MoCETSE

MoCETSE is a multi-class classifier that identifies secreted effector proteins of Gram-negative bacteria from amino acid sequence alone. Pathogens such as Pseudomonas, Legionella, and Salmonella inject effectors into host cells through dedicated secretion systems, and those effectors are the direct agents of virulence — making their identification a prerequisite for understanding pathogenesis and for nominating antimicrobial targets. Effectors are notoriously hard to recognize computationally: the labeled datasets are small, secretion signals are diffuse rather than motif-like, and the proteins share little structural conservation across species.

The model was developed by Hua Shi and colleagues at Xiamen University of Technology, with collaborators at the University of Electronic Science and Technology of China, and posted to bioRxiv in August 2025. Its argument is that feeding protein language model embeddings straight into a shallow classifier wastes them: the 1,280-dimensional per-residue representations from ESM-1b carry redundant, general-purpose signal that a task-specific network should first compress and re-focus. MoCETSE inserts two such stages — a mixture-of-convolutional-experts preprocessing network and a transformer with relative positional encoding — between the frozen language model and the classification head.

MoCETSE assigns each input to one of six classes: non-effector, or type I, II, III, IV, or VI secreted effector. That single multi-class formulation replaces the usual practice of running a separate binary predictor per secretion system.

#Key Features

• Frozen protein language model front end: ESM-1b weights are taken unchanged from the published checkpoint and are not fine-tuned, so only the downstream modules are trained.
• Mixture-of-convolutional experts: A gating network softly weights several parallel convolutional experts with different kernel sizes, reducing embedding dimensionality while capturing local motifs and broader context simultaneously.
• Relative positional encoding: The transformer block models distances between residues rather than absolute positions, targeting the long-range dependencies that carry secretion signals.
• Six-way classification: Discriminates non-effectors and five effector types in one model, rather than requiring one binary classifier per secretion system.
• Genome-scale screening: Pairs with MacSyFinder for secretion-system detection to run whole-genome effector scans.

#Technical Details

MoCETSE composes three stages: ESM-1b (a 33-layer transformer pretrained on UniRef50, emitting 1,280-dimensional residue embeddings) produces the input representation; a target preprocessing network built on a mixture-of-experts architecture with a softmax gating module reduces and re-weights it; and a relative-position multi-head attention transformer feeds a softmax classification head. Training and independent test data were inherited from DeepSecE, filtered with CD-HIT at 60% identity, with a further 90% CD-HIT pass to remove overlap with the external benchmarks. The training set holds 2,904 sequences (1,577 non-effectors, 128 T1SE, 68 T2SE, 392 T3SE, 507 T4SE, 232 T6SE) and the independent test set holds 260. Training ran in PyTorch on an A100 with Adam, a 5e-5 learning rate, batch size 32, and early stopping on F1.

Under five-fold cross-validation MoCETSE reached 0.878 accuracy and 0.850 F1, rising to 0.905 accuracy and 0.867 F1 on the independent test set; per-class AUCs ranged from 0.930 to 0.992. On external benchmarks curated by other groups it reached 91.7% accuracy and 0.835 MCC on the Bastion3 T3SE set, 98.3% accuracy and 0.939 MCC on the CNN-T4SE T4SE set (above DeepSecE's 97.8%), and 98.6% accuracy and 0.920 MCC on the Bastion6 T6SE set. Ablations attribute the gains to both new modules, most visibly for the sparsely sampled T2SE class, whose F1 rose from 0.550 to 0.636. Genome-wide screens of P. syringae DC3000, L. pneumophila Philadelphia 1, and P. aeruginosa PAO1 recovered validated effectors with higher precision or recall than DeepSecE.

#Applications

MoCETSE is a screening tool for microbiology and infection-biology labs: given a newly sequenced Gram-negative genome, it flags candidate effectors and assigns them to a secretion system, narrowing a proteome-scale list to a shortlist worth testing by translocation assay. Because it runs directly on sequence and needs no position-specific scoring matrices, it avoids the expensive PSSM generation that older tools such as Bastion3 and Bastion6 depend on, which is what makes whole-genome inference practical.

#Impact

MoCETSE illustrates a pattern now common across bioinformatics: a large frozen protein language model supplying representations, with a compact task-specific network learning what to keep. Its contribution is showing that this intermediate refinement measurably helps on small, imbalanced effector datasets where prior work fed embeddings directly to a classifier. The work was subsequently published in PLOS Computational Biology. Code, trained weights, and example data are available on GitHub, though the repository carries no license file, which limits reuse. Remaining limitations are those of the underlying data: T2SE remains the weakest class because only 68 training examples exist, and genome-wide precision is low in absolute terms, so outputs are candidate lists rather than confirmed effectors.