Wasserstein–Markov Random Forest (WMRF): A Machine Learning-Based Model for Accurate Protein Subcellular Localization

Motivation: Accurate prediction of protein subcellular localization (PSL) from sequence is central to cell biology and proteome-scale annotation. However, current approaches face a persistent trade-off: deep learning models often deliver strong accuracy but are computationally costly, while conventional feature-engineering methods are efficient yet often fail to capture higher-order sequential dependencies. To narrow this gap, we propose Wasserstein–Markov Random Forest (WMRF), a lightweight framework that maps each protein into a Markov transition space and quantifies its deviation from training-set consensus. Concretely, WMRF builds multi-order (1st–3rd) Markov transition profiles and augments them with four consensus-referenced scalar scores—Wasserstein distance to the mean matrix, Wasserstein distance to the median matrix, ℓ1distance to the mean matrix, and Kullback–Leibler divergence to the mean matrix—yielding compact features that are discriminative and straightforward to compute. These features are integrated with a Random Forest classifier to enable accurate and scalable PSL prediction. Results:We evaluate WMRF on four public SwissProt-derived benchmark datasets (Cytoplasm, Membrane, Mitochondrion, and Nucleus). Across all compartments, WMRF consistently outperforms representative feature-engineering baselines, including AAC, PseAAC, and HMM, achieving accuracies of 81.53%, 85.24%, 85.04%, and 90.24%, respectively. Compared with deep learning–based models, our method provides a practical accuracy–efficiency trade-off, sacrificing a small amount of accuracy to substantially reduce training time and computational cost. Moreover, ablation studies indicate that the four consensus-referenced deviation scores contribute positively to classification across all four benchmarks.