Table 1.
A summary of bioinformatics tools for conformational and linear BCE prediction.
Fig 1.
Schematic representation of the study workflow.
(A) Overview of the proposed framework, highlighting a comparative architecture between the Transformer-based model (upper section) and the traditional machine learning model (lower section). (B) A detailed depiction of the Transformer-based model architecture, emphasizing its structure comprising 2 encoder layers and 8 attention heads. The model processes input amino acid sequences via an embedding layer, positional encoding, multi-head self-attention, and feature optimization modules, before classifying the sequences through a fully connected layer.
Table 2.
Performance of the Transformer-based deep learning model for predicting BCEs in parasites across training and independent test datasets.
Fig 2.
Comparison of deep learning feature (deepFeature) with eight top traditional handcrafted features (AAC, ASDC, CKSAAGP, CKSAAP, DDE, DPC, GDPC, and GTPC).
(A-C) Radar plots showing SP, SE, and MCC values for each feature. (D-F) ROC curve comparing the deep learning model (deepBCE) with four traditional machine learning algorithms (GNB, LGBM, RF, and SVM) on CKSAAP, DDE, and DPC features. The x-axis represents False Positive Rate (FPR), and the y-axis represents True Positive Rate (TPR).
Table 3.
Performance comparison of our models against existing state-of-the-art models.
Fig 3.
Performance comparison between our models and existing models on benchmark datasets.
(A) Results on Test1 dataset. (B) Results on Test2 dataset.
Fig 4.
Proteomic and bioinformatics analysis of F. hepatica across four developmental stages.
(A) The lifecycle of Fasciola spp., demonstrating the developmental stages within the intermediate snail host, environmental phases, and definitive mammalian/human host. The parasite image was adapted from Servier Medical Art (https://smart.servier.com/). (B) Principal component analysis (PCA) of protein expression profiles in metacercaria, juvenile fluke (28 dpi), immature fluke (58 dpi), and adult fluke (118 dpi). (C) Distribution of subcellular localization for proteins capable of generating B-cell epitopes (BCEs), highlighting their potential immunogenic properties. (D) Heatmap depicting protein expression dynamics across the four stages, with color-coded subcellular localization categories shown on the right. (E) Protein-protein interaction network of the leucine aminopeptidase (LAP) protein, generated using the STRING database, emphasizing key interactions and its potential as a vaccine candidate.
Fig 5.
Predicted 3D structural models of potential F. hepatica vaccine molecules, highlighting the spatial distribution of putative BCEs.
(A–D) Structural models for Glutathione transferase, leucine aminopeptidase, annexin, and β-actin, respectively. The predicted template modelling (pTM) score from AlphaFold3, which reflects prediction quality, is provided in the legend (higher value indicates greater model reliability). Predicted functional domains are color-coded, while non-domain regions are shown in gray.
Fig 6.
Sequence alignment and experimental validation of the LAP protein and its putative BCEs.
(A) Multiple sequence alignment of F. hepatica LAP with five other trematode species. Predicted BCEs identified by our AI models are highlighted with red boxes. (B) Dot-blot immunoassay validation of predicted BCEs using serum from F. hepatica-positive animal, with F. hepatica-negative serum and a synthetic human peptide as controls.
Table 4.
Experimental validation of predicted BCEs derived from the LAP protein using dot-blot immunoassay.