Computational B-cell epitope identification and production of neutralizing murine antibodies against Atroxlysin-I

Abstract

Epitope identification is essential for developing effective antibodies that can detect and neutralize bioactive proteins. Computational prediction is a valuable and time-saving alternative for experimental identification. Current computational methods for epitope prediction are underused and undervalued due to their high false positive rate. In this work, we targeted common properties of linear B-cell epitopes identified in an individual protein class (metalloendopeptidases) and introduced an alternative method to reduce the false positive rate and increase accuracy, proposing to restrict predictive models to a single specific protein class. For this purpose, curated epitope sequences from metalloendopeptidases were transformed into frame-shifted Kmers (3 to 15 amino acid residues long). These Kmers were decomposed into a matrix of biochemical attributes and used to train a decision tree classifier. The resulting prediction model showed a lower false positive rate and greater area under the curve when compared to state-of-the-art methods. Our predictions were used for synthesizing peptides mimicking the predicted epitopes for immunization of mice. A predicted linear epitope that was previously undetected by an experimental immunoassay was able to induce neutralizing-antibody production in mice. Therefore, we present an improved prediction alternative and show that computationally identified epitopes can go undetected during experimental mapping.

Figure 1

Selection of Kmers as epitopes. The graphs illustrate how the selection of positive Kmers to be considered epitopes alters the rate of false positives based on compositional rules. As an example, Kmers of 6 and 15 aa were employed. The X-axis shows the amino acid sequence position. The Y-axis shows the probability of an amino acid residue to be a part of an epitope. Red lines represent a true epitope. Black lines represent computational prediction. (A) and (B) illustrate a prediction where at least one amino acid residue must be predicted as an epitope to label a Kmer as positive (C) and (D) show that when 50% of the amino acids must be predicted as an epitope to label a Kmer as positive (E) and (F) shows when all the amino acid residues from of a Kmer must be predicted as an epitope to label a Kmer as positive. The arrows indicate the portion of potentially false positives in each prediction method. On the right side of the figure, there is an example of prediction of an epitope marked red within the sequence SYVDLFIRETDFLSLDE by means of a 6 aa Kmer and the three approaches illustrated in the graphs.

Figure 2

Predicted and experimental epitope overlapping. The X-axis shows amino acid residue position. The Y-axis represents the experimental and predicted epitope score values from 0 to 100. Black lines represent the epitopes predicted by our model. The blue, orange, and green lines represent epitope mapping by SPOT-Immunoblotting using anti-Atr-I, anti-Bap1, and anti-Leuc-a antibodies respectively. Letters (a–c) represent the mapping of epitopes within the individual proteins, Bap1, Atr-I, and Leuc-a from Bothrops asper, B. atrox, and B. leucurus, respectively. The overlapping positions between black and colored lines represent successful predictions, while overlapping between colored lines indicates a cross-reaction.

Figure 3

ELISA tests of anti-CPEN, CNEN, and Atr-I sera. (A–C) show the boxplots results for ELISA plates coated individually by CPEN, CNEN, and Atr-I, respectively. |Plates were incubated for 1 h with the respective sera at 37 °C followed by another round of 3× washing before incubation with a respective secondary antibody for 1 h. An OPD substrate was added for ~20 minute incubation, and the reaction was stopped with H₂SO₄ prior plate reading. The black lines within the boxes correspond to medians. All the samples marked with (*) show to be significantly different for a p < 0.05, when comparing by the t test available in software R. (D) shows antibody binding over 9 doses (x-axis represents days).

Figure 4

Hemorrhagic Atr-I activity neutralization in vivo. Mice were challenged with 1 MHD (Minimal Hemorrhage Dose) of Atr-I diluted either in PBS (A) anti-CNEN serum (B) preimmune serum (4C) or anti-CPEN serum (D). Black circles indicate hemorrhagic areas in animal skin, for each treatment. (A–C) showed a clear hemorrhage area, while serum against CPEN (shown in D) was able to reduce hemorrhage, causing only skin irritation.

Figure 5

Localization of predicted epitopes in the Atr-I model. A cartoon view of the structural model of the protein Atr-I is displayed. The residues that belong to the computational prediction are shown in pink (A) while the experimental epitopes and the cross-reactive regions between the different serum samples tested are indicated in blue (anti-Atr1) and orange respectively (anti-Bap1 and anti-Leuc-a) (B,C). The overlap of these methods is presented in (D) where the black regions correspond to the matching computational and experimental predictions.