Unveiling Interaction Signatures Across Viral Pathogens through VASCO: Viral Antigen-Antibody Structural COmplex dataset

15 16 Viral antigen -antibody (Ag -Ab) interactions shape immune responses, drive pathogen 17 neutralization, and inform vaccine strategies. Understanding their structural basis is crucial for 18 predicting immune recognition, optimizing immunogen design to induce broadly neutralizing 19 antibodies (bnAbs), and developing antiviral therapeutics. However, curated structural 20 benchmarks for viral Ag-Ab interactions remain scarce. To address this, we present VASCO (Viral 21 Antibody-antigen Structural COmplex dataset), a high -resolution, non-redundant collection of 22 ~1225 viral Ag-Ab complexes sourced from the Protein Data Bank (PDB) and refined via energy 23 minimization. Spanning Coronaviruses, Influenza, Ebola, HIV, and others, VASCO provides a 24 comprehensive structural reference for viral immune recognition. By comparing VASCO against 25 general protein-protein interactions (GPPI), we identify distinct sequence and structural features 26 that define viral Ag -Ab binding. While conventional descriptors show broad similarities across 27 datasets, deeper analyses reveal key sequence-space interactions, secondary structure preferences, 28 and manifold-derived latent features that distinguish viral complexes. These insights highlight the 29

of GPPI -trained predictive models and the need for specialized computational 30 frameworks. VASCO serves as a critical resource for advancing viral immunology, improving 31 predictive modeling, and guiding immunogen design to elicit protective antibody responses. By 32 bridging sequence and structural immunological datasets, VASCO should enable better docking, 33 affinity prediction, and antiviral therapeutic development —key to pandemic preparedness and 34 emerging pathogen response. 35 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint

36 37 Antigen (Ag) and antibody (Ab) interaction is a fundamental feature of our immune system [1], 38 wherein antibodies evolve to recognize and bind to antigens, form ing stable complexes and 39 ultimately perturbing antigenic function and facilitat ing their neutralization. Beyond their 40 biological significance, antibodies have also become invaluable tools in laboratory settings and are 41 increasingly employed as therapeutic agents[2]. By virtue of their capabilities to recognize antigen 42 epitopes with high specificity [3], and their modular anatomy [4], these antibodies form essential 43 targets for diagnostic and therapeutic mechanisms as well as experimental biology assays [5-7]. 44 The burgeoning field of antibody engineering seeks to understand Ag-Ab binding, stability, and 45 immunogenic properties. Yet this field suffers from expensive resource requirements, biosafety 46 concerns, and research reproducibility crisis [8], making c omputational methods essential 47 alternatives in this endeavor . While computational tools have proven instrumental in antibody 48 engineering, the reliance on antibody sequence s pecific data rather than Ag-Ab structural 49 complexes has become a prevailing trend due to greater data availability [9]. To facilitate antibody-50 related computational analyses, various databases, such as DIGIT [10], IMGT [11], Abysis [12], 51 Antibodypedia[13], and Antibody Registry [14] have been established, each offering distinct 52 advantages in terms of sequence information. However, to get a molecular level understanding of 53 antibody response, it is crucial to consider structural nuances that govern Ag -Ab interactions 54 beyond just the sequence space. Such structural information can provide essential understanding 55 about the conformational topology, dynamics, and binding affinities that underpin this immune 56 response, facilitating rational design of vaccines and therapeutic antibodies. 57 58 The growth of publicly available conformational data for Ag -Ab complexes in the Protein Data 59 Bank (PDB) [15] reflects the recognition of the pivotal role of structural information of 60 interactions. The PDB has witnessed a substantial surge in antibody structures, exponentially 61 increasing over the years, and now constitutes approximately 4.2% of the total entries ( March 62 2025). Many contributions have already been made by datasets such as SAbDAb [16], Thera -63 SAbDAb[17] and IEDB-3D[18]. However, this increasing data availability opens the possibility 64 of exploring underlying patterns in antibody – antigen interaction complexes whichdemands the 65 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint benchmarked curation of structural data for specific classes of proteinsthat goes beyond sequence 66 or antibody information, but more of a focus on the interfaces of interactions. 67 68 A critical subset of Ag -Ab interactions that deserve detailed scrutiny are viral antibody-antigen 69 interaction complexes. The frontline defense in virus-host interactions entails antibodies binding 70 to viral antigens to neutralize their function or recruit other immune components for targeted 71 destruction. In the realm of viral Ag-Ab interactions, resource, safety , and reproducibility 72 challenges are heightened. The dynamics of viral infections [19], coupled with the inherent 73 variability of viral strains [20], necessitate a meticulous understanding of Ag-Ab characteristics 74 where structural characterization of antibody -antigen complexes takes center stage. Antibodies 75 have a typical two-chain modular anatomy[21] where the antigen recognition site is largely limited 76 to the complementarity determining regions (CDRs) that include three hypervariable (HV) loops 77 from each chain. Based on their structural and sequence variability, the antigen binding region 78 (paratope) of antibodies personate their binding specificity [22], with much of the uncertainty 79 coming from these variable loops[23]. Moreover, it has been shown that not all the CDRs may be 80 involved in the interactions in many cases, or some parts of the paratope also fall outside the CDRs, 81 contributing to the intricacy of binding. Identification of viral envelope antigen residues that form 82 the binding interface to antibodies (i.e., the epitope region) is even more challenging due to an 83 apparent lack of common features. While general PPI prediction models have seen success in 84 protein docking and interaction scoring, they consistently struggle when applied to Ab -Ag 85 interactions, yielding lower performance[24, 25]. 86 87 The need of the hour is to curate and systematically evaluate structural data of viral Ag -Ab 88 interaction complexes along with comparative analys es against general protein -protein 89 interactions (PPI). This approach can shed light on the unique features and challenges posed by 90 viral interactions, offering valuable insights that extend beyond the immediate context of infection. 91 Such a c urated dataset will play a pivotal role in advancing viral immunology research by 92 providing a foundation for predictive modeling, therapeutic development, and a deeper 93 understanding of the complexities inherent in viral antibody responses. Understanding structural 94 patterns in viral antibody -antigen interactions is essential for evaluating the accuracy of 95 computational tools and predictive models. Specialized benchmarks, such as DOCK6[26], 96 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint Docking Benchmark 5.0 [27], and Affinity Benchmark versions 1 and 2 [27], already provide a 97 standardized platform for assessing the performance of algorithms in general PPIs. In contrast, by 98 establishing benchmarks tailored to the unique features of viral Ag-Ab associations, researchers 99 can refine and improve predictive capabilities specific to attributes of viral responses. Comparative 100 analyses with general PPIs will enable the identification of distinguishing features, such as the 101 sequence-structure topology of CDR loops and the adaptability of antibodies to diverse viral 102 strains. Understanding these distinctions can inform the development of targeted therapies and 103 contribute to our broader understanding of immune responses. 104 105 Curated datasets of viral Ag -Ab interaction complexes, evaluated against general protein-protein 106 interactions, will find applications across various scientific research domains. Insights gained from 107 curated datasets can aid in the design of targeted therapies and vaccines against viral infections. 108 By analyzing viral antibody -antigen interactions, researchers can identify conserved structural 109 motifs that may be crucial for vaccine efficacy. Understanding the structural patterns of effective 110 antibody responses can also inform the development of antiviral drugs and prophylactic 111 measures[28, 29] . The insights derived from these comparative analyses can contribute to the 112 identification of potential biomarkers and therapeutic targets. Structural patterns derived from 113 curated datasets can further contribute to the modeling and prediction of viral diseases and their 114 evolution. 115 116 In this paper, we present VASCO (Viral Antibody-antigen Structural COmplex dataset), a curated 117 collection of 1 225 high-resolution, non -redundant viral antigen -antibody (Ag -Ab) interaction 118 complexes in an energy-relaxed conformation close to their crystal structure local minimum, with 119 resolutions better than 5 Å. The dataset encompasses a diverse range of viral species, including 120 SARS-CoV-2, Influenza, Ebola, and HIV, providing a comprehensive structural reference for viral 121 immune recognition. Additionally, VASCO includes a detailed set of structural and 122 physicochemical features relevant for training, validation, and testing of machine learning models, 123 facilitating predictive modeling of antibody-antigen interactions. To assess the distinct properties 124 of viral Ag-Ab interactions, we compared VASCO against two control datasets of general protein-125 protein interactions (GPPI), comprising 2000 heterodimeric and homodimeric complexes. While 126 conventional structural features —such as contact surface area, hydrogen bonding, and non -127 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint covalent interactions—showed broad similarities between datasets, deeper analyses uncovered key 128 sequence-structure signatures and manifold -derived latent features that distinguish viral Ag -Ab 129 interfaces. These findings underscore the limitations of existing GPPI-based predictive models in 130 capturing the complexities of Ag -Ab binding and emphasize the need for specialized 131 computational frameworks tailored to antibody -antigen interactions. By providing a benchmark 132 dataset and structural insights into viral immune recognition, VASCO serves as a valuable resource 133 for advancing viral antibody engineering, vaccine design, and computational immunology. 134 135

136 137 Constructing VASCO as a Dataset to Bridge the Gap in Ab-Ag Interaction Modeling 138 139 The VASCO dataset was curated through an exhaustive search of the Protein Data Bank (PDB) by 140 querying its API for structural data of viral antigen-antibody (Ag-Ab) complexes. We focused on 141 high-resolution complexes, selecting those with a resolution better than 5Å, which ensures 142 structural reliability for downstream analyses . The dataset comprises 1 225 non-redundant viral 143 antibody-antigen interaction complexes. To ensure structural consistency and reduce artifacts from 144 crystallographic, cryo-EM, or NMR-derived conformations, all structures in VASCO underwent 145 local energy minimization, preserving their native-like binding states while resolving steric clashes 146 or structural irregularities. The viral species represented in the dataset were further categorized as 147 follows (See Figure 1): SARS-CoV and CoV-2, MERS and related coronaviruses: 544 structures; 148 HIV: 183 structures; Influenza : 144 structures; Ebola: 26 structures; other miscellaneous viral Ag-149 Ab complexes: 103 structures. This dataset, referred to as VASCO (Viral Antibody antigen 150 Structural COmplex dataset), aims to support predictive modeling efforts for antibody -antigen 151 interactions, which are distinct from general protein -protein interactions (GPPI) due to the 152 structural complexity of antibody binding interfaces. In particular, antibody -antigen binding sites 153 exhibit highly variable topologies due to the complementarity -determining regions (CDRs) and 154 variable loops, making accurate predictions challenging using models developed for more general 155 protein-protein interactions. To highlight these differences, we complemented the VASCO dataset 156 with two control datasets of GPPI interfaces: heterodimers and homodimers, each comprising 2000 157 randomly selected interactions from the PDB, similarly filtered for structural resolution below 5 158 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint Å. These general protein-protein interaction datasets (GPPI) provide a basis for comparison to help 159 assess the unique structural properties of viral antibody-antigen interactions. The VASCO dataset 160 includes a range of typical protein interface descriptors such as contact surface areas, hydrogen 161 bond counts, and secondary structure involvement, which are standard in PPI classification. 162 However, initial comparisons between VASCO and the GPPI datasets showed no significant 163 deviations in most structural features across the datasets. Then the obvious question becomes, how 164 do the current state of the art GPPI interaction prediction methods fail to perform as well Ab -Ag 165 interaction prediction? Interestingly, deeper analysis revealed notable differences in specific 166 feature distributions: (i) Sequence space of the contact residues: Antibody -antigen interfaces 167 showed distinct patterns in sequence composition at the contact points compared to GPPI. (ii) 168 Secondary structure features: Certain differences emerged in the involvement of secondary 169 structural elements at the interfaces, with viral antibody -antigen interactions displaying more 170 variability. (iii) Manifold-derived latent features: Using dimensionality reduction techniques, we 171 identified latent features —complex representations derived from structural manifolds —that 172 exhibited significant differences between VASCO and GPPI datasets. These findings indicate that 173 while many general protein interface features overlap between viral Ag-Ab interactions and GPPI, 174 specific characteristics, especially in the sequence and structural context, set antibody -antigen 175 interactions apart, underscoring the need for specialized predictive models for these complexes. 176 177 Distinctive Amino Acid Contact Profiles Define Viral Antibody-Antigen Interface Signatures 178 179 Figure 2 presents the sequence contact maps comparing the frequency of amino acid contacts 180 between viral antibody-antigen (Ab-Ag) interfaces and general protein-protein interactions (GPPI) 181 from homodimer and heterodimer control sets. A striking observation is that while the contact 182 maps for the homo and hetero control sets are nearly indistinguishable (Figure 2b,c), the viral Ab-183 Ag contact profiles exhibit clear and distinct patterns. In the viral Ab-Ag interfaces, serine (Ser) 184 and tyrosine (Tyr) residues from the antibodies consistently form high -frequency contacts with a 185 variety of antigen residues. Additionally, glycine (Gly) from the antibody shows a high occurrence 186 of contacts with asparagine (Asn), tyrosine (Tyr), and glycine (Gly) on the antigen surface (Figure 187 2a). The significance of Tyr, Ser and Gly on Ab CDRs have been widely noted in literature [30-188 32]. Among charged residues, the Lys-Asp (lysine from antigen and aspartic acid from antibody) 189 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint pair is prominent, standing out as a frequent contact. In contrast, the control sets show a much 190 more diverse range of contact pairs, with arginine (Arg) and leucine (Leu) dominating the profile 191 by making frequent contacts with several amino acids, followed by Arg-Phe (phenylalanine) pairs 192 from both sides of the protein interface. Interestingly, in the HIV Ab -Ag interfaces, the serine 193 population diminishes while glycine contacts from both the antibody and antigen sides increase 194 significantly (Figure 2d). Glycine mutations have been known to increase the potency and breadth 195 of HIV-1 broadly neutralizing antibodies [33, 34].Despite this shift, the overall contact profiles 196 remain fairly consistent across the different viral species, underscoring the unique signature of 197 viral antibody-antigen interactions. This distinct contact profile, dominated by Ser, Tyr, and Gly 198 from antibodies interacting with specific antigen residues, highlights the unique structural features 199 of viral Ab-Ag interfaces. The recurrence of these specific residue interactions across viral species 200 suggests that pairwise contact frequency signatures encode critical information about antibody 201 recognition. These patterns should be explicitly incorporated into predictive modeling approaches, 202 as they may offer a more biologically relevant and interpretable basis for scoring and ranking 203 antibody-antigen interactions. 204 205 Ab-Ag Interactions Share Interface Properties with General PPIs but Highlight the Need for 206 Specialized Models 207 208 To investigate the structural characteristics of the viral antibody -antigen (Ab-Ag) interactions in 209 the VASCO dataset, we compared the contact surface area (CSA), number of inter-residue non-210 covalent contacts, hydrogen bonds, and intermolecular contacts per unit area with those of general 211 protein-protein interactions (GPPI) from homodimer and heterodimer datasets. These comparisons 212 are illustrated in Figure 3, with panels showing the results for each feature. In Figure 3a, we 213 compare the CSA across the full VASCO set, individual viral species (SARS, Influenza, Ebola, 214 HIV), and the GPPI control sets. While the median CSA is not significantly different between the 215 viral Ab-Ag complexes and the GPPI datasets, the variability is noticeably greater in the control 216 sets. This broader spread of CSA values in the homodimer and heterodimer datasets reflects more 217 218 diverse surface area exposure in general PPIs, whereas viral antibody-antigen interactions are more 219 uniform in terms of accessible surface area. This is not very surprising, given the strong structural 220 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint similarities between different Fabs. Figure 3b presents the number of non -covalent contacts 221 between the binding molecules . While the me dian number of contacts is consistent across all 222 groups, the quartile ranges is again significantly larger for the homodimer and heterodimer 223 datasets, indicating that general protein-protein interfaces exhibit greater variability in the number 224 of non-covalent contacts compared to the more constrained viral Ab -Ag interactions. Figure 3c 225 presents the number of hydrogen bonds at the interfaces. Here too, the me dian values are 226 comparable across the viral and GPPI datasets , with significantly larger lower-to-upper quartile 227 spread in the GPPI. Lastly, in Figure 3d, we compare the average presence of contacts per unit 228 area of interaction. Unlike the other three measures, here the variabilities are comparable between 229 the VASCO and the control GPPI sets. Similarly, H-bonds per unit area of contact also remained 230 similar between different viral Ag-Ab and general proteins (Supplementary Figure S1a). These 231 findings indicate that commonly used interface descriptors —contact surface area, non -covalent 232 interactions, and hydrogen bonds —are not fundamentally different between viral A g-Ab 233 interactions and general PPIs. Instead, they largely scale with the available interaction surface area. 234 Given this similarity, the consistently poorer performance of ML -based docking and scoring 235 models on Ab -Ag complexes remains an intriguing challenge [REF]. This suggests that the key 236 distinguishing factors for Ab-Ag interactions may lie beyond these traditional structural features, 237 underscoring the need for specialized predictive models and alternative feature selection strategies 238 tailored to the unique constraints of antibody-antigen recognition. 239 240 β-Strands and Turns Dominate Viral Ag-Ab Interfaces Over Helical Content 241 242 Figure 4 illustrates the secondary structure composition of the viral antibody -antigen (Ab-Ag) 243 interfaces in the VASCO dataset compared to the general protein -protein interaction (GPPI) 244 control sets. Our analysis reveals a higher propensity for β -bridges and turns within the viral 245 interfaces (Figure 4a), while showing significantly lower helical content in comparison to both 246 homodimer and heterodimer datasets (Figure 4b,c). Both homodimeric and heterodimeric protein 247 interactions on the other hand have helices as the main structural features at interfaces, also 248 reported earlier[35]. The comparative data indicate that the viral Ab-Ag interfaces favor β-strands 249 and turns, which may contribute to the stability and specificity of these interactions. This is in stark 250 contrast to the GPPI control sets, where a more balanced distribution of secondary structures is 251 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint observed, characterized by a greater prevalence of helical content. The higher beta -bridge 252 occurrence stems from the viral antigen surface, not from the antibody, which mainly contacts 253 through the disordered CDRs. Notably, the control sets exhibit similar values across different 254 secondary structure distributions, reinforcing the uniformity of their structural characteristics , 255 independent of homomeric or heteromeric composition . Interestingly, the Ebola interactions 256 display a unique profile, with helical content values that fall within the range observed in the GPPI 257 control sets (Supplementary Figure S1b). This suggests that the Ebola virus may possess distinct 258 structural features at its Ab -Ag interfaces, potentially influencing its immune evasion strategies 259 and the design of therapeutic antibodies. Overall, these findings highlight the specific structural 260 adaptations present in viral antibody -antigen interactions, marked by an increased occurrence of 261 β-strands and turns alongside reduced helical content compared to general protein -protein 262 interactions. This distinct secondary structure composition is critical for understanding the 263 dynamics and mechanisms underlying viral immune responses. 264 265 266 Theoretical Interaction Energies Reveal Stability Trends between Viral and General 267 Interfaces 268 269 Interaction energies and affinities can be highly informative for understanding the strength and 270 specificity of antibody -antigen interactions, providing quantitative measures that can guide 271 predictive modeling, improve docking accuracy, and aid in the design of high-affinity therapeutic 272 antibodies. Although 1225 Ab-Ag structures were curated, acquiring corresponding experimental 273 binding affinities or energies proved challenging. We could find less than 5% of these complexes 274 that had experimentally determined binding energies available in the literature, demonstrating the 275 difficulty of obtaining standardized, comparable energy values for such interactions. Given the 276

of this study —to provide a comprehensive dataset for data -driven Ab -Ag interaction 277 prediction—it is crucial to include a theoretical measure of comparative binding stability. Despite 278 its limitations in predicting absolute energy values, Molecular Mechanics Poisson -Boltzmann 279 Surface Area (MMPBSA) method is effective in estimating relative binding energy differences, 280 making it suitable for this analysis. Figure 5 presents the results of theoretically calculated binding 281 free energies for antibody -antigen (Ab -Ag) complexes in the VASCO dataset, using implicit 282 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint solvent MMPBSA energy calculations. The calculated binding energies were subsequently scaled 283 based on the available experimental data (~5% of the dataset) to ensure a reasonable comparison. 284 As observed in several other structural features in this study, the trends in binding energies between 285 the VASCO dataset and the control sets (homodimer and heterodimer protein-protein interactions) 286 are within standard deviations of each other. However, there are several noteworthy observations. 287 The median binding energies of viral Ag -Ab complexes fall between those of homomeric and 288 heteromeric protein-protein interactions, with homodimers being the least tightly bound (Figure 289 5, top left). Electrostatic interactions contribute the most to binding stabilization in both VASCO 290 and heteromeric complexes. In contrast, as expected, van der Waals forces serve as the primary 291 stabilizing factor in homomeric interactions but play a comparatively smaller role in viral Ag -Ab 292 binding stability. Additionally, viral complexes exhibit distinct solvation energy characteristics 293 (Figure 5, bottom row) —polar solvation energies are less destabilizing, while nonpolar solvation 294 energies provide less stabilization compared to general PPIs. 295 296 One notable observation is the larger variation in binding energies within the control sets, which 297 exhibit a broader range of contact surfaces and interactions. In some cases, particularly in the 298 control sets, energies even show positive values, indicating unstable interactions. These instances 299 highlight the necessity of caution when using experimentally resolved structures (from X -ray 300 crystallography, NMR, or cryoEM) for predictive modeling, as some structures may represent 301 suboptimal, non -stabilized conformations. All energy calculations were conducted following 302 extensive local structure minimizations using the CHARMM36m force field, ensuring accurate 303 theoretical predictions. However, this analysis emphasizes the need for careful treatment of 304 structural and energetic data in any predictive or scoring model, as some structures deviate from 305 their most stable orientations. 306 307 308 Manifold Learning Reveals Hidden Differences Between VASCO and GPPI interfaces that 309 otherwise remain elusive. 310 311 Despite limited distinctions found using conventional structural features to describe antibody -312 antigen (Ab-Ag) interactions, these complexes continue to underperform in general protein-protein 313 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint interaction (GPPI) binding predictions [REF]. This discrepancy suggests that critical variations in 314 interface architecture may remain hidden in higher -dimensional space, necessitating a more 315 nuanced analysis. To investigate this, we applied several dimensionality reduction techniques to 316 explore patterns in the data that traditional handcrafted descriptors may miss (Figure 6 and 317 Supplementary Figure S2). Principal component analysis (PCA), one of the most commonly used 318 dimension reduction techniques[36], failed to capture any significant differences between the viral 319 Ab-Ag interfaces and the GPPI control sets (Supplementary Figure S2a). This lack of distinction 320 was further echoed when we employed nonlinear dimensionality reduction technique of t-321 distributed stochastic neighbor embedding (t-SNE)[37] (Supplementary Figure S2 b). Neither 322

could meaningfully separate viral interactions from the homo and hetero GPPI controls in 323 the reduced feature space. However, employing the isomap method [38] of checking geodesic 324 distances rather than flat Euclidian variations between datapoints, we observed clear distinctions 325 in the architecture of Ab -Ag interfaces. This was further verified by spectral embedding of the 326 manifold[39]. In both methods, the top two embedding coordinates revealed that viral Ab -Ag 327 interactions form a distinctive cluster, markedly different from those of GPPI controls (Figure 6a, 328 6b). This suggests that these nonlinear techniques can capture subtle variations in the interface that 329 remain elusive in linear and conventional approaches. Interestingly, within the VASCO dataset, 330 HIV Ab-Ag complexes also formed a distinct cluster (Figure 6c,d). This observation aligns with 331 findings from the sequence contact map analysis (Figure 4), further reinforcing the uniqueness of 332 HIV interactions relative to other viral complexes. Based on these manifold-driven latent features, 333 we developed a probability map in the spectral embedding space, spanned by the top two 334 coordinates. This map assigns a negative log -probability score to any interface, providing a 335 measure as to whether the interface likely belongs to an Ab-Ag interaction (Figure 6e) or a GPPI 336 control (Supplementary Figure S2c), or even the specific subset of HIV interfaces (Figure 6f). 337 This classification tool can be applied to novel protein interfaces, offering a powerful way to 338 categorize them based on structural likelihood. Thus, by employing nonlinear dimensionality 339 reduction techniques, we uncovered critical differences in viral Ab -Ag interface architecture that 340 traditional hand-picked descriptors and linear methods failed to detect. These findings underscore 341 the value of manifold learning as a feature selection tool, capable of capturing complex structural 342 patterns that are crucial for Ab-Ag interaction prediction. 343 344 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint Spectral Embedding Captures Distinct Structural Modes of Viral Ab-Ag Interfaces 345 346 To further explore the distinctions uncovered by spectral embedding, we extracted the most 347 representative structures from the three extreme corners in the triangular manifold projection of 348 the spectral embedding space. This allowed us to visualize how the Ab -Ag interfaces from the 349 VASCO dataset vary along the most informative dimensions ( Figure 7 ). These representative 350 structures were selected based on their proximity to the mean conformation (by RMSD) within 351 each extreme region of the embedding space ( Figure 7a), providing insights into the structural 352 diversity across viral interfaces. Among these three extremes, the top cluster (Set-T) is the most 353 populated, containing 50% of the VASCO dataset, while the bottom left cluster (Set-L) comprises 354 15% of the interfaces. The bottom right cluster (Set-R) is sparsely populated, representing only 355 3% of the dataset, with the remaining structures distributed throughout the embedded space. 356 Examining the structural differences between these clusters, we find that the Fab architecture 357 remains largely consistent, with the primary variations occurring at the antibody-antigen interface 358 (Figure 7b). The CDR loops in Set-T and Set-L exhibit similar alignments, though Set-T shows a 359 slightly more splayed -out heavy chain conformation. In contrast, the Set -R interfaces display a 360 more disordered loop arrangement, with CDR loops protruding outward more prominently. 361 The nature of antigen contact varies significantly across these clusters (Figure 7c). In Set-T, the 362 antigen surface engages both the heavy (H) and light (L) chain CDR loops equally, forming a large 363 surface area of interaction. In Set -L, the contact area is relatively smaller and is slightly biased 364 toward the light chain of the Fab. Meanwhile, in Set-R, which is the least populated conformational 365 state, the antigen makes minimal contact, predominantly with the heavy chain, leading to a 366 significantly reduced binding interface. These findings suggest that while the majority of viral Ab-367 Ag interactions favor stable, broad-contact conformations (as seen in Set-T), a subset adopts more 368 asymmetric or sparse interactions (Set -L and Set -R). The relatively low occurrence of Set -R 369 structures highlights the instability or rarity of such binding modes, reinforcing the importance of 370 interface-specific constraints in viral antibody recognition. This analysis demonstrates that spectral 371 embedding effectively captures structural heterogeneity in antibody-antigen interfaces, offering a 372 useful framework for categorizing binding modes that could inform predictive modeling and 373 antibody design. 374 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint

375 376 In this study, we have presented VASCO, a curated dataset of viral anti gen-antibody (Ag-Ab) 377 structural complexes, comprising over 1 225 non-redundant structures that have been energy -378 minimized to resolve steric clashes and improve geometric consistency. This dataset spans 379 multiple viral families, including SARS, Influenza, Ebola, and HIV, and captures high -resolution 380 structural details of antibody interactions with viral antigens. To contextualize these interactions, 381 we compare the dataset against general protein -protein interaction (GPPI) control sets, including 382 heterodimeric and homodimeric protein complexes. VASCO provides a comprehensive structural 383 representation of human antibody interaction with viruses , offering insights into antibody 384 recognition patterns, viral epitope characteristics, and motifs unique to viral A g-Ab binding. By 385 integrating traditional structural descriptors with advanced manifold techniques, this dataset serves 386 as a valuable resource for studying antibody-antigen interactions and developing predictive models 387 for viral antibody engineering and therapeutic design. 388 389 Our study reveals that while many conventional descriptors do not show significant differences 390 between Ab -Ag and GPPI interfaces, key latent features extracted through nonlinear 391 dimensionality reduction (e.g., isomap and spectral embedding) clearly differentiate the viral 392 complexes. These findings suggest that these manifold embeddings should be used to complement 393 the traditional descriptors in designing a predictive model for Ab -Ag interfaces to attain superior 394 performance from existing GPPI scoring methods. 395 396 The curated dataset presented here holds significant implications for advancing the field of 397 antibody engineering and pandemic preparedness. While the predominant contribution to VASCO 398 comes from SARS -related antibodies, this does not limit the dataset's applicability, as key 399 sequence-structure features remain largely consistent across different viral families, as 400 demonstrated in our comparative analyses. The shared binding properties of viral Ag -Ab 401 interactions, including contact residue compositions and secondary structure preferences, suggest 402 that SARS -CoV-2 serves as a representative model for studying antibody -antigen interactions 403 more broadly. Moreover, the recent COVID -19 pandemic has significantly expanded the 404 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint availability of high -resolution SARS -CoV-2 antibody-antigen complexes, enriching the dataset 405 and allowing for deeper structural insights that extend to other viral systems. 406 407 By providing detailed insights into the structural determinants of viral antibody -antigen 408 interactions, this dataset can serve as a vital resource for researchers working on therapeutic 409 antibody design. Structural differences captured in this dataset will help in improving 410 computational models for predicting antibody binding, guiding rational antibody design with 411 enhanced specificity and binding affinity. The VASCO dataset specifically focuses on viral Ab -412 Ag interactions, a critical component for understanding immune responses during viral infections. 413 In the context of pandemics, such as COVID -19, studying how antibodies bind to viral antigens 414 can enhance our ability to predict the course of infections and to design therapeutic interventions, 415 including vaccines and monoclonal antibody therapies. This dataset provides the structural 416 foundation for evaluating engineered antibodies and optimizing their therapeutic potential. Given 417 the global relevance of antibody design in combating viral threats, this dataset can inform the 418 development of more effective engineered antibodies. By analyzing the structural features of 419 successful antibody interactions with viral epitopes, this resource will enable the engineering of 420 antibodies with improved binding affinities and therapeutic potential. 421 422 While computational techniques such as docking algorithms and binding affinity predictions are 423 essential tools for studying protein-protein interactions, the specific challenges of modeling viral 424 Ab-Ag interfaces require tailored approaches. Benchmarks like Docking Benchmark 5.0, Affinity 425 Benchmark 1 and 2, and DOCKGROUND have provided useful platforms for testing 426 computational models, but the representation of viral Ab -Ag complexes in these benchmarks 427 remains limited. Our dataset aims to bridge this gap by providing a viral -specific benchmark for 428 evaluating the performance of docking and affinity prediction algorithms. Some datasets have been 429 developed for studying antibody -antigen (Ab-Ag) interactions, though they are not specifically 430 focused on viral Ab -Ag complexes. Notable examples include the PECAN dataset [40], which 431 consists of 460 structures, partitioned into 205 for training, 103 for validation, and 152 for testing; 432 the PARAGRAPH dataset[41], containing 1,086 complexes, with 60% allocated for training, 20% 433 for validation, and 20% for testing; and the MIPE dataset [42], comprising 626 structures, with 434 90% used for five-fold cross-validation and 10% reserved for testing. While these datasets provide 435 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint valuable benchmarks for Ab -Ag modeling, they encompass a broad range of antigen types, 436 including non-viral targets. In contrast, VASCO is specifically tailored to viral Ab-Ag interactions, 437 capturing the unique structural characteristics of antibodies binding to viral antigens. This 438 specialization makes it a crucial resource for improving predictive models in the context of viral 439 immunity and therapeutic antibody design. The diversity and structural uniqueness of viral Ab-Ag 440 interfaces pose significant challenges for current computational methods. By incorporating the 441 variability and dynamics of viral Ab -Ag complexes into predictive models, this dataset has the 442 potential to improve the accuracy of antibody design and virus neutralization predictions. 443 444 One of the key obstacles in Ab-Ag interaction modeling is the structural flexibility of CDR loops, 445 which play a central role in antigen recognition. Viral antibodies often exhibit a higher degree of 446 structural flexibility to adapt to the diverse and dynamic nature of viral antigens. Understanding 447 and incorporating these dynamics into predictive models is critical for capturing the full range of 448 possible interactions. As a result, docking and scoring methods trained on general PPI datasets fail 449 to capture the key determinants of antibody binding, leading to substantial performance gaps. A 450 review of docking studies, including RosettaDock [43], ZDOCK[44], IRAD[45], and AlphaFold-451 Multimer[46], reveals that while reported success rates for general PPI predictions range from 35% 452 to 90%,[43, 44] the performance drops considerably for Ab -Ag interactions, with success rates 453 between 19% and 63% [43, 46]. This disparity highlights the limitations of existing models and 454 underscores the need for dedicated structural datasets, such as VASCO, to enable the development 455 of machine learning and computational approaches specifically tailored for antibody -antigen 456 interactions. 457 The key findings from the dataset analysis are as follows. The pairwise contact sequence signature 458 in viral Ag -Ab interactions is distinct from general GPPI, reflecting the specialized nature of 459 immune recognition. Secondary structure analysis indicates not only disordered coils and turns, 460 but also a preference for β -bridges in antigen binding regions, contrasting with the more diverse 461 structural motifs in GPPI. Manifold embedding reveals consistent patterns across viral species, 462 clustering distinctly from GPPI that shows greater divergence, highlighting fundamental 463 differences in interaction landscapes. Interestingly, HIV exhibits a unique interaction pattern, 464 likely due to its heavily glycosylated envelope, which constrains accessible antibody-binding sites. 465 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint 466 Moving forward, the VASCO dataset offers numerous opportunities for research and application. 467 As researchers develop more sophisticated machine learning algorithms for antibody -antigen 468 interaction prediction, this dataset will be instrumental in training and testing these models. By 469 integrating data-driven approaches with the structural features highlighted in this study, future 470 predictive tools will be better equipped to model the complexities of viral immunity. Moreover, 471 the dataset's focus on viral interactions provides an invaluable resource for pandemic readiness. 472 By expanding our understanding of how antibodies interact with viral antigens, we can better 473 anticipate viral mutation impacts, refine therapeutic interventions, and prepare for emerging viral 474 threats. 475 476 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint FIGURES 477 Figure 1 478 Figure 1: The VASCO set of energy -minimized viral antigen -antibody structural complexes. (a) Schematic and actual structural representations of viral antigens interacting with antibody Fab domains, with a zoom -in on the example contact interface of SARS -CoV2 spike interacting with H014 Fab (PDB 7CAC). The immunoglobulin full structure was aligned by the fab domain for representation purposes (PDB 1IGY). (b) Composition of the VASCO dataset. (c) Composition of the comparison set of viral Ag-Ab with general PPI complexes. .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint Figure 2 479 480 Figure 2: Pairwise contact profile between different amino acids at interaction interfaces. (a) Sequence contact pair frequencies at viral Ag -Ab interface, (b) homodimer, and (c) heterodimer interaction surfaces. (d) Virus-specific sequence contact pair profile. 481 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint Figure 3 482 483 Figure 3: Structural contact features compared between VASCO and control general inter-protein interactions. (a) Median and quartile ranges of surface area (angstrom-squared) of contact between proteins. (b) Median and quartile ranges of intermolecular residue -residue non-covalent contacts. (c) Intermolecular hydrogen bond count. (d) Fractional count of inter - residue contacts between participating proteins, per unit surface area of contact. 484 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint Figure 4 485 486 Figure 4: Secondary structure distribution at protein interaction interfaces. (a) Median and quartile ranges of secondary structure distributions over interface residues within the VASCO set. (b) Median and quartile ranges of secondary structure distributions over interface residues within control homodimer and (c) heterodimer interaction interfaces. 487 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint Figure 5 488 Figure 5: Theoretical binding energy comparison between viral Ag -Ab and general PPIs. The median and quartile ranges are shown. Total binding interaction energy, along with decomposition into electrostatics, van der Waals, and solvation (polar/non -polar) energies are represented in different panels. 489 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint Figure 6 490 Figure 6: Manifold reduction helps to quantify intrinsic differences between viral Ag -Ab and general PPIs. (a) Isomap and (b) Spectral Embedding of the combined set of VASCO, and general protein-protein interface structures. (c) Spectral embedding and (d) isomap projections of different viral types reveal that HIV Ag-Ab interfaces (cyan) cluster together to form a distinct subclass – shown by circles. (e) Negative log probability map of an interface, measuring its likelihood of being a viral Ag-Ab interface or, specifically, (f) a HIV interface. .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint Figure 7 491 492 Figure 7. Structural variations in viral Ab -Ag interfaces across spectral embedding extremes. (a) Triangular spectral embedding projection of the VASCO dataset, colored by a weighted histogram PMF representation. More red regions indicate higher population sites. The three extreme clusters (Set-T, Set-L, and Set-R) are marked and colored red, green an d yellow respectively. Representative structures were selected based on proximity to the cluster mean by RMSD. (b) Superposition of representative antibody structures from each cluster, illustrating differences in CDR loop conformations and interface orien tations. (c) Antigen binding modes across the three clusters, showing differences in surface of contact and chain preference (balanced in Set-T, L-chain biased in Set-L, and H-chain dominant in Set-R). 493 494 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint FUNDING: K.W., I.S., E.S. and S.C. were partially supported by NIH NIGMS grant 495 R35GM151231-01. and through NEU Faculty Startup Funds. ES was also supported by NEU 496 PEAK fellowship. This research used computational resources from Northeastern Discovery 497 cluster at MGHPCC. 498 499 AUTHOR CONTRIBUTIONS: S.C. conceptualized the study. K.W., I.S., E.S. and S.C designed 500 the experiments. K.W., I.S., E.S. and S.C performed the data collection, analysis, interpretation, 501 and figure preparation. K.W., I.S., E.S. and S.C wrote and reviewed the manuscript. S.C. is the 502 corresponding author. 503 504 COMPETING INERESTS: The authors declare no competing interests. 505 506 DATA AND MATERIALS AVAILABILITY: Data and materials are available from the 507 corresponding authors upon request. 508 509 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint

510 511 512 1. Kapingidza, A.B., K. Kowal, and M. Chruszcz, Antigen-Antibody Complexes. Subcell 513 Biochem, 2020. 94: p. 465-497. 514 2. Walsh, G. and E. Walsh, Biopharmaceutical benchmarks 2022. Nat Biotechnol, 2022. 515 40(12): p. 1722-1760. 516 3. Jain, D. and D.M. Salunke, Antibody specificity and promiscuity. Biochem J, 2019. 517 476(3): p. 433-447. 518 4. Stanfield, R.L. and I.A. Wilson, Antibody Structure. Microbiol Spectr, 2014. 2(2). 519 5. Goulet, D.R. and W.M. Atkins, Considerations for the Design of Antibody-Based 520 Therapeutics. J Pharm Sci, 2020. 109(1): p. 74-103. 521 6. Kaplon, H. and J.M. Reichert, Antibodies to watch in 2021. MAbs, 2021. 13(1): p. 522 1860476. 523 7. Day, M.J., Introduction to Antigen and Antibody Assays. Top Companion Anim Med, 524 2015. 30(4): p. 128-31. 525 8. Voskuil, J.L., The challenges with the validation of research antibodies. F1000Res, 2017. 526 6: p. 161. 527 9. Olsen, T.H., F. Boyles, and C.M. Deane, Observed Antibody Space: A diverse database 528 of cleaned, annotated, and translated unpaired and paired antibody sequences. Protein 529 Sci, 2022. 31(1): p. 141-146. 530 10. Chailyan, A., A. Tramontano, and P. Marcatili, A database of immunoglobulins with 531 integrated tools: DIGIT. Nucleic Acids Res, 2012. 40(Database issue): p. D1230-4. 532 11. Lefranc, M.P. and G. Lefranc, Immunoglobulins or Antibodies: IMGT((R)) Bridging 533 Genes, Structures and Functions. Biomedicines, 2020. 8(9). 534 12. Swindells, M.B., et al., abYsis: Integrated Antibody Sequence and Structure-535 Management, Analysis, and Prediction. J Mol Biol, 2017. 429(3): p. 356-364. 536 13. Bjorling, E. and M. Uhlen, Antibodypedia, a portal for sharing antibody and antigen 537 validation data. Mol Cell Proteomics, 2008. 7(10): p. 2028-37. 538 14. Bandrowski, A., et al., The Antibody Registry: ten years of registering antibodies. 539 Nucleic Acids Res, 2023. 51(D1): p. D358-D367. 540 15. Burley, S.K., et al., Protein Data Bank (PDB): The Single Global Macromolecular 541 Structure Archive. Methods Mol Biol, 2017. 1607: p. 627-641. 542 16. Dunbar, J., et al., SAbDab: the structural antibody database. Nucleic Acids Res, 2014. 543 42(Database issue): p. D1140-6. 544 17. Raybould, M.I.J., et al., Thera-SAbDab: the Therapeutic Structural Antibody Database. 545 Nucleic Acids Res, 2020. 48(D1): p. D383-D388. 546 18. Mendes, M., et al., IEDB-3D 2.0: Structural data analysis within the Immune Epitope 547 Database. Protein Sci, 2023. 32(4): p. e4605. 548 19. Pybus, O.G. and A. Rambaut, Evolutionary analysis of the dynamics of viral infectious 549 disease. Nat Rev Genet, 2009. 10(8): p. 540-50. 550 20. Mosa, A.I., Antigenic variability. Frontiers in Immunology, 2020. 11: p. 2057. 551 21. Ma, H. and R. O'Kennedy, The Structure of Natural and Recombinant Antibodies. 552

Mol Biol, 2015. 1348: p. 7-11. 553 22. Al-Lazikani, B., A.M. Lesk, and C. Chothia, Standard conformations for the canonical 554 structures of immunoglobulins. J Mol Biol, 1997. 273(4): p. 927-48. 555 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint 23. Weitzner, B.D., R.L. Dunbrack, Jr., and J.J. Gray, The origin of CDR H3 structural 556 diversity. Structure, 2015. 23(2): p. 302-11. 557 24. Ambrosetti, F., et al., Modeling Antibody-Antigen Complexes by Information-Driven 558 Docking. Structure, 2020. 28(1): p. 119-129 e2. 559 25. Bryant, P., G. Pozzati, and A. Elofsson, Improved prediction of protein-protein 560 interactions using AlphaFold2. Nat Commun, 2022. 13(1): p. 1265. 561 26. Allen, W.J., et al., DOCK 6: Impact of new features and current docking performance. J 562 Comput Chem, 2015. 36(15): p. 1132-56. 563 27. Vreven, T., et al., Updates to the Integrated Protein-Protein Interaction Benchmarks: 564 Docking Benchmark Version 5 and Affinity Benchmark Version 2. J Mol Biol, 2015. 565 427(19): p. 3031-41. 566 28. Corti, D. and A. Lanzavecchia, Broadly neutralizing antiviral antibodies. Annu Rev 567 Immunol, 2013. 31: p. 705-42. 568 29. Walker, L.M. and D.R. Burton, Passive immunotherapy of viral infections: 'super-569 antibodies' enter the fray. Nat Rev Immunol, 2018. 18(5): p. 297-308. 570 30. Koide, S. and S.S. Sidhu, The importance of being tyrosine: lessons in molecular 571 recognition from minimalist synthetic binding proteins. ACS Chem Biol, 2009. 4(5): p. 572 325-34. 573 31. Osajima, T., et al., Computational and statistical study on the molecular interaction 574 between antigen and antibody. J Mol Graph Model, 2014. 53: p. 128-139. 575 32. Birtalan, S., et al., The intrinsic contributions of tyrosine, serine, glycine and arginine to 576 the affinity and specificity of antibodies. J Mol Biol, 2008. 377(5): p. 1518-28. 577 33. Diskin, R., et al., Increasing the potency and breadth of an HIV antibody by using 578 structure-based rational design. Science, 2011. 334(6060): p. 1289-93. 579 34. Zhou, T., et al., Multidonor analysis reveals structural elements, genetic determinants, 580 and maturation pathway for HIV-1 neutralization by VRC01-class antibodies. Immunity, 581 2013. 39(2): p. 245-58. 582 35. Yan, C., et al., Characterization of protein-protein interfaces. Protein J, 2008. 27(1): p. 583 59-70. 584 36. David, C.C. and D.J. Jacobs, Principal component analysis: a method for determining the 585 essential dynamics of proteins. Methods Mol Biol, 2014. 1084: p. 193-226. 586 37. Zhou, H., F. Wang, and P. Tao, t-Distributed Stochastic Neighbor Embedding Method 587 with the Least Information Loss for Macromolecular Simulations. J Chem Theory 588 Comput, 2018. 14(11): p. 5499-5510. 589 38. Das, P., et al., Low-dimensional, free-energy landscapes of protein-folding reactions by 590 nonlinear dimensionality reduction. Proc Natl Acad Sci U S A, 2006. 103(26): p. 9885-591 90. 592 39. Aflalo, Y. and R. Kimmel, Spectral multidimensional scaling. Proc Natl Acad Sci U S A, 593 2013. 110(45): p. 18052-7. 594 40. Pittala, S. and C. Bailey-Kellogg, Learning context-aware structural representations to 595 predict antigen and antibody binding interfaces. Bioinformatics, 2020. 36(13): p. 3996-596 4003. 597 41. Chinery, L., et al., Paragraph-antibody paratope prediction using graph neural networks 598 with minimal feature vectors. Bioinformatics, 2023. 39(1). 599 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint 42. Wang, Z., Y. Wang, and W. Zhang, Improving paratope and epitope prediction by multi-600 modal contrastive learning and interaction informativeness estimation. arXiv preprint 601 arXiv:2405.20668, 2024. 602 43. Chaudhury, S., et al., Benchmarking and analysis of protein docking performance in 603 Rosetta v3.2. PLoS One, 2011. 6(8): p. e22477. 604 44. Chen, R., L. Li, and Z. Weng, ZDOCK: an initial-stage protein-docking algorithm. 605 Proteins, 2003. 52(1): p. 80-7. 606 45. Vreven, T., H. Hwang, and Z. Weng, Integrating atom-based and residue-based scoring 607 functions for protein-protein docking. Protein Sci, 2011. 20(9): p. 1576-86. 608 46. McCoy, K.M., M.E. Ackerman, and G. Grigoryan, A comparison of antibody-antigen 609 complex sequence-to-structure prediction methods and their systematic biases. Protein 610 Sci, 2024. 33(9): p. e5127. 611 612 .CC-BY-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted March 14, 2025. ; https://doi.org/10.1101/2025.03.11.642737doi: bioRxiv preprint