DECODE enables high-throughput mapping of antibody epitopes at single amino acid resolution

To achieve high-throughput and detailed epitope analysis, we developed the DECODE method, which consists of an improved mRNA display method and GPU-based next-generation sequencing (NGS) analysis. This chapter provides details about the optimized mRNA display method. Conventional mRNA display methods have low yields of products from each reaction step, including transcription, ligation, and translation so that these methods require multiple purification steps to remove unreacted regents and change to the appropriate reaction buffer, which hinder improvements in throughput [17,18,20]. We improved the buffer composition and reaction conditions for all these steps and optimized them to improve product yields as shown below.

Initially, a template DNA library was designed using a random sequence of 12 amino acids to sufficiently accommodate the nearly linear epitope length and using NNK to reduce the frequency of stop codons (S1A Fig). Our significant improvement was the enhanced ligation efficiency between mRNA and puromycin-conjugated DNA (pu-DNA). With conventional methods, this ligation efficiency was low, requiring electrophoresis and purification, which we considered to be an obstacle to achieving high-throughput selection and a cause of decreased library yield. We investigated the cause of the low ligation efficiency between mRNA and pu-DNA. We confirmed the ligation between mRNA and pu-DNA using polyacrylamide gel electrophoresis with single-nucleotide resolution and found that approximately half of the mRNA had one base added during the transcription reaction, which interfered with ligation (S1B Fig). Previous papers have shown that base addition occurs due to run-off of T7 RNA polymerase [21]. To reduce base addition by T7 RNA polymerase, we used template DNA with two 2′-O-methyl-guanosine (Gm) residues introduced at the 5′ end of the complementary strand. As a result, the transcript ligation efficiency increased from 53% to 91% (Figs 1B, S1B, and S1C).

Next, we examined the ligation efficiency of mRNA and peptide depending on the type of ligation approach. When we compared the linear ligation approach with the hairpin-format ligation approach, we found that linear ligation using Gm-containing templates had the highest ligation efficiency (S1D–S1F Fig). Regarding the recovery of peptide selection, linear ligation also showed higher recovery than hairpin ligation (S1G Fig). Additionally, we were concerned about the degradation of the mRNA-peptide complex and used the RNase- and Protease-free PURE system for in vitro translation [22] and performed the binding reaction to targets before reverse transcription. As described above, we succeeded in omitting the electrophoresis and purification steps by reducing the amounts of unreacted products in each step.

To demonstrate this optimized protocol for epitope analysis, peptide selection was performed to the anti-FLAG antibody (clone M2) as a target. It has already been reported that the epitope motif recognized by the anti-FLAG antibody is DYKXXD [20,23]. The NGS results showed that the proportion of the known epitope motif DYKXXD increased through each round of peptide selection by the DECODE method, reaching over 95% in the third round (Fig 1C). This result indicates that the peptide enrichment efficiency of the DECODE method is almost the same as that of conventional methods, which require the gel electrophoresis purification of mRNA and Pu-DNA complexes. This demonstrates that the DECODE method can achieve higher throughput without the laborious electrophoresis and purification step, while maintaining peptide enrichment efficiency. Sequence logos were retrieved from the top 1,000 peptide sequences of the NGS data by WebLogo [24–26] (Fig 1D). These results indicated that the new improved mRNA display method can sufficiently recover the epitope peptides of the antibody. This protocol is designed so that the reaction solution composition at each step does not affect the next reaction, allowing the process to proceed with simple pipetting. Ultimately, it will be possible to perform peptide selection targeting at least approximately 100 individual antibodies per week using multiwell plates, increasing the throughput by more than 100 times using existing methods.

To predict the specificity and genome-wide cross-reactivity of a large number of antibodies from the NGS results obtained by DECODE, we developed a GPU-based analysis algorithm with reference to previous study [27]. First, we used the BLOSUM62 table, which is a homology score matrix between amino acids commonly used to compare amino acid sequences, for quantitative evaluation of sequence similarity, and the negative values were modified to zero because DECODE selection used the randomized library (S2L Fig). Next, we calculated a frequency list of peptides which had 20¹² random amino acids generated from an ideally random cDNA library (4³⁶ bases) and derived sequence similarities (similarity scores) between this list of peptides and all positions of the protein referenced from the UniProt database (S2A Fig). The theoretical inverse cumulative probability distribution (P) map was prepared from the similarity scores and the peptide frequency list (Fig 2A, top). The experimental inverse cumulative probability distribution (Q) maps were similarly calculated from the cDNA sequences of the DECODE selections with 100,000 to 1 million reads from the NGS results (Fig 2A, bottom and S2B Fig). Fig 2 shows the analysis results of the anti-c-Fos antibody (clone 9F6) as an example. In comparison with the P-map, the similarity scores around the binding site of Q-map were increased (Fig 2B and 2C). To quantify this difference, we explored various formulas commonly used in machine learning to calculate the distances between various distributions. As a result, the Pearson Χ2 formula showed the highest distance around the high similarity score area, while the low similarity area showed the smallest distance (Figs 2D and S2C). These results indicated that the presence of more similar sequences could be detected with high sensitivity. The sum of the distances of similarity score for each protein position between the P and the Q was named the DECODE score. The DECODE scores at all positions of the c-Fos protein (Fig 2E) showed a remarkably high peak around the binding site. We also plotted the DECODE scores of all mouse proteins (Fig 2F), and the highest DECODE score was found for the c-Fos protein. The top 100 DECODE scores are plotted in Fig 2G, which shows that only a portion of the c-Fos protein had a high DECODE score. These results suggest that clone 9F6 may have high target specificity. Other clones of the c-Fos antibody were also analyzed, and the DECODE scores of the c-Fos protein were visualized, showing a peak for each antibody (Figs 2H and S2D–S2H). DECODE score calculations were performed on the GPU, which was coded using CUDA C/C++ to perform the calculations faster for multiple NGS results, and 1 million lead NGS results were achieved in 3 min. DECODE scores were also calculated using the WAC table, another amino acid similarity score table, which has a maximum value of 4 and a narrower range of similarity scores that enabled a reduction in computation time, and similar results were obtained (S2I–S2K and S2M Fig). The WAC table could be used as an alternative to BLOSUM62 if there is a need to reduce the computational time. As above, this analysis method has enabled a comprehensive estimation of the protein, site, and amino acid to which an antibody binds among all proteins across the genome.

Fig 2. A GPU-based deep epitope analysis in DECODE using the genome-wide protein database.

(A) Calculation flows of the theoretical inverse cumulative probability distribution (P) between the protein database and all possible peptides from the random DNA library (upper) and experimental inverse cumulative probability distribution (Q) between the protein database and screened peptides by DECODE selection (lower). (B) Heat map of the inverse cumulative probability distribution of each position of the mouse c-fos protein. Theoretical distribution (upper) or experimental distribution by DECODE selection for anti-c-fos antibody clone 9F6 (bottom). (C) Inverse cumulative probability distribution plots on the non-binding site (left) and the binding site (right) of anti-c-fos antibody clone 9F6. (D) Plots of the distances at each similarity score between P and Q on the binding site are calculated by various distance functions. (E) A bar graph of the DECODE score on the mouse c-fos protein sequence for anti-c-fos antibody clone 9F6. (F) Manhattan plot of all mouse proteins for anti-c-fos antibody clone 9F6. Red dots and highlights indicate the c-fos protein. This plot was visualized with downsampled data to 1/50 using LTTB [51]. (G) The top 100 DECODE score showed in (F). Red dots indicate the c-fos protein. (H) Epitopes map of 4 monoclonal anti-c-fos antibodies against the c-fos protein. These DECODE scores were normalized by the ratio with the means on c-fos protein. The data underlying for panels A–H shown in the figure can be found in S1 Data or https://doi.org/10.5281/zenodo.14286317. DECODE, Decoding Epitope Composition by Optimized-mRNA-display, Data analysis, and Expression sequencing; LTTB, Largest Triangle Three Buckets.

https://doi.org/10.1371/journal.pbio.3002707.g002

Epitope analysis of anti-c-Fos antibodies (clones 2H2, 9F6, and C-10) was conducted using the DECODE method to identify the sites where antibodies bind and hotspot residues (Fig 3A). Sequence logos were visualized only for amino acids that converged by 20% or more. For all antibodies, the same epitope was reproducibly obtained in 3 independent peptide selection (S3A Fig). The analysis results also showed that each antibody specifically recognized a different site on the c-fos protein (S3B Fig), and there was no correlation between clones (S3C Fig). We performed the ELISA to verify whether the predicted epitopes are actually specifically bound by the 3 anti-c-Fos antibodies. First, we chemically synthesized the identified epitope peptides, which were the 12 consecutive amino acid sequence that had the highest DECODE score in the target protein sequences including several residues before and after, and investigated their binding reactions with each antibody. The results showed that each antibody specifically recognized the sequences identified by DECODE (S3D Fig). We next performed ELISA using a double mutant c-Fos protein with 2 predicted hotspot residues mutated (Figs 3B, S3E and S3F). As a result, each antibody bound to the wild type but not to the mutant c-Fos, suggesting that the sites where antibodies bind was accurately predicted by DECODE. Also, when antibodies with different epitopes were used, there was a binding reaction for both the wild type and the mutant (S3G and S3H Fig). These results supported that each antibody actually binds specifically to the site identified by DECODE. Furthermore, to verify the accuracy of hotspot residues identified by DECODE, we performed competitive ELISA against c-Fos protein using epitope peptides with single amino acid mutations (Figs 3C, S4A and S4B). We confirmed a correlation between the amount of change in DECODE score due to mutations and the IC50 calculated as a result of competitive ELISA (Fig 3D). These correlations suggested that peptides with mutations in the hotspot residue significantly bind to antibodies less effectively than peptides with the other mutations. The above results mean that DECODE identify accurate epitopes at the single amino acid level and quantify variations in recognition by antibodies. Finally, to estimate genome-wide cross-reactivity for the 3 antibodies, we calculated DECODE scores for all mouse proteins derived from UniProt (Fig 3E). The results showed that clones 2H2 and 9F6 had high specificity for the c-fos protein, whereas C-10 showed cross-reactivity with other proteins. Among the proteins predicted to cross-react with C-10, we particularly focused on 4 proteins (AHNAK2, LAMA1, FOSL2, and PLCD4) that had higher DECODE scores than the target protein (c-fos) (Fig 3E). There were observed that the sites with the maximum DECODE scores on these proteins contain sequences very similar to the motif of C-10. In particular, FOSL2 belongs to the same FOS family as the c-fos protein, and its sequence is very similar, with a close DECODE score value. The ELISA results for these sequences showed that anti-c-fos (C-10) cross-reacts with AHNAK2, LAMA1, and FOSL2 (S3I Fig). The antigen peptide derived from PLCD4 (pep4) was not soluble and could not be tested in ELISA. These results indicate that in immunohistochemistry, C-10 may stain not only the target protein c-fos, a neuronal activity marker, but also other proteins, leading to a risk of misinterpretation of research results. This highlights the importance of predicting cross-reactivity in immunohistochemical research.

Fig 3. DECODE identifies the binding sites and hot spot residues.

(A) Sequence logos at the highest DECODE score position at round 3 of the DECODE selection for anit-c-fos monoclonal antibodies (9F6, 2H2, and C-10). (B) Binding assay for anti-c-fos antibodies (clones 2H2, 9F6, and C-10) at each concentration against mouse c-fos protein by direct ELISAs. There was immobilized wild-type protein (gray) or double mutated proteins (light blue, yellow, and pink), respectively. The A450 was normalized by the saturated value of antibodies with different epitopes provided as S3F–S3H Fig. Data are shown as means ± STD (n = 3). Lines represent fitting with Michaelis–Menten equation. (C) Competitive assay for anti-c-fos antibodies (clones 9F6, 2H2, and C-10) to mouse c-fos protein with every single amino acid mutated peptide at various concentrations. A450 values were normalized by non-competitive conditions shown in S4A Fig. Data are shown as means ± STD (n = 3). Lines represent fitting with Michaelis–Menten equation. Mutant peptide sequences are provided as S4 Table. (D) Scatter plot of the correlation between the difference in converged amino acid ratios between wild and mutant amino acids (Δconvergence ratio (WT-MT)) and the difference of IC50 calculated in (C) between mutant and wild type (Log10 ΔIC50 (WT-MT)). Plots, lines, and shaded areas represent actual data, the regression line, and the 95% confidence bounds, respectively. Spearman correlation coefficients about 3 clones 9F6, 2H2, and C-10 were respectively R = 0.61, R = 0.63, and R = 0.82. (E) Manhattan plots of the DECODE scores on the human protein database for each anti-c-fos antibody (clones 9F6, 2H2, and C-10). Plots for the c-fos protein are shown in red. These plots visualize data downsampled to 1/50 using LTTB. The data underlying for panels A–E shown in the figure can be found in S1 Data or https://doi.org/10.5281/zenodo.14286317. DECODE, Decoding Epitope Composition by Optimized-mRNA-display, Data analysis, and Expression sequencing; ELISA, enzyme-linked immunosorbent assay; LTTB, Largest Triangle Three Buckets.

https://doi.org/10.1371/journal.pbio.3002707.g003

Therefore, it can be said that the DECODE results may serve as a standard for evaluating the quality of antibodies, including their target specificity. The above series of experiments with the anti-NeuN antibody clone A60, which targets a different protein, NeuN, demonstrated consistency of DECODE (S4C–S4G Fig).

DECODE was performed on a total of 230 antibodies, and several remarkable findings were obtained (S1 Table). The peptide selection for these antibodies was conducted in 3 rounds. The cDNA-peptide complexes obtained in each round were quantified by qPCR, and it was found that sufficient peptides had converged by the third round. Focusing on the epitopes of the anti-p53 monoclonal antibodies, we found similar epitopes despite their different clones, between clones Bp53-12, DO-1, and DO-7, and also between C-11 and D-11 (Figs 4A, 4C, S5A and S5B). First, results of ELISA indicated that 8 kinds of anti-p53 antibodies specifically bound to each individual epitope peptide and p53 protein (Figs 4D and S6A). Competitive ELISA using single amino acid mutant epitope peptides for clones Bp53-12, DO-1, and DO-7, which showed strong affinity, ensured the accuracy of the identified epitope hotspot residues (Figs 4B and S5C–S5E). These 3 clones also showed similar Km values to p53 protein (Figs 4E and S6B). As described above, both epitope analysis results and ELISA results showed that Bp53-12, DO-1, and DO-7 recognize the same binding site at the single amino acid level. Therefore, we next investigated whether the antibodies in these 3 clones with different name were actually the same or different. We performed quantitative LC-MS analysis to compare the sequences between the proteins for Bp53-12 and DO-1, which are both mouse IgG2a kappa proteins, to determine whether they were the same antibody or not. The antibodies and the internal control proteins, lysozyme, and cytochrome C, were digested with trypsin, and the resulting peptides were labeled with light (28 Da) and heavy (32 Da) isotopes containing dimethyl groups to identify the source proteins. As a result, the signal intensities of the peptides from the same antibody and internal control were almost the same for both the light and the heavy fragment; however, the variable region-derived peptides of clone Bp53-12 and DO-1 showed distinctive signals, respectively (Fig 4F). DO-7 (mouse IgG2b kappa) and Pab240 (mouse IgG1 kappa) were also compared with Bp53-12 by LC-MS and were confirmed to contain different peptides (S6C and S6D Fig). These results indicate that antibodies with different variable regions and isotypes may recognize nearly identical amino acids, a finding that can be readily revealed by DECODE.

Fig 4. DECODE reveals epitope similarities among monoclonal antibodies and differences among distinct clones for polyclonal antibodies.

(A) Epitope sequence logos of anti-p53 monoclonal antibodies (Bp53-12, DO-1, and DO-7) at third round of DECODE selection. (B) Correlation between the difference in convergent amino acid ratios between wild and mutant amino acids (Δconvergence ratio (WT-MT)) and the difference of IC50 calculated in S4C Fig between mutant and wild type (Log10 ΔIC50 (WT-MT)). Lines and shaded areas represent the regression line and the 95% confidence limits, respectively. Spearman correlation coefficients for 3 clones (Bp53-12, DO-1, and DO-7) were respectively R = 0.90, R = 0.94, and R = 0.90. (C) Synthesized peptides for binding assays and epitope logos for each anti-p53 monoclonal antibody. (D) Results of the direct ELISAs for each anti-p53 antibody against each peptide shown in (C), p53 protein, and BSA. Data are shown as means ± STD (n = 3). (E) Binding curves of direct ELISAs and Km values for anti-p53 antibodies (clones Bp53-12, DO-1, DO-7). Data are shown as means ± STD (n = 3). Lines represent fitting with Michaelis–Menten equation. (F) Correlation of the signal intensities of LC-MS during anti-p53 antibodies. Antibodies were digested by trypsin and their amino groups were labeled by dimethyl group with isotope reagents to identify each antibody. Blue circles and orange circles indicate the constant and variable regions in each antibody. The gray triangles and the cross show the peptides of lysozyme and cytochrome C proteins, respectively, which were added to each sample as internal standards. (G) The catalog number, lot number, synthesized peptides for binding assays, and epitope logos for anti-c-Fos polyclonal antibodies. (H) Binding assay with direct ELISAs for each antibody against each peptide shown in (G). Data are shown as means (n = 3). The data underlying for panels A–H shown in the figure can be found in S1 Data or https://doi.org/10.5281/zenodo.14286317. DECODE, Decoding Epitope Composition by Optimized-mRNA-display, Data analysis, and Expression sequencing; ELISA, enzyme-linked immunosorbent assay.

https://doi.org/10.1371/journal.pbio.3002707.g004

Polyclonal antibodies are still used in biological research, although it is widely perceived that there is poor reproducibility among lots. We analyzed 2 different products of anti-c-Fos polyclonal antibodies’ products, which respectively had 2 lots made from different individuals: cat#ab190289 (lot# GR3253255-1, GR3313102-1, Abcam) and cat#ab209794 (lot# GR3198011-8, GR3266315-7, Abcam). We confirmed the binding between these antibodies and 15 types of c-Fos protein fragment peptides identified by DECODE (Figs 4G and S7A–S7C). As an ELISA result, we found that despite having the same product number, both lots bound to almost different positions on the target and the correlation between lots was low (Figs 4H and S7D).

We attempted to apply the epitope information from DECODE to 3D immunostaining. First, we demonstrated the prediction of the sensitivity of formalin/PFA fixation in immunostaining based on the epitope information for 2 anti-c-Fos monoclonal antibodies, 2H2 and 9F6. According to DECODE results, clone 9F6 recognized the hotspot residues contained Tyr and His, which form irreversible methylene bridges via primary amines and formalin/PFA [28], that is why it was predicted to decrease the signals for strong fixed samples (S8A Fig). We certified the correctness of these predictions by ELISA (S8B Fig), and 3D immunostaining using normally fixed and strongly fixed mouse whole brains (Figs S8C–S8F and S9). As a result, 9F6 decreased the signals than 2H2 due to the influence of fixation, and antigen retrieval was also ineffective. Thus, it is possible to select an appropriate antibody for each formalin fixation condition based on the detailed epitope information obtained by DECODE. This is expected to enhance the reproducibility of experiments.

Furthermore, by applying epitope information, we can address the issue of antibody permeability, which is a significant challenge in 3D immunostaining. Generally, when performing 3D immunostaining on a cleared whole mouse brain or large organ without slicing, the antibodies tend to be trapped on the tissue surface and do not penetrate into the center, making uniform immunostaining difficult. To investigate the important factors for the penetration of the antibody into large organs, referring to previous studies [29], we simulated the concentration of the antigen-antibody complexes using various parameters: antibody concentration, antigen concentration, diffusion rate, and staining duration (Figs 5A and S10A–S10D). The results suggested that the low k_on improved the penetration speed, but the complex concentration was decreased (i.e., the signal). However, in many cases, low antibody permeability is due to the antigen concentration being too high, which means that even if permeability is increased, sufficient signal intensity can be expected for observation. Therefore, we hypothesized that a strategy to improve antibody permeability by reducing the apparent k_on of antibodies using chemically synthesized epitope peptides might be effective for 3D immunostaining. We first obtained the epitope information of anti-NeuN and anti-tyrosine hydroxylase (TH) monoclonal antibodies, which are difficult to stain in the whole mouse brain, and the accuracy of these epitopes was demonstrated by competitive ELISA (Figs 5B–5D and 5G–5I). We next immunostained NeuN or TH proteins in whole mouse brains with each antibody, both in the presence and absence of competitive epitope peptides that reduces the apparent k_on. To compare the antibody penetration between with and without epitope peptides, we aligned brain samples to a representative sample as a fixed image using ANTs software and autofluorescence images (S10E and S10F Fig) [30,31]. The results indicated that the permeabilities of these antibodies increased in the cerebellum and striatum, regions typically difficult for antibody penetration, when a competing peptide was added. Interestingly, when using the peptides with double mutations in hotspot residues, the penetration speed was not increased (Figs 5E, 5F, 5J–5M and S11). Furthermore, we performed co-staining for the NeuN and TH proteins using the optimal concentration of the full-match peptide. The penetration speed of each antibody was compared in the cross-sections and was confirmed to be improved (Fig 5N and 5O). These results support the hypothesis that epitope peptides identified by DECODE enhance the permeability of 3D immunostaining by individually controlling the apparent k_on of each antibody. This highly antibody-specific strategy demonstrated that co-staining can also be readily achieved.

Fig 5. DECODE improves the penetration speed of antibodies and contributes to uniform staining in 3D immunostaining.

(A) Simulation of the penetration of the antibody for evaluating the 3D staining pattern. The upper panels show staining patterns at the different values of the association rate constant (k_on). The lower graph shows the concentration of antibody-antigen complex (red) and antigen (gray) in the cross-sectional region at Y = 0. (B, G) Epitope sequence logos of the anit-NeuN monoclonal antibody (clone A60) (B) and anti-tyrosine hydroxylase (TH) antibody (clone EP1532Y) (G). (C, H) Synthesized epitope peptides of A60(c) and EP1532Y(H) based on (B, G). Upper and lower respectively show a wild-type sequence and double mutant. (D, I) Competitive ELISAs for A60 (D) and EP1532Y (I) with full-match or mismatch peptides against target protein. Data are shown as means ± STD (n = 3). Lines represent fitting with Michaelis–Menten equation. (E, L) Immunostained images of the sagittal plane of the registered mouse brain. Black, light color, and deep color indicate no peptide, full-match peptide, and mismatch peptide, respectively. Immunostaining was performed by A60 (100 nM) (E) or EP1532Y (67 nM) (L) with various concentrations of epitope peptide. The color lines on the cerebellum in the images of (E) show cross-sectional locations. The color lines from A to B on the striatum of (L) show cross-sectional location. Scale bars, 2 mm. (F, M) Signal intensities of the cross-section by the cross-sectional location in (E, L). Data are shown as means ± STD (n = 3). (J) 3D regional cropping of the striatum. Scale bars, 2 mm. (K) Profiles showing the mean intensity of each depth from the surface of the striatum indicated as the red region in (J). Means ± STD (n = 3). (N) Co-staining of NeuN and TH using full-match epitope peptides at optimal concentration. Scale bars, 2 mm. (O) The images show the co-stained sample with or without epitope peptides. Scale bars, 2 mm. Graphs show the mean intensity of the cross-section by the line in the upper images. The data underlying for panels A, B, D, F, G, I, K, M, and O shown in the figure can be found in S1 Data or https://doi.org/10.5281/zenodo.14286317. ELISA, enzyme-linked immunosorbent assay.

https://doi.org/10.1371/journal.pbio.3002707.g005

To demonstrated the applicability of DECODE to more complex blood antibodies, we performed epitope analysis of serum antibodies of EAE mouse. Assuming that the cause of the disease is unknown, here we attempted to re-identify a sequence that matched a peptide to induce the disease without relying on prior antigen information, methods such as constructing a library based on the antigen protein sequence, purifying specific antibodies using the antigen protein, or extracting peptides are not utilized. Initially, we confirmed the sensitivity of DECODE against a highly complex antibody mixture by titrating the anti-FLAG antibody, which mimics the same antigen’s antibody group in healthy mouse serum. As a result, the FLAG motif was detected at antibody concentrations exceeding 1.1% and 0.28% in the second and third rounds, respectively (Figs 6B and S12A–S12C). Subsequently, we prepared EAE mice by injecting the myelin oligodendrocyte glycoprotein (MOG) fragment peptide (MOG35-55) with complete Freund’s adjuvant (CFA). We immobilized IgG antibodies from EAE mouse serum with a score of 3 or non-immunized healthy mice on protein G magnetic beads (Fig 6A) and performed DECODE for these antibodies from the EAE mice (n = 8) and healthy mice (n = 6). After the second and third rounds of DECODE, some motifs on the immunized peptide were detected only in EAE mice (Fig 6C and 6D). Moreover, these epitopes showed high DECODE scores among all mouse proteins (Figs 6E and S13). In addition, the maximum DECODE scores and ranks of the MOG protein were significantly higher in EAE mice than in healthy mice, as determined by the Mann–Whitney U test (Figs 6F and S12D–S12F). The presence of anti-MOG35-55 antibodies in the serum of EAE and healthy mice was confirmed by ELISAs (Figs 6G and S12G). These results indicate the feasibility of re-identifying pathogen-specific epitopes from blood antibodies using the DECODE method, without relying on the antigen itself.

Fig 6. DECODE identifies the pathogenic epitope for serum antibodies in EAE mice.

(A) Workflow of unbiased epitope analysis of serum antibodies using DECODE. (i) EAE mice preparation by injecting MOG35-55 peptide with CFA. (ii) Immobilization of the serum antibodies on protein G magnetic beads. (iii) DECODE selection. (iv) Comparison of the DECODE results between patient and healthy mouse. (B) Evaluation of the sensitivity of the DECODE was performed using sera added with various concentrations of anti-Flag antibody. The plot shows percentages of Flag motifs in each selection round. Data are shown as means ± STD (n = 3). (C) DECODE scores on MOG protein of each EAE mouse serum and healthy mouse serum. (D) Epitope sequence logos at third DECODE selection round. (E) Manhattan plots of the DECODE scores on mouse protein database. Each panel indicate individual serum antibody (EAE mouse (n = 8), healthy mouse (WT, n = 6), anti-FLAG M2 magnetic beads (M2 beads), and no antibody immobilized beads (No antibody)) in the second DECODE selection round, respectively. MOG protein is shown in red. (F) Box-whisker plot showing the highest DECODE score on MOG protein for each DECODE selection round comparing EAE (green, n = 8) and WT (gray, n = 6). P values were obtained by Mann–Whitney’s U test. (G) Binding assays by direct ELISAs for serum antibodies against MOG peptide. Data are shown as means ± STD (n = 3). Lines represent fitting with Michaelis–Menten equation. Table shows calculated Km values. The data underlying for panels B–G shown in the figure can be found in S1 Data or https://doi.org/10.5281/zenodo.14286317. CFA, complete Freund’s adjuvant; DECODE, Decoding Epitope Composition by Optimized-mRNA-display, Data analysis, and Expression sequencing; EAE, experimental autoimmune encephalomyelitis; ELISA, enzyme-linked immunosorbent assay; MOG, myelin oligodendrocyte glycoprotein.

https://doi.org/10.1371/journal.pbio.3002707.g006

Single-strand template DNA library CCTAATACGACTCACTATAGGGTTAACTTTAAGAAGGAGATATACATATG(NNK)nTGCGGCAGCGGCAGCGGCAGCTACTTTGATCCGCCGACC (n = 12) was obtained from FASMAC Inc. Further, 25 pmol of the ss-DNA library was formed duplex and amplified using Phusion High-Fidelity DNA Polymerase (NEB), according to the manufacturer’s recommended protocol with 0.5 μm Primer 1 (CCTAATACGACTCACTATAGGGTTAACTTTAAGAAGGAGATACATATATG) and 0.5 μm Primer 2 (ggTCGGCGGATCAAAGTAG, g = 2′-O-methylguanosine). The PCR conditions were as follows: 500 μl scale and 4 cycles of temperature changes (95°C for 10 s, 58°C for 10 s, and 75°C for 30 s in a thermal cycler). The PCR products were electrophoresed (135 V, 15 min) on a 2% agarose gel containing 0.1% ethidium bromide to validate the amplification of the DNA library and were purified with Ampure XP. The control DNA template for the anti-Flag antibody (mc1’) was prepared based on a previous study [44] (S2 Table).

The purified template DNA library was transcribed with the following transcription mixture: 40 mM HEPES-KOH (pH 7.6), 20 mM MgCl₂, 2 mM spermidine, 5 mM DTT, 12.5 mM NTPs, and 0.43 μg/μl of the T7 RNA polymerase. The condition of the transcription reaction that can finally inactivate the T7 RNA polymerase was as follows: 37°C for 1 h, 75°C for 5 min. All DNA purified above was used in the first round, and 5 μl (approximately 10 pmol) of DNA was utilized in the second and subsequent rounds. The transcription products were annealed with the following mixture: 5 μm Pu-DNA (5′-[PHO] CTCCCGCCCCCCCCGTCC [SpC18]5CC [Puromycin]) and 5 μm splint DNA (5′-GGGCGGGAGGGTCGGCGGATCAA) (S2 Table). The condition of the annealing reaction was as follows: 95°C for 1 min and 75°C for 30 s. Then, the temperature was lowered to 25°C with a constant gradient of 1°C/15 s. The annealed product was added with 35 U of T4 DNA Ligase (Takara Bio Inc.) and incubated at 37°C for 1 h and then at 16°C for O/N. Transcription and ligation products were electrophoresed on 10% acrylamide gel containing 7 M urea (180 V, 40 min) to confirm transcription reaction and ligation efficiency. The Pu-DNA-linked mRNA library of the first round was purified with RNAClean XP. After the second round, this purification step was skipped. The first round of transcription was performed at a 500 μl scale. After the second round, it was performed at a 10 μl scale. The first round of ligation was performed at a 500 μl scale. After the second round, it was performed at a 10 μl scale. Transcription and ligation products derived from the DNA template with and without Gm modification were electrophoresed (180 V, 3 h) on a 30 cm length of 10% acrylamide gel containing 7 M urea to validate transcription reaction and ligation efficiency in a single-base resolution.

In total, 2 μl of the recovered cDNA library was labeled with barcode sequence via PCR using barcode primers (S3 Table). PCR was performed under a similar condition during the amplification step described above. The barcode-labeled cDNAs were mixed and purified with Ampure XP. Sequences were obtained via NGS (Genome Analysis Office, Genetic Information Research Center, Institute for Microbial Diseases, Osaka University) using Illumina MiSeq, HiSeq 3000, or NovaSeq 6000 with approximately 1 to 10 million reads per antibody.

Peptides were synthesized with an automated peptide synthesizer (Syro Wave; Biotage) using Fmoc solid-phase chemistry. Synthesized peptides were treated with a cleavage cocktail (trifluoroacetic acid [88.5% v/v], phenol [4.4% v/v], 1,2-ethanedithiol [2.2% v/v], thioanisole [0.4% v/v], and water [4.4% v/v]) to facilitate cleavage and de-protection at r.t. for 3 h. The molecular weight of the synthesized peptides was confirmed using MALDI-TOF Mass (MALDI-TOFMS AXMA ASSURANCE [Shimadzu]) with α-cyano-4-hydroxycinnamic acid matrix (Sigma-Aldrich). Synthesized peptide sequences for ELISA are shown in S4 Table.

In-solution protein digestion was performed based on the phase-transfer surfactant protocol [45]. Antibody (Ep53-12 or DO-1, 100 ng each), cytochrome c from the equine heart (40 ng), and lysozyme from egg white (40 ng) were solubilized in 20 μl of phase-transfer surfactant lysis buffer (50-mM NH4HCO3, 12 mM sodium deoxycholate, and 12 mM sodium N-lauroylsarcosinate) and were incubated at 95°C for 5 min. Cysteine reduction and alkylation were performed via incubation with 20 mM tris ((2-carboxyethyl) phosphine) and 20 mM iodoacetamide at 37°C for 1 h in the dark. Then, quenching was conducted with 20 mM dithiothreitol at r.t. for 20 min. The sample was digested with 40 ng of trypsin (proteomics grade, PROMEGA) at 37°C overnight. Via the addition of 1% TFA and centrifugation at 10,000 g for 10 min, the reaction was stopped, and the detergents were removed. In addition, in-gel protein digestion was performed using the gel bands of heavy and light chains of 800-ng antibodies (Bp53-12, C-11, D11, DO-1, DO-2, DO-7, Pab1801, and Pab240) separated via sodium dodecyl-sulfate polyacrylamide gel electrophoresis. The tryptic digest was loaded on a C18 pipette tip prepared, as reported in a previous study [46]. Then, dimethyl labeling was performed according to a previous report [47]. Formaldehyde (CH₂O, Sigma-Aldrich) or isotope-labeled formaldehyde (13CD2O, ISOTEC) was used to light or heavy label the sample with each antibody. The dimethyl-labeled peptides on the tip were eluted with 80% acetonitrile and 0.1% TFA and dried using SpeedVac. Then, they were stored at −80°C until mass spectrometry. The dimethyl-labeled peptides were reconstituted by 2% acetonitrile and 0.1% TFA. Then, equal amounts of light- and heavy-labeled samples were mixed just before MS measurement. The mixture was analyzed with the Orbitrap LTQ Velos mass spectrometer (Thermo Scientific) and the UltiMate 3000 RSLCnano-flow HPLC system (Thermo Scientific) under CID MS/MS mode. Raw data were searched with MaxQuant (version 2.0.1.0) against the UniProt mouse database (17,082 genes downloaded from Swiss-Prot in 2021.7) combined with the mouse immunoglobulin genes (1,213 genes from TrEMBL and 861 genes from VBASE2) [48]. Peptide identification was filtered at FDR 1%, and carbamidomethylation of Cys was set under a fixed modification. Oxidation of Met and determination of Asn/Gln was set for variable modifications. The MS intensities of the peptides were obtained via label-free quantification implemented in MaxQuant [49].

For whole-brain staining, the brains of adult mice (C57BL/6N) were used. MK-801 was injected into the mice. Dissection, fixation, PBS washing, tissue clearing, staining, imaging, and analysis were performed according to previous studies [29,30]. For immunostaining in Fig 6, the respective concentrations of peptide were used instead of the quadrol. For each experiment, tissue clearing imaging was confirmed to be reproducible on individual mouse whole brain samples (n = 3).

Via the subcutaneous injection of MOG peptide (UT) antigen at 2 back sites with CFA, 10- to 12-week-old B6N mice (n = 8) were generated. Pertussis toxin was administered on days 0, 2, and 7. After achieving a score of ≥3, the whole-blood sample was collected to obtain the serum [50].