes Extended ncRNA benchmark corpus (45,447 total transcripts). datasets/41/ CSV 166,048 16,521 41 Families Multi-family Rfam benchmark corpus (also includes 16,542 validation sequences). datasets/177/ CSV 219,221 29,472 177 Families Open-set screening benchmark (104 in-distribution families, held-out uncharacterized). datasets/nRC-public/ FASTA 6,160 2,185 13 Classes Standard comparative corpus from convolutional RNA literature. datasets/demo/ FASTA — 14 13 Classes + 1 OOD Curated representative transcripts for instant testing of Bayesian Chow triage. Quick Start & Reproduction Commands The repository contains standalone executable bash scripts enabling one-command execution across all stages: # 1. Clone repository git clone https://github.com/BoukeliaAbdelbasset/MV-MGB.git cd MV-MGB # 2. Install dependencies pip install -r requirements.txt pip install -e . # 3. Master Pipeline Execution (Runs tests, model verification, benchmark evaluation, and triage) ./run_all.sh --quick # Fast end-to-end verification (< 30s) ./run_all.sh # Complete full pipeline reproduction # 4. Step-by-Step Runners: ./run_training.sh --dataset dataset1 # Train on Dataset 1 ./run_training.sh --dataset dataset2 # Train on Dataset 2 ./run_predict.sh --demo # Run inference on bundled demo transcripts ./run_predict.sh --sequence "GGGCCCGUAGCUCAGUGGU..." # Classify single RNA sequence ./run_predict.sh --fasta my_transcripts.fasta -o out.csv # Batch FASTA screening ./run_evaluate.sh --dataset dataset1 # Evaluate test set accuracy and MCC Archive Contents & File Manifest MVBDSource/ ├── README.md # Repository documentation and benchmark results ├── TUTORIAL.md # Comprehensive user guide and biophysical tutorial ├── tutorial_mvmgb_inference.ipynb # Interactive Jupyter Notebook Tutorial (pre-rendered) ├── LICENSE # Apache License 2.0 ├── requirements.txt # Frozen Python library dependencies ├── environment.yml # Conda virtual environment definition ├── pyproject.toml # Modern build and pytest configuration ├── setup.py # Python package setup script ├── Makefile # Unified make targets (make all, test, train, predict) ├── run_training.sh # Executable bash script: Model Training ├── run_predict.sh # Executable bash script: Inference & Diagnostic Triage ├── run_evaluate.sh # Executable bash script: Benchmark Evaluation ├── run_all.sh # Executable bash script: Master End-to-End Pipeline ├── models/ │ ├── mvmgb_dataset1.joblib # Serialized production model artifact (126 MB) │ └── model_metadata.json # Architecture parameters and registered metrics ├── datasets/ # All 7 bundled benchmark datasets (FASTA, CSV, DOT, PKL) ├── src/ # Core Python library │ ├── pipeline.py # High-level MVMGBClassifier API │ ├── canonicalization.py # Transcript canonicalisation and 64-bit MD5 hashing │ ├── data_loader.py # Multi-format sequence and label loaders (FASTA / CSV) │ ├── config.py # Global seeds and automated path discovery │ ├── views/ # Four feature space encoders (TF-IDF, NV, BSF, BSP) │ └── models/ # Multi-view stacking engine & Chow triage rule ├── scripts/ # Python CLI entry points and mirrored bash runners └── tests/ # Automated pytest suite (leakage, views, inference) Scientific Metadata & Authorship Title: MV-MGB: Multi-View Meta-Gradient Boosting for Non-Coding RNA Classification Authors: Rabah Lebsir (Department of Computer Science, Université 8 Mai 1945 Guelma, Algeria; Email: [e-mail]) Abdelbasset Boukelia (Centre de Recherche en Biotechnologie (CRBT), Constantine, Algeria; Email: [e-mail]; ORCID: 0000-0003-1827-212X) Persistent Record DOI: 10.5281/zenodo.23724491 Code Repository: https://github.com/BoukeliaAbdelbasset/MV-MGB Target Journal: Expert Systems with Applications (ESWA), Elsevier Science License: Apache License 2.0 (Code) / Creative Commons Attribution 4.0 International (Datasets & Documentation) Citation @article{lebsir2026mvmgb, title={An Explainable Multi-View Expert Decision Support System for Alignment-Free Non-Coding RNA Classification Using Biophysical and Geometric Feature Spaces} note={Code and data available at https://github.com/BoukeliaAbdelbasset/MV-MGB, Zenodo DOI: 10.5281/zenodo.23724491} }