Abstract
Bacterial secreted proteins, particularly effectors delivered by specialized secretion systems, are key mediators of virulence and host-pathogen interactions. However, accurate computational identification remains challenging, as many existing methods rely heavily on sequence similarity or handcrafted features, and often focus on a single secretion system. Recent studies have reported that some bacterial effectors may be associated with more than one secretion system, highlighting the complexity of secretion system annotation and motivating the development of system-aware computational prediction approaches. Here, we present PLM-Effector, a hybrid deep learning framework that integrates modern protein language models (PLMs) with multiple neural architectures via a two-layer ensemble stacking strategy. By extracting complementary features from N- and C-terminal regions, PLM-Effector enables secretion-type-aware prediction across five major bacterial secretion systems (T1SS-T4SS and T6SS), with each system modeled independently. Systematic benchmarking shows that embeddings from protein-specific PLMs (ESM-1b, ESM2_t33, ProtT5) are more discriminative than those from general-purpose language models (e.g. BERT, BioBERT). Leveraging these representations, PLM-Effector achieves superior performance on an independent test set, with macro F1-scores of 0.9848, 0.8649, 0.9899, 0.9620, and 0.9728 for secreted proteins of T1SS-T4SS and T6SS, respectively, outperforming existing tools and homology-based baselines. Implemented as an accessible web server (http://www.mgc.ac.cn/PLM-Effector/) with source code and datasets available (https://github.com/zhengdd0422/PLM-Effector/), PLM-Effector provides a reproducible and user-friendly platform for both small-scale and genome-wide secreted protein discovery, facilitating advances in the study of bacterial pathogenesis.