<p>Predicting the subcellular localization of multi-site bacterial proteins remains challenging because label correlations, limited sample size, and low sequence similarity reduce the effectiveness of conventional multi-label classifiers. In this study, we propose a multi-label cascade flexible neural forest (MLCFN Forest) that combines label-powerset-style coding–classification–decoding with a cascade ensemble of flexible neural tree (FNT) groups. The framework preserves label dependencies, decomposes the induced multi-class task into coordinated binary FNT outputs, and progressively reallocates model capacity to low-confidence samples through confidence-guided sample propagation and feature enhancement across layers. We explicitly note that the cascade FNT backbone builds on our previous CFNForest studies for cancer subtype classification, whereas the present work adapts that backbone to multi-label bacterial protein localization by introducing label-set encoding/decoding, multiclass FNT grouping, and a multi-site protein subcellular localization evaluation workflow. Using Gram-negative and Gram-positive benchmark datasets together with AECA-PSSM and PSSM-DWT features, MLCFN Forest consistently outperformed ML-RBF, ML-KNN, ML-LOC, INSDIF, and MLASSO under jackknife evaluation. For example, it achieved OLA/OAA values of 78.3%/76.4% on the Gram-negative AECA-PSSM dataset, 80.7%/78.5% on the Gram-negative PSSM-DWT dataset, and 80.2%/77.6% on the Gram-positive AECA-PSSM dataset. PCA-based low-dimensional experiments further suggested that the model retained good predictive ability after substantial dimensionality reduction. The present study is limited to classical benchmark datasets and does not yet include an independent external test set, strict nested model selection, or direct benchmarking against protein language model predictors; these points are therefore discussed as limitations and future directions.</p>