Understanding where and how an antibody binds to its target protein is important for understanding how the antibody performs its function, whether that function is neutralizing a pathogen during an immune response, binding an epitope in immunoassays, or labeling a target molecule in a live-cell imaging experiment. However, determining the binding epitope of an antibody can be a time and labor-intensive endeavor with significant expense. Traditionally, antibody epitopes on target proteins have been identified by performing deletion analysis on the target protein to determine if the antibody loses reactivity for the deletion mutants in various immunoassays, which provides the general region of the target protein the antibody binds to. With the advent of widely available chemical peptide synthesis, sequence-specific synthetic peptides can be used for competitive immunoassays (such as enzyme-linked immunosorbent assays (ELISA)) to establish sequences that can effectively compete with the antigen for antibody binding. Peptide mapping experiments are a powerful method for determining the fine sequence of linear antibody epitopes, but these experiments can be relatively expensive and the time between experimental design and data acquisition can be weeks to months due to the need to design and chemically synthesize peptides. Once a peptide has been identified that binds with high affinity and specificity to an antibody antigen binding fragment (Fab), crystal structures can be determined that demonstrate intermolecular interactions between the peptide and antibody. These can then provide a molecular-level explanation for an antibody’s binding mode. Finally, with the advent of rapid single B-cell sequencing technologies to analyze humoral immune responses towards vaccination or infection, determining where specific antibody clones bind on an antigen becomes even more challenging due to the need to isolate or synthesize specific antibody genes, produce antibodies, and then perform deletion or epitope mapping experiments described above to fully understand how and where antibodies bind. These challenges make determining antibody epitopes expensive and time-consuming, and limit the number of antibodies that are characterized in detail.
Antibodies that bind to linear epitopes represent an important subset to molecular biology, as they can be added to recombinant proteins for use in various types of immunoassays. A number of linear antibodies have been developed for use in various immunoassays (ELISA, western blot, immunofluorescence, etc.). The development of computational methods for linear epitope determination could increase the number and quality of new linear epitopes available to the field. Most epitope prediction tools (such as BepiPred (1), ElliPro (2), and ABCpred (3)) are generally designed to predict regions of an antigen that could be recognized by any antibody rather than a specific antibody. These programs also provide no insight into the structural match of the epitope and antibody, potentially making decisions without key structural information that otherwise may be relevant. The challenge in predicting epitopes for a specific antibody lies in the complexity of protein-protein interaction dynamics, which includes conformational changes, binding affinities, and thermodynamic stability. Structure based approaches including HADDOCK (4, 5) and ZDOCK (4, 6) can be used to dock peptides into antibody structures, but these require known peptides for binding. Significant progress has been made to address this problem via deep learning: some of the new and exciting tools are GearBind (7), PALM and A2binder (8), and DSMBind (9). We point the reader to this review for an excellent overview of some of the tools that have existed for some time, along with a comparison of these tools (10).
Determining antibody-epitope interactions is, at its most basic level, a structural biology problem. Determining what molecular interactions are present between an antibody and its antigen can define the epitope, determine what portions of the epitope and CDR sequences are responsible for molecular interactions, and provide clues to antibody specificity and affinity. With the advent of highly accurate structural predictions, including the AlphaFold2 (AF2) neural networks (11, 12), the ability to accurately predict protein structures, and potential protein-protein interactions, has dramatically increased. AlphaFold2 was trained on existing protein structures and can effectively model new protein structures. Numerous antibodies, antibody Fab regions, and other related constructs with bound target peptides or proteins have been crystalized and deposited into the Protein Data Bank (PDB) (for example (13–16)). These PDB entries represent a valuable training set that may increase the likelihood that AlphaFold2 can successfully predict the structure for antibody-epitope complexes (12, 17–19). The authors of AlphaFold2 multimer (12) comment on the difficulty of predicting antibody-epitope complexes, and results for this are indeed mixed at best (17–19). One way in which this current report is distinct is our focus on linear epitopes. We hypothesize that the lack of strong competing structure within the short peptide may boost AF2 prediction of scFv-epitope binding predictions relative to conformational epitopes. This problem has precedent, as AlphaFold2 has previously been used to study the interactions between proteins and peptides (17, 18). AlphaFold2’s ability to correctly dock independent protein chains can be repurposed to predict how strongly two proteins interact together and extends to predicting the interaction between an antibody and short flexible peptides (linear epitopes) drawn from a larger protein antigen.
To maximize compute efficiency, it is helpful to minimize the size of the system subject to structure prediction. The computational expense of AlphaFold2 scales with the square of the length of the concatenated sequences involved. Fortunately, not all portions of the antibody are critical. Antigen binding by antibodies is primarily dictated by the antigen binding fragment (Fab) containing the variable light (VL) and variable heavy (VH) fragments. Conversion of full antibody sequences into single chain variable fragments (scFv) can significantly reduce structure prediction complexity and compute time. A wildtype scFv sequence can easily be generated directly from translated antibody heavy and light chain DNA sequences. Briefly, the sequences are first divided into framework and complementarity determining regions (CDRs) using Kabat (20) or IMGT (21) nomenclature.
A flexible linker sequence (GGGGSGGGGSGGGGS, 15 a.a.) is then added between the new C-terminus of the truncated light chain and the original N-terminus of the shortened heavy chain to generate a single protein sequence that incorporates both antigen-binding chains. The resulting fusion protein often functions in a similar fashion to the original antibody. Another well-known protein engineering strategy for antibodies is “loop grafting”, where the CDR loops from one antibody are grafted onto a different framework region. We have recently used this approach to develop scFvs with improved in vivo performance (22). The structures of the novel scFv chimeras can be rapidly and confidently be predicted by AlphaFold2 due to their small size and the extensive immunoglobin representation within sequence databases and the PDB. Excluding the time needed to obtain a multiple sequence alignment (MSA), predicting the structure for a single scFv in complex with a 10-a.a. peptide requires only 1.5 minutes on an NVIDIA A5000 graphics processing unit (GPU). This modest compute time allows a GPU-laden server or workstation to handle large-scale structure prediction of hundreds of related systems. As for the MSA input, a high quality MSA can quickly be obtained via ColabFold (23), which relies on the MMseqs2 MSA server. In our workflow, we repeatedly predict the structure for a fixed single scFv sequence in complex with varying peptide partners. In this case, we do not expect the peptide portion of the MSA to be useful. Therefore, to avoid sending hundreds of nearly identical MSA requests to MMseqs2 MSA server, and to avoid varying information in the MSA, we slightly modified the LocalColabFold code to include the option to cache the MSA (install available on the GitHub). We generate one cached MSA per epitope scan, where each residue in the query peptide is a glycine.
Several recent papers have attempted to use AlphaFold2 to identify antibody epitopes (24–26), but have primarily focused on computational identification and have not verified their results using new antibodies that are not within the PDB training set. While there are many other structure prediction models other than AlphaFold2 (27, 28), including some specifically dedicated to predicting antibodies or antibody-like structures (29–32), we chose AlphaFold2 to directly test its ability to correctly identify and place epitopes into an antibody binding cleft. We selected AlphaFold2 due to its widespread use throughout the literature, as well as its ease of installation and modification via the LocalColabFold implementation (23). In this project we test a method we call PAbFold, a LocalColabFold-based pipeline to identify epitopes for several well-known linear-epitope antibodies from sequence information only. There was a strong correlation between AlphaFold2’s confidence in the peptide structure (pLDDT) (33) and the experimentally verified epitope binding sequence. Additionally, we found that AlphaFold2 very accurately predicted the linear epitope of a novel SARS-CoV-2 nucleocapsid-specific antibody (mBG17) with minimal prior epitope information. The molecular interactions predicted by AlphaFold2 were experimentally validated using peptide mapping ELISA experiments. Overall, this work demonstrates that AlphaFold2 has compelling promise for linear antibody epitope discovery from sequence information alone. We also have observed that this emergent linear epitope prediction ability is sensitive to the peptide length and that the performance was optimal when using AlphaFold2-multimer version 2 and older MSAs generated by MMSEQS version 2202 server, rather than the more recent AlphaFold2-multimer version 3 models and MMSEQS version 2302 server.
Software
All structure predictions were completed on a single AMD EPYC 7443 server with two NVIDIA RTX A5000 GPU cards. PAbFold code was written in Python 3.7 and Bash. The only extra Python dependencies are NumPy and Matplotlib. AlphaFold2 calculations were run using an installation of LocalColabFold (23). Briefly, PAbFold contains 3 stages. In the first stage, a python script ‘A_PeptideMapping_prep_submission_files.py’ writes FASTA input files for ColabFold. Each FASTA file contains the entire sequence of the subject scFv, a colon “:”, and then the candidate linear epitope which represents a small section of the target antigen protein that changes dependent upon both the epitope length (default 10 a.a.) and a sliding window (default 1 a.a.).
After completion of the ColabFold jobs, two different analysis methods are presented in this paper, and both are accessible via the ‘B_PeptideMapping_plddt_perres_analysis.py’ python script. The first is the ‘Simple Max’ method, which assesses each peptide window with only the output model that is top ranked by ColabFold (on the basis of ipTM). The AlphaFold2 confidence pLDDT (33) is recorded for each residue within the peptide. Other than the N-and C-terminal residues, each residue is observed within multiple windows. We proceed to calculate (and plot) the maximum pLDDT observed for each residue across the set of sliding window peptides that contain that residue. Thus, in the ‘Simple Max’ method each residue is considered independently. To obtain aggregate scores for each peptide window, we sum the maximum pLDDT associated with each member residue. This method is sensitive in that any isolated high-confidence residue placements in the top ranked AlphaFold2 peptide prediction can increase the score, but a high aggregate peptide score could arise from multiple, mutually inconsistent peptide binding poses. Our second, complementary analysis method instead focuses on recognizing full peptide poses of elevated AlphaFold2 confidence. We refer to the second method as the ‘Consensus method’ because it begins by averaging the per-residue pLDDT across the five AlphaFold2 models. We then compute the average pLDDT for each peptide. For visual inspection, scripts output a heat map for the average per-residue pLDDT and a bar-chart that for the subsequent per-peptide average pLDDT. In this case, we simply rank top peptides based on the per-peptide average pLDDT. Scripts are available at https://github.com/jbderoo/PAbFold.
Antibody sequences
Sequences and references for antibodies, scFvs, and antigens can be found in Supplemental Table 1A. To create an scFv, the complementarity determining regions or loops of an antibody are identified via the Kabat numbering scheme. The loops are then spliced onto the scFv backbones of the 15F11 and 2E2 as previously described by our group (22). The scFv sequences are aligned with their CDR loops and flexible linkers highlighted in Supplemental Table 1B.
Monoclonal Antibody Production
Anti-SARS-CoV-2 nucleocapsid protein (NP) monoclonal mouse antibody mBG17 was previously developed and characterized (34). Briefly, two BALB/c mice immunized with recombinant NP were sacrificed and primary splenocytes isolated. Splenocytes were fused with Sp2/0 Ag14 myeloma cells and individual hybridoma clones were isolated after eleven days. Hybridoma clones were tested for antibody production against NP via enzyme-linked immunosorbent assay (ELISA) and western blot. Clones were further tested for isotype and cross-reactivity, and VH and VL sequences were determined. The hybridoma clone mBG17 was identified as a SARS-CoV-2 nucleocapsid-specific antibody targeting linear epitope via ELISA and western blot (34). Generation of recombinant mBG17 and production of recombinant antibody in 293F cells was performed as described in (35). The approximate epitope region for mBG17 was determined via western blot with modified recombinant NP proteins containing 40 to 50 amino acid deletions. The epitope location was determined to reside between SARS-CoV-2 nucleocapsid residues a.a. 381-419 based on loss of western blot signal with the a.a. 381-419 deletion (34).
Peptide Competition ELISA
The anti-SARS-CoV-2 nucleocapsid protein mBG17 antibody epitope was experimentally identified using competition enzyme-linked immunosorbent assay (ELISA). Using the previously determined 39 nucleocapsid protein amino acid range for the mBG17 epitope as a starting point, seven overlapping peptides were synthesized (Thermo Scientific) spanning the 39 amino acid region with overlaps of 5 amino acids. These peptides were termed Fragment 1 through 7 (Table 1). A 96-well ELISA plate was coated with 0.1ug/ml of recombinant SARS-CoV-2 NP (34) overnight at 4°C. The plate was blocked with 4% (w/v) dry non-fat milk in 1X PBS with 0.1% (v/v) Tween-20 for 1 h shaking at room temperature. While blocking, inhibited recombinant mBG17 antibody samples were produced by incubating 40 μL of antibody with 40 μg (approximately 30 nMol) of a single peptide fragment for one hour at room temperature. Following this, peptide-incubated mBG17 was applied to the blocked nucleocapsid protein coated plate in triplicate and allowed to incubate for 1 h at room temperature while shaking. The plate was rinsed with 0.1% (v/v) Tween-20 in 1X PBS and washed three more times for 5 minutes shaking at room temperature. The plate was then incubated with HRP-conjugated goat anti-mouse polyclonal antibody solution diluted at 1:20,000 in 1X PBS for 1 h shaking at room temperature. After another rinse and three more washes the plate was developed with 1-Step™ Ultra TMB-ELISA Solution (ThermoFisher) before stopping the reaction with an equal volume of 2M H2SO4. Solution absorbance at 450 nm was measured using a PerkinElmer Victor X5 multilabel plate reader. Absorbances were averaged within fragment-inhibited sample groups and corrected with the average value of the negative control. These absorbances were then normalized against the absorbance from the group with the highest value before multiplying by 100 to obtain percentage of potential signal.
The effect of single alanine substitutions on fragment 5 (DDFSKQLQQS) peptide binding was determined by competition ELISA using a series of ten alanine-substituted peptides (Table 1) at a range of concentrations to determine relative competition activity. A modified version of the previously described inhibition ELISA was performed using the unmodified Fragment 5 peptide and the ten alanine-substituted peptides. During the mBG17 inhibition step, the mBG17 antibody solution was incubated with a 4-fold serial dilution of peptides beginning at 40 μg and continuing to ∼2.5 ng before being applied to the NP coated plates in triplicate. The remainder of the competition ELISA was carried out as described above.
Assessment of AlphaFold2 generated scFv structures
We first verified that AlphaFold2 could generate scFv structures that have similar structures to their parent monoclonal antibodies. We chose the 9E10 clone of the anti-Myc antibody as an initial test system, as the scFv sequence is available (36) and has a well-known linear epitope (EQKLSEEDL)(37). We predicted the wild-type Myc scFv structure and aligned this model to the corresponding Fab crystal structure (PDB entry 2orb) via the align command in PyMOL (Supplemental Figure 1A). The AlphaFold2 predicted scFv was very similar (RMSD value of 0.42Å) to the anti-Myc Fab structure, suggesting that the predicted scFv structure was a suitable starting point for epitope prediction. We also examined the structures of the Myc CDRs loop grafted onto the 15F11 (38) and 2E2 (22) frameworks, as we have previously observed that loop grafting onto these frameworks can enhance protein folding and solubility (22). The loop-grafted Myc-2E2 and Myc-15F11 and structures were also similar to the Myc Fab structure (PDB 2ORB) (37) with similar RMSD values of 0.45Å (Supplemental Figure 1B), indicating that they are also reasonable starting points for epitope prediction.
Development of Python-based scripts for automated scFv:peptide structure prediction
We developed a series of Python scripts that automate the process of epitope prediction and analysis with AF2. A_Peptide_Mapping_prep_submission_files.py accepts a linear scFv sequence and a linear full-length antigen sequence, and processes the antigen sequence into a series of short peptides with custom peptide length and sliding window sizes (default parameters are 10 amino acid peptides with a 1 amino acid sliding window). It then adds lines for each scFv:peptide pair to a FASTA file. Structures are then predicted via LocalColabFold for each scFv:peptide pair with AlphaFold2 in parallel on two NVIDIA RTX A5000 GPUs. The python script B_PeptideMapping_plddt_perres_analysis.py parses the AlphaFold2 output structures to extract per-residue pLDDT for the peptide residues in each scFv:peptide pair. Conf_plot_and_top10.py will plot the maximum pLDDT (across all host peptides) scores as a function of amino acid position within the antigen sequence and ranks predicted peptides based on ΣpLDDT scores for the ‘Simple max’ method. To use the ‘Consensus’ method, include the –all-models flag when running B_PeptideMapping_plddt_perres_analysis.py. We also supply a python script that replicates how we present the data called all_model_analysis.py for use. An overview of the method is shown in Figure 1. AF2’s failure to predict whole antigen structure coupled with the scFv is highlighted in Supplemental Figure 2.
Testing of scFv:peptide structure prediction method using the Myc Epitope
We first tested the PAbFold method with the anti-Myc-scFv described in (39), using the full-length human Myc proto-oncogene protein sequence as the antigen. We initially used an antigen peptide length of 10 and a 1 amino acid sliding window. Given these parameters, the 9 a.a. Myc epitope motif (EQKLSEEDL) appeared intact within one of the 10-mer peptides, with subsets of the 8, 9, 11, and 12 a.a. appearing in neighboring sliding peptide windows. PAbFold generated predicted structures, each of which took an average of ∼200 seconds to process. The entire process took approximately 12 hours on our GPU server. AlphaFold2 placed all peptides into or near the traditional antigen binding site between the CDR loops (Supplemental Figure 3). The average confidence (mean pLDDT across residues) for these peptides ranged from 20 to 90. When we inspect the consensus confidence for each residue in each sliding window (Figure 2A), the expected Myc peptide epitope (EQKLSEEDL) was one of several peptides with high average pLDDT. The second highest ranked peptide in this analysis (QKLSEEDLL) was a near perfect match for the expected epitope. We consider this window to be a successful prediction. Perhaps surprisingly, the peptide window with the exact match (EQKLSEEDL) did not score particularly well due to its average pLDDT of 51.0. In this instance, the expected epitope sequence did not stand out when plotting the maximum observed per-residue pLDDT for each residue (Figure 2A, bottom).
We proceeded to test predictions with two engineered scFv chimeras where loop grafting was used to place the Myc recognition CDRs onto two antibody framework regions with high in vivo performance, generating Myc-15F11 and Myc-2E2 scFv sequences. Epitope prediction performance was markedly improved with the chimeric scFvs (Figure 2B and 2C). Specifically, the QKLSEEDLL peptide window became the top ranked peptide on the basis of average consensus pLDDT. In the case of Myc-2E2 (Figure 2C), the average confidence for the correctly predicted epitope was particularly high compared to alternate peptide windows, and another close match to the expected epitope (EEQKLSEED) was ranked within the top 5 peptides (Figure 2D). Ranking epitopes using the Simple Max analysis was similar; the region containing the correct epitope was nearly top ranked for Myc-15F11 and was top ranked for Myc-2E2 (Figure 2E). Thus, AlphaFold2 was able to more clearly detect authentic Myc antibody epitope using CDRs loop grafted onto the 2E2 or 15F11 frameworks, relative to the native Myc scFv framework.
To investigate the superior epitope recognition performance of the chimeric Myc scFvs, we aligned the Cα coordinates for the predicted scFv structures (predicted with and without the target epitope) to the reference crystal structure and calculated the RMSD for all backbone positions (N, Cα, C, O) and the loops (Supplemental Figure 4). Notably, regardless of the Myc scFv variant, the CDR loop RMSD improved by more than 1Å when the epitope was present. Secondly, consistent with the improved epitope prediction performance for the chimeric scFvs (15F11 and 2E2), the epitope peptide QKLISEEDL was placed more accurately for those predicted structures than in the WT scFv (Supplemental Figure 4). We could not discern an obvious structural difference between the WT and chimeric scFvs that explains the structure prediction performance gap.
Assessment of peptide length, sliding window size, and position on AlphaFold2 scFv:peptide structure prediction
Our initial selection of the 10 a.a. window was intended to match or slightly exceed the size of known epitopes such as Myc and HA. We next assessed how different peptide sizes and sliding window lengths would affect epitope prediction accuracy and run time. We re-ran the Myc-2E2-scFv:peptide complex prediction calculations varying peptide size between 8, 9, 10, and 11 (with a fixed sliding window size of 2) or varying the sliding window size to 1 or 5 (with a fixed peptide size of 10). We observed that using a sliding window of 2 a.a. provided nearly the same level of accuracy and resolution as the 1 a.a. Ultimately, we determined that our original peptide size of 10 amino acids and sliding window of 1 a.a. provided highest resolution data possible (Supplemental Figure 5), and therefore maintained a peptide size of 10 and a sliding window length of 1 for our remaining experiments.
We then predicted the complex structure for Myc-2E2 with various negative control peptides: A10, (GS)5, (GGGGS)2, and G10 to determine how non-binding peptides are docked and scored (Supplemental Figure 5I and 5J). We again observed that AlphaFold2 placed all peptides into the traditional antigen binding between the CDR loops, but the reported peptide scores for the negative controls were particularly low (29–41). These results indicate that AlphaFold2 “knows” where antigens bind in antibody or scFv structures and attempts to model any peptide partner into this region, but the low pLDDT scores indicate confidence in the interactions are quite low.
We also tested if AlphaFold2 could detect the Myc epitope if it was inserted as an epitope tag within different positions of a heterologous protein. We created a synthetic antigen by adding the Myc epitope within the 99-a.a. unrelated HIV-1 Gag protease protein sequence at either the N-or C-terminus or in the middle of the protein sequence, and used PAbFold to detect the Myc peptide (Supplemental Figure 6). In each case, the average consensus pLDDT was highest for the inserted epitope, such that the authentic epitope would be top ranked and prioritized for testing. Thus, as expected for a sliding window analysis, the epitope position within the antigen was no barrier to detection.
Testing of the PAbFold method using the HA Epitope
Based on our success detecting the Myc epitope, we sought to determine if our method could detect a different well-known linear peptide, HA, derived from positions 114-126 within the Influenza A virus hemagglutinin protein (YDVPDYASLR). Using an anti-HA scFv sequence that had been previously generated (22, 39), we generated new HA-15F11 and HA-2E2 scFvs loop grafted sequences. We used the same procedure described above to predict structures for influenza A virus HA derived peptides on HA-scFv (Supplemental Figure 7A), HA-15F11-scFv (Supplemental Figure 7B) and HA-2E2-scFv (Supplemental Figure 7C). In the HA case, the expected epitope was ranked highly for all three scFv variants, but when assessing entire peptides by average consensus pLDDT was only ranked in the top 5 for the HA-15F11-scFv. These results, in combination with the Myc results described above, indicate that AlphaFold2 can accurately detect linear antibody epitopes in antigen sequences, and that grafting CDR loops onto alternative scFv backbones may increase the noise-to-signal ratio, making the identification of correct epitopes more accurate.
Like the Myc system, trends are observed with the HA system regarding loop placement. Although not as extreme, the loops for all HA scFvs undergo movement that make it more closely match the crystal structure (PDB entry 1frg). Again, the epitope placement of predicted structures of the chimeric scFvs more closely mimicked the deposited crystal structure than the WT scFv (Supplemental Figure 4B).
Determination and experimental validation of a novel linear antibody epitope
The Myc and HA monoclonal antibodies are well known and several crystal structures (Myc PDB: 2or9, peptide bound (2009) | HA PDB:1frg, peptide bound (1994)) have been solved (22, 37, 39, 40), raising the possibility that AlphaFold2 has incorporated these antibody or epitope structures into its training set. The AlphaFold2 training set was reported to exclude chains of less than 10, which would eliminate the myc and HA epitope peptides. Nonetheless, to guard against the possibility that the AlphaFold2 models have incorporated specific knowledge into the training set thereby directly probing if PAbFold epitope scanning can predict a linear antibody epitope without a priori knowledge of the antibody or antigen sequence, we tested if PAbFold can predict the epitope sequence of a recently developed antibody lacking structural information available in the Protein Data Bank. The mBG17 mouse monoclonal antibody was generated in response to the COVID-19 pandemic, the antibody VH and VL sequences were determined, and the epitope was localized to a. a. 381-419 via Western blot analysis of deletion mutants of the nucleocapsid protein (34). mBG17 was not included in AlphaFold2’s training or test set, making it an ideal test case for de novo epitope prediction.
The mBG17 monoclonal antibody was converted to wild-type scFv, 15F11-scFv, and 2E2-scFv using the same procedures used for Myc and HA scFv. As an additional control calculation (labeled “3-body”), we used AlphaFold2 to predict the structure for a 3-protein complex (the peptide, and the disconnected nontruncated mBG17 VH and VL variable domain sequences). All 4 Fab variants (WT scFv mBG17, 15F11-mBG17 scFv, 2E2-mBG17 scFv, and 3-body mBG17) were screened against all 10 a.a. peptides with a 1 a.a. sliding window, as with Myc and HA. In all 4 cases, AlphaFold2 predicted that the top ranked peptides were located in the a.a. 381-419 region of the SARS-CoV-2 nucleocapsid protein, and more specifically residues a.a. 400-415 (Figure 3A,3B,3C, and 3D). The top scoring peptide for all three scFv variants was the 402-411 window (DFSKQLQQSM) (Figure 3E and 3F). The strong AF2 preference for peptides from this C-terminal segment was particularly evident in the average consensus pLDDT analysis.
We next sought to experimentally verify the minimal linear epitope for mBG17 to determine how closely the AlphaFold2 prediction corresponded to our experimental data. Seven 10 a.a. peptides that overlapped by 5 a.a. each were synthesized and used in competition ELISAs with mBG17 monoclonal antibody and recombinant SARS-CoV-2 nucleocapsid protein (Figure 3G and 3H). The peptide corresponding to a.a. 401-410 showed almost complete competition of mBG17 binding to the SARS-CoV-2 nucleocapsid protein in the ELISA, whereas none of the other peptides were able to compete for mBG17 binding to nucleocapsid. Peptides a.a. 296-405 and a.a. 406-415 overlap a.a. 401-410 at the N-and C-terminus, respectively, but neither was able to compete, indicating that mBG17 binds a.a. 401-410 on both sides of a.a. 405 and a.a. 406. An alignment of all the peptides used in the overlapping peptide competition ELISA experiments showed that peptide sequence DDFSKQLQQS represents the experimentally determined epitope for mBG17, nearly identical to the epitope predicted by AlphaFold2 (Figure 3H: DDFSKQLQQS). These results demonstrate that the PAbFold pipeline was able to very accurately predict the region that an antibody binds to a novel linear epitope that is not present in AlphaFold2’s training set.
Fine-characterization of the mBG17 epitope and comparison to the predicted AlphaFold2 model
To further experimentally characterize the binding of the mBG17 to the a.a. 401-410 (DDFSKQLQQS) peptide and compare experimental data with the predicted AlphaFold2 model, we designed and synthesized ten additional peptides, each containing an alanine point mutation at one position in the a.a. 401-410 peptide. The peptides are labeled D1A, D2A, F3A, S4A, K5A, Q6A, L7A, Q8A, Q9A, and S10A. Competition ELISAs were performed using increasing concentrations of each peptide to better assess differential binding (Figure 4A). As expected, WT (a.a. 401-410) peptide showed strong competition, although Q9A showed slightly better competition. This could be attributed to alanine’s propensity to be in an alpha-helical coil (PropA, AHC = 0) vs glutamine’s propensity to escape it (PropQ, AHC = 0.39) (41), thus further stabilizing the Q9A alpha helix. D1A showed no change in competition, indicating that D1 was not involved in binding. Peptides with substitutions K5A, Q6A, and S10A showed minor reductions in competition, S4A showed a moderate reduction on competition, whereas resides D2A, F3A, L7A, and Q8A all showed strong reductions in competition. These data indicate that the key interactions between mBG17 and the a.a. 401-410 peptide are residues D2, F3, L7, and Q8, with S4 playing a moderate role and D1, K5, Q6, Q9 and S10 playing negligible roles in binding.
Finally, we compared the experimental data shown above with the best scoring mBG17:DDFSKQLQQ model generated by AlphaFold2 (Figure 4B and 4C). The AlphaFold2 model suggests that residue D2 forms a hydrogen bond with mBG17 a.a. Y34, residue F3 forms a hydrophobic interaction with mBG17 a.a. L185, residue S4 lacks a hydrogen bond partner, residue L7 forms a hydrophobic interaction at the base of the binding cleft with mBG17 a.a. A104, and residue Q8 hydrogen bonds with the backbone carbonyl of Y34 and the backbone amide of W35. Residues that experimentally showed no or minimal effects on competition (D1, K5, Q6, Q9) are all predicted to interact primarily with the solvent and lacked visible interactions between the peptide and scFv sequence. In summary, the AlphaFold2-driven PAbFold prediction was remarkably consistent with the experimental alanine scanning data, suggesting that the prediction of the mBG17 linear epitope location was accurate due to the correct prediction of the structural details for how that linear epitope binds to the antibody.
In this project we assessed the ability of an AlphaFold2-based linear epitope scan pipeline we call PAbFold (Peptide:Antibody Fold) to predict linear antibody epitopes using just antibody and antigen sequences. We first assessed the quality of scFv models produced by AlphaFold2. We then developed a series of Python scripts that accept scFv and whole antigen protein sequences as inputs, parse the antigen protein sequences into short overlapping peptides, run batch predictions for each scFv:peptide pair, and output two peptide scoring schemes based on the peptide per-residue pLDDT scores as a metric for AlphaFold2 model confidence. Binding of the expected epitope to the WT-Myc scFv could only be detected via the consensus method, but either analysis method could readily detect the expected epitope bound to the chimeric Myc scFvs. Conversely, the alternate analysis method (Simple Max) performed better with respect to ranking the expected HA epitope binding to the WT and chimeric anti-HA scFv variants. In the HA case, performance was comparable for both the WT and chimeric scFv variants.
It is important to note that binding of scFv variants to sequences other than the expected epitopes may be statistically unlikely but not impossible. For example, consider the peptide ATMPLNVSFT near the N-terminus of the Myc proto-oncogene protein sequence. In the context of the WT anti-Myc scFv this peptide had slightly higher average consensus pLDDT (52.4 rather than 51.0) than a peptide (QKLISEEDLL) that closely matched the expected epitope. In the absence of direct experimental evidence, predicted affinity for this unexpected sequence is not necessarily incorrect, though the lack of comparable predicted binding to the 15F11 and 2E2 chimeric scFv variants further decreases the likelihood. In the future, it might be useful to assess peptide binding via consensus across scFv variants.
Lastly, we tested this process on a novel antibody generated by our group targeting the SARS-CoV-2 nucleocapsid protein (mBG17) and found the method performed significantly better than with Myc and HA. Either analysis method could very easily flag peptide windows containing the authentic experimentally validated epitope. This worked for the WT scFv, the chimeric scFv variants, and even a structure with disconnected heavy and light chain domains. Experimentally, we cleanly validated the AlphaFold2 prediction using a peptide competition ELISA assay to experimentally determine the mBG17 epitope. Confidence in the AlphaFold2 prediction was further buoyed via alanine scanning peptide competition ELISAs that verified the importance of the key binding interactions predicted by AlphaFold2.
Identification of antibody VH and VL sequences from monoclonal B-cells has become a routine task, with sequence information obtainable via various sequencing technologies such as next generation sequencing and nanopore sequencing for a relatively low cost. As a result, the determination of the epitope in service of a deeper understanding of how antibodies bind their antigen is an increasingly notable bottleneck. An experimental epitope determination campaign can take weeks or months of work, but with the advent of AlphaFold2 and the epitope prediction method we describe here, an antibody and its antigen could be sequenced in a few days (often through contract research organizations for low cost) and accurate linear epitope predictions generated within less than a day, dramatically epitope validation throughput as well as providing detailed predictions for the molecular features of antibody-epitope interaction.
Conformational epitopes are structured antigens that are found during many immune responses, and prediction of these epitopes from antibody and antigen sequences would be a significant boon to the field of biology. For example, conformational epitope prediction coupled with single-cell B-cell sequencing would allow for detailed analysis of antibody maturation during immune responses to vaccines or pathogen infection, helping better define how the immune response to infection evolves over time and how evolution of antigen sequences affects the antibody response. In this work we did not focus on using AlphaFold2 to predict conformational epitopes primarily because of the complex structures that conformational epitopes possess. Literature reports suggest that prediction of the complexes between antibodies and both whole antigens and conformational epitope proteins has proven to be very difficult for AlphaFold2, and indeed the authors themselves make this observation (12, 42, 43). Notably, the structures that proved most difficult to predict for AF2 and other tools in the CASP15-CAPRI154 challenges were antibody-antigen complexes (44). Reports suggest that a mix of both statistics-based approaches (neural networks like AF2) and physics-based approaches (such as Rosetta) predict optimal antibody-antigen complexes (45). Indeed, if we attempt to predict binding of our scFvs to intact antigen proteins (Supplemental Figure 2), we find no predictive capability. When predicting scFv:peptide complexes, it may be the case that AlphaFold2 is able to thoroughly evaluate an induced fit for the peptide due to both its length (small sample space) and its propensity to not adopt a strong competing structure. In contrast, a larger and complicated structure may be more challenging to move during the AlphaFold2 structure prediction or recycle steps. Additional complexities may arise in extreme induced conformational changes during docking. Recent reports indicate that progress is being made in predicting the binding locations of conformational epitopes (46, 47).
We observed that the ability of AlphaFold2 to successfully predict the epitope peptide binding is quite delicate. First, epitope prediction was highly sensitive to the peptide length (Supplemental Figure 5), with minimal predictive power for peptide length other than 10 a.a. Further investigation of this sensitivity would be a useful avenue for future research. Perhaps with enhanced sampling, epitopes can be detected within longer peptides (e.g. 11 a.a., 12 a.a., etc.). Methodological tuning of this type could ultimately help illuminate the path to increasingly difficult protein-protein binding prediction problems. Similarly, we have likewise determined that epitope scanning performance was sensitive to changes in the underlying AlphaFold2 neural networks and the MSA. Specifically, unless otherwise noted, all data in this report was obtained using ColabFold version 1.5.2 and the 5 neural networks that comprise AlphaFold2 multimer version 2 (mm2). Likewise, the MSAs we use were obtained from the MMSEQS server (and cached) when the default sequence databases were UniRef30 2202 and PDB70 220313. They have since been updated to PDB30 2302 and PDB100 230517. For a complete description, see the change logs on the github for ColabFold (https://github.com/sokrypton/ColabFold#colabfold---v152).
Insofar as protein-peptide prediction is an emergent “off-label” capability for AlphaFold2 that is not part of the training sets, further training of the models or other changes can degrade performance. Benchmarking performance can be difficult when there are multiple moving targets. The most recent calculations we have analyzed were using ColabFold version 1.5.2 which was current as of February 19, 2023. The changes from ColabFold 1.5.2 to 1.5.5 (current as of this writing) are limited to version control and ensuring ColabFold still works on Google Colab, and therefore will not change the calculation performance. Relative to ColabFold 1.3 (the current method at the outset of this project), ColabFold 1.5.2 embodied two substantial changes. First, ColabFold 1.5.2 used the updated AlphaFold multimer (mm) version 3 by default. Second, the backend server MMSEQS ((48) and (https://github.com/soedinglab/MMseqs2)) that supplies MSAs also underwent updates, namely the database updates. Upon evaluation, we found that the recent default methods (ColabFold 1.5.2) still predicted the epitope successfully for the mBG17 system (Supplemental Figure 8). However, the ColabFold 1.5.2 default methods had a pronounced decline in PAbFold performance for the HA and Myc systems. Specifically, the combination of mm3 and the revamped ColabFold MSA server tended to be less discriminating compared to the default settings for ColabFold 1.3 (ColabFold 1.3 was the most up to date version when this project was initialized). The updated configuration flagged diverse peptide sequences with elevated pLDDT values (Supplemental Figures 9 and 10) resulting in the loss of successful epitope predictive power. While testing ColabFold 1.5.2 with the most recent MSA server, but reverting the AlphaFold2 models to mm2, the outcome improved, with experimentally validated sequences rising to the top more frequently than when using mm3, but still falling short in ranking the experimentally validated epitope sequence embedded within the antigen. However, when previously cached MSAs were paired with mm2 (using ColabFold 1.5.2), performance was maximized. Furthermore, we attempted to recreate the MSA databases locally with similar but not identical results to queueing the server with databases UniRef30 2202 and PDB70 220313 (Supplemental Figure 11). Additionally, the MMSEQS team ((48) and (https://github.com/soedinglab/MMseqs2)) graciously rebuilt a server we could query using LocalColabFold that mimicked the original UniRef30 2202 and PDB70 220313 database set up as closely as possible on their end. The MSA that was generated from these databases was used, and still did not perform as well as the original MSAs that were generated upon first retrieval and generation (Supplemental Figure 12). As a negative control, we repeated all calculations without using any MSAs, and only relying upon the sequence to make a structural prediction. As expected, all epitopes were scored very poorly (Supplemental Figure 13). Despite our significant efforts, it is unclear why our initial results cannot be perfectly recapitulated, but the difference has been traced to detailed MSA contents (Supplemental Figure 14), resulting in differences in correct epitope identification. These results are summarized in (Supplemental Figure 15).
One key lesson of this research effort is that caching the MSAs proved to be very useful as a method to guard against changes in the performance of 3rd party tools. We recommend that future methods development work using LocalColabFold adopt the strategy of caching MSAs when feasible. It is also our hope that by describing the latent ability of AlphaFold2 to predict scFv-binding epitopes that this ability will be preserved and enhanced in future iterations.
The authors would like to thank members of the Snow, Stasevich, and Geiss groups for helpful discussions at TagTeam meetings. This project was funded in part by NIH R01AI132668 (Geiss) and NIH R56AI155897 and R01AI168459 (Snow, Stasevich, Geiss). The authors would also like to thank Dr. Milot Mirdita for being our point of contact for the ColabFold and MMSEQS teams, and assisting us with setting up databases and generating MSAs for our use.
Contents:
Table 1A: Full sequence information for all scFv and antigen proteins
Table 1B: MSA for the scFv chimera variants, with loop and linker region annotation
Figure 1: Comparison of AlphaFold2 Myc scFv predictions to Fab crystal structure
Figure 2: AlphaFold2 predictions for scFv interacting with full length antigen proteins
Figure 3: Illustration of AlphaFold2 peptide predicted placements and confidence thereof
Figure 4: Structure superposition analysis for Myc and HA scFv variants relative to reference crystal structures
Figure 5: In the context of Myc, testing prediction performance versus sliding peptide window parameters
Figure 6: Testing detection of Myc epitope inserted into three locations in an unrelated 3rd-party protein
Figure 7: HA epitope prediction for three anti-HA scFvs
Figure 8: Comparing prediction performance for mBG17 using multimer-v2 and multimer-v3
Figure 9: Comparing prediction performance for Myc using multimer-v2 and multimer-v3 and the new MSA
Figure 10: Comparing prediction performance for HA using multimer-v2 and multimer-v3 and the new MSA
Figure 11: Comparison of 9 major systems after recreating MSAs locally with downloaded databases
Figure 12: Comparison of 9 major systems after recreating MSAs with colabfold after MMSEQS rebuilt the old databases for our use
Figure 13: Comparison of the 9 major systems without using any MSA, using only the single sequence
Figure 14: Comparison of the contents of the MSA for Myc-2E2 after being generated by the 4 major methods: old generation, new generation, local generation, and MMSEQS rebuilt specialty server.
Figure 15: Overview table of whether or not each MSA generation type could accurately detect the experimentally determined epitope in each of the 9 major systems.
