Exo-Tox: Identifying Exotoxins from secreted bacterial proteins

Abstract

Background: Bacterial exotoxins are secreted proteins able to affect target cells, and associated with diseases. Their accurate identification can enhance drug discovery and ensure the safety of bacteria-based medical applications. However, current toxin predictors prioritize broad coverage by mixing toxins from multiple biological kingdoms and diverse control sets. This general approach has proven sub-optimal for identifying niche toxins, such as bacterial exotoxins. Recent Protein Language Models offer an opportunity to improve toxin prediction by capturing global sequence context and biochemical properties from protein sequences.

Results: We introduce Exo-Tox, a specialized predictor trained exclusively on curated datasets of bacterial exotoxins and secreted non-toxic bacterial proteins, represented as embeddings by Protein Language Models. Compared to Basic Local Alignment Search Tool (BLAST)-based methods and generalized toxin predictors, Exo-Tox outperforms across multiple metrics, achieving a Matthews correlation coefficient > 0.9. Notably, Exo-Tox's performance remains robust regardless of protein length or the presence of signal peptides. We analyze its limited transferability to bacteriophage proteins and non-secreted proteins.

Conclusion: Exo-Tox reliably identifies bacterial exotoxins, filling a niche overlooked by generalized predictors. Our findings highlight the importance of domain-specific training data and emphasize that specialized predictors are necessary for accurate classification. We provide open access to the model, training data, and usage guidelines via the LMU Munich Open Data repository.

Keywords: Bacterial; Embeddings; Exotoxins; Predictor; Proteins; Toxins.

Fig. 1

Embeddings information separates better toxins from non-toxins than aac. A 2D Principal Component Analysis(PCA) projection of amino acid composition (aac) (B) and per protein embeddings by ProtT5. The aac was scaled before the PCA was carried out, embeddings were not scaled (see Material and methods). Bacterial exotoxins (red) and secreted bacterial non-toxins (blue) form overlapping clusters

Fig. 2

Toxicity of secreted proteins predicted accurately from embeddings. Different exotoxins prediction methods are compared. This includes three sets of input features. Methods that use the amino acid composition as input are depicted in shades of red. Methods using the first 20 Principle Components that retain the most information from Pseudo Amino Acid Composition are shown in hues of yellow. Methods using the first 20 Principle Components that retain the most information from ProtT5 protein embeddings are in shades of blue. The different input features approaches are compared to BLAST and Foldseek as a baselines (gray lines) and CSM-Toxin [47] and MultiToxPred 1.0 [48], two state of the art generalized toxin predictor not specialized on bacterial proteins (green). Data: hold-out test set of bacterial exotoxins and bacterial, secreted non-toxins with less than 30% sequences similarity to training set of CSM-Toxin and training set of our proposed predictor. Metric: Matthew’s correlation coefficient (MCC). Model architectures: kNN: K-Nearest Neighbors, LR: Logistic Regression, SVC: Support Vector Classifier, RF: Random Forest and XGB: Extreme Gradient Boosting. The model architectures are differentiated by color intensity. From light to dark: kNN, LR, SVC, RF, and XGB. Black whiskers mark the 95% interval with ± the 1.96 the standard error

Fig. 3

Signal Peptides do not influence toxin prediction Panel A: Exploration of signal peptides predicted with SignalP-6.0. Bacterial exotoxins are in red, the bacterial secreted non-toxins are in blue. Signal Peptides distinguish between two translocation routs Sec and Tat and three Signal Peptidases SPI-III. Prediction include SP: Sec/SPI, LIPO: Sec/SPII, TAT: Tat/SPI, LIPOTAT: Tat/SPII, PILIN: Sec/SPIII and OTHER indicates no known signal peptides. The majority of exotoxins does not have a predicted signal peptide. Panel B: Performance comparison with and without signal Peptides. Model architecture was introduced in Fig. 2 Support Vector Classifier (SVC) using the first 20 Principal Components calculated on per protein protT5 embeddings (Embs20). Two versions of the test set are compared. Light blue are the original test set sequences. Dark blue: the test sequences without the predicted signal peptides. Embs20/SVC performs equally well on both test set versions

Fig. 4

Several proteins are identified as toxins by both predictors, but not all. Size of the sets is shown in the lower left corner. The intersection of each set’s prediction are shown by the dots and bars located below each of the bars on the bar plot above. Four sets were tested 1) secreted: secreted bacterial proteins, 2) exotoxins: bacterial exotoxins, 3) controls: bacterial proteins independent of secretion status and 4) phages: phage proteins not containing prophages. Two predictors were tested 1) aac: was trained on amino acid composition and 2) Embs20: was trained on the first 20 Principle Components of per-protein embeddings