Benchmarking sequence-based and AlphaFold-based methods for pMHC-II binding core prediction: distinct strengths and consensus approaches Authors: Soobon Ko, Honglan Li, Hongeun Kim, Woong-Hee Shin, Junsu Ko, Yoonjoo Choi Published: 2026 ABSTRACT AbstractBackgroundInteractions between peptide and MHC class II (pMHC-II) are crucial for T-cell recognition and immune responses, as MHC-II molecules present peptide fragments to T cells, enabling the distinction between self and non-self antigens. Accurately predicting the pMHC-II binding core is particularly important because it provides insights into pMHC-II interactions and T-cell receptor engagement. Given the high polymorphism and peptide-binding promiscuity of MHC-II molecules, computational prediction methods are essential for understanding pMHC-II interactions. While sequence-based methods are widely used, recent advances in AlphaFold-based structure prediction have opened new possibilities for improving pMHC-II binding core predictions.MethodsWe constructed a non-redundant dataset of 72 pMHC-II complexes from the IMGT database, supplemented with shuffled negative peptides and curated non-binders. Two sequence-based methods (NetMHCIIpan-4.3 and DeepMHCII) and two structure-based methods (AlphaFold2 fine-tuned (AF-FT) and AlphaFold3 (AF3)) were benchmarked for binding and core prediction. Performance was evaluated using standard metrics (precision, recall, F1 score), and consensus strategies integrating sequence- and structure-based models were developed for scenarios with known and unknown binding status.ResultsThe AlphaFold-based methods showed strong performance in predicting positive binders, with AF3 achieving the highest positive recall (0.86) and AF2-FT performing similarly (0.81). However, both methods frequently misclassified unbound peptides as binders. NetMHCIIpan excelled at identifying non-binders, achieving the highest negative recall (0.93), but had lower positive recall (0.44). In contrast, DeepMHCII demonstrated moderate performance without any notable strength. Consensus approaches combining AlphaFold-based methods for binder identification with filtering using NetMHCIIpan improved overall prediction precision (0.94 and 0.87 for known and unknown binding status, respectively).ConclusionsThis study highlights the complementary strengths of AlphaFold-based and sequence-based methods for predicting pMHC-II binding core regions. AlphaFold-based methods excel in predicting positive binders, while NetMHCIIpan is highly effective at identifying non-binders. Future research should focus on improving the prediction of unbound peptides for AlphaFold-based models. Since NetMHCIIpan’s binding core predictive ability is already high, future efforts should concentrate on enhancing its binding prediction to further improve overall accuracy.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12920-026-02432-4.Keywords: pMHC-II, AlphaFold, Binding core, Consensus, Peptide binding Background Antigen presentation is a critical process in the immune response, where antigen-presenting cells, such as dendritic cells, macrophages, and B cells, display peptide fragments on their surface bound to major histocompatibility complex (MHC) molecules. This interaction is pivotal for the activation of T cells, which recognize the presented peptides via their T-cell receptors (TCRs). The process is essential for distinguishing between self and non-self antigens, enabling the immune system to mount responses against a wide variety of pathogenic invaders [1]. Therefore, an accurate understanding of peptide-MHC (pMHC) interactions is crucial, not only for elucidating the mechanisms of immune responses but also for the development of effective vaccines and immunotherapies [2–6]. MHC molecules are highly polymorphic [7, 8], possibly as an evolutionary strategy to protect against the virtually limitless number of potential pathogens [9, 10]. This extreme polymorphism allows MHC molecules to present a broad spectrum of antigenic peptides. In addition to the polymorphism, a single MHC molecule can bind to a variety of different peptides [11]. This broad peptide-binding capacity, or peptide-binding promiscuity, further enhances the immune system’s flexibility, allowing it to respond to many different pathogens despite the limited number of MHC alleles in any given individual. However, the polymorphism and peptide-binding promiscuity of MHC molecules make it practically impossible to explore the entire pMHC interaction space experimentally, presenting a significant challenge for immunological research [12, 13]. Given the impracticality of experimentally determining all pMHC interactions, computational prediction methods have emerged as essential tools for investigating these interactions. There are two main approaches for predicting peptide-MHC interactions: sequence-based and structure-based methods [14, 15]. Traditionally, sequence-based methods have been more widely used due to their speed and accuracy. These methods rely on large datasets of known pMHC binding affinities to train machine learning models [16–19] or apply position-specific scoring matrices [20–22]. Sequence-based methods are generally more effective for predicting pMHC-I interactions, where the peptide length is relatively fixed (typically 8–11 amino acids). This relatively fixed length allows for more straightforward modeling of the binding core region, leading to high prediction accuracy. However, predicting pMHC-II interactions is more challenging. Unlike MHC-I, MHC-II molecules bind longer peptides with additional flanking regions at each terminus, making it difficult to accurately identify the binding core region. As a result, sequence-based methods often struggle with the greater flexibility of pMHC-II complexes [23–25]. Structure-based approaches have gained increasing attention due to recent breakthroughs in artificial intelligence and its application to protein structure prediction. Tools such as AlphaFold [26] and RoseTTAFold [27] have revolutionized the field, enabling the reliable prediction of protein structures with remarkable accuracy. However, structure-based methods have been less effective in modeling immune molecules, primarily because co-evolutionary information is often unavailable for these interactions [28]. To address this limitation, the development of AlphaFold2 fine-tuned for peptide interactions has shown promising results [29]. Additionally, the most recent version, AlphaFold3 [30], has demonstrated enhanced prediction accuracy, delivering impressive performance in molecular modeling. In this study, we benchmarked two widely-used sequence-based methods, NetMHCIIpan [31] and DeepMHCII [32], and two structure-based methods, AlphaFold2 fine-tuned for peptide interactions [29] and AlphaFold3 [30], with a specific focus on identifying the binding core regions (Fig. 1A). While there have been numerous benchmark and performance comparison studies for pMHC-II prediction, to date, no studies have specifically focused on the accuracy of binding core region prediction. Examples of pMHC-II binding and prediction assessment. The MHC-II complex is shown in white, the non-core portion of the peptide is colored in blue, and the core region is in red. A The pMHC-II complex with the correctly predicted core region. B The peptide binds to MHC-II, but the predicted core region is incorrect. C Incorrect peptide binding conformation is regarded as unbound. D Complete mislocation of the peptide is also considered unbound To address this gap, we prepared an extensive test set and evaluated the computational prediction methods in various settings, providing a comprehensive comparison. Our results show that each method has distinct characteristics in prediction. We propose a consensus approach that combines the strengths of both sequence-based and structure-based methods to improve overall prediction accuracy and provide more reliable results for pMHC-II binding core region identification. Beyond methodological benchmarking, the accurate identification of the pMHC-II binding core serves as a key enabler for broader applications in immunological modeling and design. Reliable core delineation defines the molecular interface recognized by the T-cell receptor and underpins numerous downstream applications, including the precise mapping of T-cell epitopes for vaccine development, the structural modelling of TCR-pMHC recognition in immunotherapy research, and the elucidation of antigen presentation mechanisms relevant to autoimmune and infectious diseases. By establishing a unified benchmarking framework, this study provides not only an evaluation of current computational methods but also a foundation for advancing structure-based immunological modelling and translational research. Methods Dataset preparation A total of 126 peptide-MHC-II (pMHC-II) complex structures, including those with T-cell receptors, were collected from the IMGT database [33]. Any structures containing non-standard amino acids within the pMHC-II were excluded. For the MHC-II structure, only the groove region, as defined by the IMGT, was used for structural modeling [34]. The MHC types and sequences were identified based on IMGT annotations [35]. From this initial set, a non-redundant set of 72 unique sequences was prepared for benchmarking (Additional File 1). The final non-redundant dataset comprised 72 pMHC-II complexes: 51 from human, 9 from mouse, and 12 from chimeric MHC molecules. Unbound (negative) peptides were generated by randomly shuffling the positive peptides while preserving both the peptide length and the overall amino acid composition of the original sequences. For example, if the original positive binder sequence was “AAAAADDDD,” a corresponding shuffled negative peptide could be “ADADDDAAA.” This procedure ensures that the negative peptides are compositionally realistic yet unlikely to retain known binding motifs. In addition, we created another set of experimentally validated negative peptides from the IEDB database [36]. The dataset was constructed through several filtration steps to ensure high confidence. First, peptides were required to share the same MHC-II allele and peptide length as their corresponding binders. Second, only peptides with binding affinities experimentally determined by quantitative assays were considered. Among these, we selected those reported as non-binders or exhibiting no or very weak binding in qualitative binding or dissociation constant (KD/IC₅₀) assays. Finally, only peptides with no detectable binding or extremely weak affinity (KD ≥ 50 μM) were retained. The resulting dataset represents a high-confidence negative set composed exclusively of experimentally validated non-binders. For each positive peptide, a negative peptide of the same length bound to the same MHC type was randomly selected. In this case, chimeric MHC molecules were not included in the validation. Among the 72 benchmark complexes, 28 lacked corresponding IEDB entries (12 chimeric and 16 human or mouse alleles); thus, 44 complexes were evaluated in the IEDB-based non-binder analysis. Computational prediction methods and assessment of binding To benchmark predictions of the binding core region in pMHC-II complexes, we employed two sequence-based methods (NetMHCIIpan-4.3 [31] and DeepMHCII [32]) alongside two structure-based methods, AlphaFold2 finetuned for peptide binding (AF2-FT) [29] and AlphaFold3 (AF3) [30]. The default settings for each prediction tool were used, and binding predictions were assessed based on specific criteria for each method. For NetMHCIIpan-4.3, peptides with a predicted binding rank ≤ 5% were considered binders, corresponding to the commonly used “weak binder” threshold recommended in the official documentation [31]. DeepMHCII classified peptides as binders if their predicted IC50 binding values were below 500 nM, corresponding to the default cutoff reported in the original publication [32]. For the two AlphaFold-based prediction methods, top-ranked predicted structures were used to estimate pMHC-II binding. All AF3 predictions were performed using the webserver (https://alphafoldserver.com/). For AF2-FT, the query peptide sequence consisted of 11 residues, with a core nine-residue segment and one residue at each terminus. For peptides longer than 11 residues, overlapping windows were generated by sliding the window along the peptide sequence. The top four templates were selected based on MHC alignment scores from a template search and used to construct structural features for AlphaFold modeling. For the selection of final structures, the predicted aligned error (PAE) scores were used, with the window slide showing the lowest PAE being chosen for further analysis of binding and core prediction. In the structure-based methods, binding was assessed manually by inspecting the orientation of peptide side chains relative to the target crystal structure. A peptide was classified as binding only when all nine residues of the predicted core were positioned inside the groove, with side-chain orientations compatible with the reference crystal structure (e.g., Fig. 1C and D). Consensus prediction strategy To integrate the complementary strengths of sequence-based and structure-based methods, we implemented a consensus prediction strategy combining NetMHCIIpan-4.3, AlphaFold2 fine-tuned for peptide interactions (AF2-FT), and AlphaFold3 (AF3). Two scenarios were considered: 1) when the peptide’s binding status was experimentally known and 2) when it was unknown. For cases with known binding status, the consensus approach aimed to maximize the accuracy of core region identification among confirmed binders. Core predictions from AF2-FT and AF3 were compared, and a consensus core was defined only when both models predicted overlapping core registers within the predefined nine-residue binding groove. This intersection-based strategy ensured that the final consensus represented structural agreement between the two models, thereby minimizing potential model-specific biases. For cases with unknown binding status, the consensus approach was implemented in two stages. In the first stage, all peptides were screened using NetMHCIIpan-4.3, which served as an initial filter leveraging its high specificity for non-binders. Peptides with a predicted rank score greater than 5.0 were excluded from further evaluation. In the second stage, the remaining candidate peptides were analyzed using AF2-FT and AF3, and a peptide was classified as a positive binder only if both structure-based methods predicted a binding conformation consistent with canonical MHC-II pocket occupancy. Peptides meeting this dual criterion formed the final consensus set of predicted binders. Results Prediction performance of pMHC-II binding We first benchmarked the predictive abilities of the four computational prediction methods, AlphaFold3 (AF3) and AlphaFold2 fine-tuned for peptide interactions (AF2-FT), alongside two sequence-based methods, NetMHCIIpan and DeepMHCII, for peptide binding (without considering true binding cores). The computational methods were tested on a non-redundant set of 72 pMHC-II complexes with determined crystal structures. The sequence-based methods completed each prediction within a few seconds on a standard CPU. In contrast, AF2-FT required less than one minute per complex when executed on an NVIDIA A100 GPU, while AF3 generated each structure within one to two minutes using the official web server. Although structure-based predictions are inherently more computationally demanding than sequence-based ones, these requirements remain well within the capabilities of modern GPU workstations and web-based implementations. We opted for random shuffling rather than using experimentally validated non-binders because, though they are known non-binders, such datasets may have already been incorporated into the training data of one or more of the benchmarked methods, potentially introducing bias. We thus compared performance trends using an additional set of experimentally validated non-binders from the IEDB database. The results showed that the overall trends remain largely consistent while the performance of AF2-FT marginally increased (Table 1). For the rest of the benchmarking analysis, we used the random shuffle negative peptides to ensure that they were unseen by all models and that the evaluation remained unbiased. Table 1pMHC-II binding and core region prediction results. Values in brackets represent binding prediction results only, while values outside parentheses correspond to core-binding prediction. The discrepancies between peptide binding prediction and core prediction are more pronounced in sequence-based methods compared to structure-based methods. Notably, AF2-FT shows no differences between binding and core prediction. Chimeric MHC-II alleles were excluded from the IEDB-based non-binder evaluation due to the absence of corresponding experimental dataMetricAF3AF2-FTNetMHCIIpanDeepMHCIISpeciesPositive Prediction (Binder Prediction)Precision0.643(0.648)0.667(0.667)0.857(0.865)0.6(0.642)Homo sapiensSensitivity0.882(0.902)0.824(0.824)0.588(0.627)0.706(0.843)F10.744(0.754)0.737(0.737)0.698(0.727)0.649(0.729)Precision0.727(0.727)0.636(0.636)0.0(0.0)0.4(0.5)Mus musculusSensitivity0.889(0.889)0.778(0.778)0.0(0.0)0.222(0.333)F10.8(0.8)0.7(0.7)0.0(0.0)0.286(0.4)Precision0.692(0.692)0.818(0.818)0.0(0.0)0.273(0.429)ChimericSensitivity0.75(0.75)0.75(0.75)0.0(0.0)0.25(0.5)F10.72(0.72)0.783(0.783)0.0(0.0)0.261(0.462)Negative Prediction (Nonbinder-shuffle Prediction)Precision0.812(0.839)0.769(0.769)0.687(0.708)0.643(0.771)Homo sapiensSpecificity0.51(0.51)0.588(0.588)0.902(0.902)0.529(0.529)F10.627(0.634)0.667(0.667)0.78(0.793)0.581(0.628)Precision0.857(0.857)0.714(0.714)0.5(0.5)0.462(0.5)Mus musculusSpecificity0.667(0.667)0.556(0.556)1.0(1.0)0.667(0.667)F10.75(0.75)0.625(0.625)0.667(0.667)0.545(0.571)Precision0.727(0.727)0.769(0.769)0.5(0.5)0.308(0.4)ChimericSpecificity0.667(0.667)0.833(0.833)1.0(1.0)0.333(0.333)F10.696(0.696)0.8(0.8)0.667(0.667)0.32(0.364)Negative Prediction (Nonbinder-IEDB Prediction)Precision0.739(0.773)0.769(0.769)0.611(0.635)0.531(0.68)Homo sapiensSpecificity0.459(0.459)0.811(0.811)0.892(0.892)0.459(0.459)F10.567(0.576)0.789(0.789)0.725(0.742)0.493(0.548)Precision0.833(0.833)0.75(0.75)0.438(0.438)0.462(0.5)Mus musculusSpecificity0.714(0.714)0.857(0.857)1.0(1.0)0.857(0.857)F10.769(0.769)0.8(0.8)0.609(0.609)0.6(0.632)OverallAccuracy0.696(0.706)0.706(0.706)0.745(0.765)0.618(0.686)Homo sapiensBalanced Accuracy0.696(0.706)0.706(0.706)0.745(0.765)0.618(0.686)MCC0.423(0.448)0.424(0.424)0.516(0.551)0.239(0.392)Accuracy0.778(0.778)0.667(0.667)0.5(0.5)0.444(0.5)Mus musculusBalanced Accuracy0.778(0.778)0.667(0.667)0.5(0.5)0.444(0.5)MCC0.57(0.57)0.342(0.342)0.0(0.0)−0.124(0.0)Accuracy0.708(0.708)0.792(0.792)0.5(0.5)0.292(0.417)ChimericBalanced Accuracy0.708(0.708)0.792(0.792)0.5(0.5)0.292(0.417MCC0.418(0.418)0.585(0.585)0.0(0.0)−0.418(−0.169)The highest values are in bold pMHC-II binding and core region prediction results. Values in brackets represent binding prediction results only, while values outside parentheses correspond to core-binding prediction. The discrepancies between peptide binding prediction and core prediction are more pronounced in sequence-based methods compared to structure-based methods. Notably, AF2-FT shows no differences between binding and core prediction. Chimeric MHC-II alleles were excluded from the IEDB-based non-binder evaluation due to the absence of corresponding experimental data The highest values are in bold The structure-based methods excelled in predicting positive binders (Fig. 2 and Table 1). AF3 achieved the highest sensitivity (0.88), reflecting its ability to correctly identify the majority of true binders. AF2-FT performed similarly well, with both methods producing comparable positive F1 scores (AF3: 0.75, AF2-FT: 0.73), demonstrating strong performance in capturing pMHC-II binding events. However, both models exhibited a tendency to misclassify unbound peptides as binders, as indicated by their relatively low specificity values (AF3: 0.56, AF2-FT: 0.63). This behavior likely stems from the nature of AlphaFold-based structure-based modeling, where the method tends to put peptides into the MHC groove, even when a perfect binding fit is not available [37]. Performance comparison of pMHC-II core prediction methods A Recall, B Precision, C: F1, and D MCC. The colors represent the different prediction methods: blue for AF3, sky blue for AF2-FT, red for NetMHCIIpan, and orange for DeepMHCII. The dotted bars indicate the performance for binding prediction. When AF2-FT correctly predicts peptide binding, the binding core is always accurately identified, while AF3 misses only one case. In contrast, the sequence-based methods often incorrectly predict core regions even when pMHC-II binding is correctly predicted In contrast, NetMHCIIpan demonstrated superior performance in identifying non-binders. It achieved the highest specificity (0.93) and a strong negative F1 score (0.75), indicating its strength in predicting non-binding peptides. Additionally, NetMHCIIpan's high positive precision (0.86) suggests that when it predicts a peptide as a binder, the prediction is generally accurate. However, its low sensitivity (0.44) indicates that it may overlook some true binders, likely due to its strict criteria for determining binding. DeepMHCII, while performing moderately across most metrics, was the least accurate method overall, with an accuracy of 0.62 and an MCC of 0.24. Other metrics (AUC-ROC and PR-AUC) also indicate that NetMHCIIpan is generally more accurate than DeepMHCII (Fig. S1). Its balanced performance, though lower compared to the other methods, suggests that it lacks the distinctive strengths of the other models, particularly in accurately predicting either binders or non-binders. The results indicate that the structure-based methods are particularly good at identifying positive binders but tend to overestimate binding by misclassifying non-binders. On the other hand, NetMHCIIpan excels at correctly identifying non-binders while maintaining high precision for the binders. Combining these two approaches, leveraging the structure-based methods' high sensitivity values and NetMHCIIpan's high precision, could provide a synergistic solution, improving overall prediction accuracy and reliability. Prediction performance of pMHC-II binding core We next benchmarked the prediction methods for the identification of the binding core region in pMHC-II complexes. As with the binding prediction assessment, AF2-FT and AF3 demonstrated strong performance in binding core prediction (Fig. 2 and Table 1). AF3 achieved a sensitivity of 0.86 (62/72), slightly better than AF2-FT (0.81, 58/72) in overall core identification. Both models showed comparable positive F1 scores (AF3: 0.75 and AF2-FT: 0.74), indicating their similar predictive abilities in capturing binding core events. Notably, for AF2-FT, whenever it predicts a peptide as a binder, the binding core is always correctly identified. A similar trend is observed for AF3, with only one case where the binding core is incorrectly predicted. In the case of unbound peptides, AF2-FT performed slightly better than AF3 (specificity 0.69 and 0.66, respectively). The two sequence-based methods showed relatively lower performance in predicting pMHC-II binding core, with recall values roughly half of those seen with the structure-based methods. While DeepMHCII generally underperformed in both binder and non-binder prediction, as reflected in its MCC value (0.08), NetMHCIIpan was notably effective at capturing non-binders, achieving a specificity of 0.93. Similar to the binding prediction results, the structure-based methods (AF3 and AF2-FT) and NetMHCIIpan demonstrated distinct strengths in different aspects of binding core prediction. The structure-based methods were highly capable of identifying positive binders, but they frequently misclassified non-binding cores as binding. In contrast, NetMHCIIpan's ability to accurately identify non-binders makes it a reliable method for negative prediction. Given the complementary strengths of the structure-based methods and NetMHCIIpan, a consensus approach combining the high sensitivity of AF3 and AF2-FT with NetMHCIIpan's superior specificity and positive precision could provide a more accurate and reliable method for predicting the pMHC-II binding core region. To investigate the failure cases, we further analyzed the prediction results. The first indicator is the query peptide length. The average peptide length of our test set was approximately 17 residues. For AF2-FT and AF3, the average length of failure cases was around 21 residues (23 and 19 residues for NetMHCIIpan and DeepMHCII, respectively). These results suggest that longer peptides are more challenging to predict accurately. Figure 3A illustrates the relationship between query peptide length and misprediction rate. Failure analysis of pMHC-II binding core prediction by peptide length and MHC species. A Relationship between query peptide length and misprediction rate for pMHC-II binding core prediction. The average peptide length of the test set was approximately 17 residues, and longer peptides posed a greater challenge for accurate prediction across all methods. B Prediction accuracy for different MHC species. Prediction accuracy was significantly lower for peptides bound to chimeric MHCs, particularly for sequence-based methods Using 17 residues as a cut-off value for distinguishing long and short peptides, the misprediction rate of AF3 was 9% for short peptides, but increased to 25% for longer peptides. Similarly, AF2-FT recorded 13% for short peptides, which rose to 35% for longer ones. This increase in misprediction rates was even more pronounced for the sequence-based methods. NetMHCIIpan showed a misprediction rate of 7% for short peptides, but the rate increased significantly to 50% for longer ones (For DeepMHCII, 17% and 40% for short and long peptides respectively). The species of the MHC molecule also plays a significant role in prediction accuracy. Specifically, the prediction accuracy was noticeably lower for peptides bound to chimeric MHCs (Fig. 3B). This trend was particularly pronounced for the sequence-based methods, likely due to their inability to effectively handle non-standard MHCs. In contrast, the structure-based methods were relatively less affected by the presence of chimeric MHCs, possibly because they were either trained on a broader dataset or are less data-dependent. Consensus approach for pMHC-II binding core prediction Considering the superior recall of the structure-based methods, we explored two possible scenarios for pMHC-II binding core prediction. In the first scenario, assuming the binding status is experimentally known, we evaluated the performance of each method in predicting the correct core region (Fig. 4A and Table 2). The structure-based methods showed superior overall performance. Among the 72 binding cases, AF3 correctly predicted the core in 62 cases (86.1%), AF2-FT in 58 (80.6%), NetMHCIIpan in 30 (41.7%), and DeepMHCII in 41 (56.9%). Interestingly, there were several instances where AF3 and AF2-FT did not agree on the predicted core region. By combining the predictions of AF3 and AF2-FT, we obtained the best results, with 68 out of 72 (94.4%) cores correctly predicted. Performance of pMHC-II binding core prediction using consensus approaches. A Binding core prediction performance for each method when the binding status is experimentally known. The structure-based methods (AF3 and AF2-FT) demonstrated superior accuracy, with a union consensus approach achieving the highest correct core prediction rate of 94.4%. B Binding core prediction performance when the binding status is completely unknown. The combination of AF3 and AF2-FT with NetMHCIIpan filtering further improved precision to 86.5%. The consensus approach demonstrated perfect recall (1.00) and the highest F1 score (0.93) Table 2Performance metrics of pMHC-II binding core prediction with known vs. unknown binding status.AF3AF2-FTNetMHCIIpanDeepMHCIIConsensusBinding status known Precision0.8610.8060.4170.5690.944Binding status unknown Recall0.9060.9690.9380.9061.000 Precision0.9060.8610.8570.9060.865 F10.9060.9120.8960.9060.928The highest values are in bold Performance metrics of pMHC-II binding core prediction with known vs. unknown binding status. The highest values are in bold We further investigated the four failure cases. Notably, two core peptides (WGAEGQRPG for 3T0E and 3O6F, and RLYLVCGEE for 6DFS and 6DFW) were consistently incorrectly predicted. In the former case, the actual query peptide was FSWGAEGQRPGFGSGGGSLVPRGSGGGG, and the sequence-based methods agreed on the core region being FGSGGGSLV. This suggests that the sequence-based methods identified a 9mer substring whose termini are hydrophobic, possibly favoring the selection of this region. In the latter case (6DFS and 6DFW), the MHC-II molecule was a chimera between human and mouse, which may explain the failure, as chimeric MHC molecules were likely not included in the training sets of the prediction methods. We also tested the scenario where the binding status was completely unknown. We first applied a union combination of AF3 and AF2-FT, given their promising performance in the previous setting. This combined approach correctly identified 94.4% of the binders (Fig. 4A) but also misclassified 31 out of 72 unbound peptides as binders. Considering that NetMHCIIpan had the highest recall for unbound peptides, we applied a filtering step using NetMHCIIpan to exclude negatively predicted peptides. After filtering, 37 peptides remained, of which 32 were correctly identified, resulting in a precision of approximately 86.5%. The consensus approach effectively improved overall performance metrics, as shown in Fig. 4B and Table 2. The consensus method achieved perfect recall (1.00) for the remaining binders, demonstrating its capability to avoid missing any true positive binding cores among the filtered peptides. The F1 score for the consensus approach was also the highest (0.93), highlighting the benefits of integrating complementary strengths to enhance predictive accuracy and reliability. Discussion This study highlights the strengths and limitations of cutting-edge sequence-based and structure-based methods for pMHC-II binding core prediction. While structure-based methods demonstrate high effectiveness in predicting positive binders with correct core identification, they tend to struggle with accurately identifying non-binders. Specifically, these methods often "force" peptides into the MHC groove, leading to an overestimation of binding events. In contrast, sequence-based methods like NetMHCIIpan excel at filtering non-binders but frequently miss true binders. DeepMHCII showed moderate performance across the board, lacking the pronounced strengths seen in the other methods. The AlphaFold-based structure-based methods tested here were clearly superior in predicting pMHC-II binding cores. AF3, with the highest sensitivity and F1 scores, consistently identified most true binders and correct core regions, while AF2-FT performed similarly well. Interestingly, AF3 achieved this high accuracy without any fine-tuning for peptide-MHC interactions, demonstrating the powerful predictive potential of the advanced prediction method. However, their relatively low specificity suggests that they often misclassify unbound peptides as binders, likely due to the inherent nature of AlphaFold-based modeling that fits peptides into the MHC groove even when the fit is imperfect. To further elucidate the causes of core misclassification, we analyzed binder cases that were consistently mispredicted by all evaluated methods. Although these peptides were experimentally confirmed as binders, both sequence- and structure-based approaches failed to identify their correct core positions (Fig. S2). For sequence-based methods, errors were primarily observed in long peptides or chimeric MHC-II alleles, reflecting their limited generalization to non-human or atypical allele contexts. In contrast, AF2-FT and AF3 frequently exhibited recurrent structural failure modes, not only in these representative examples in Fig. S2 but also across other mispredicted complexes. Specifically, Gly/Ser-rich flanking segments (e.g., GA, GGG, GSGGS, SGS) tended to form flexible loop-like extensions, while short motifs such as AEG or EGA induced helix-like turns that distorted the peptide orientation within the MHC groove. Moreover, several mispredictions showed instability at the P9 anchor pocket, where small residues such as Gly or Ala failed to fully occupy the hydrophobic cavity, resulting in partial detachment or displacement of the bound core. Collectively, these structural tendencies such as loop formation, secondary-structure induction, and weak terminal anchoring, were particularly pronounced in long peptides and chimeric MHC-II alleles, contributing to the systematic overestimation of binding propensity observed in structure-based models. NetMHCIIpan, on the other hand, performed exceptionally well in predicting the binding core region if peptide binding is not considered, with an accuracy comparable to that of AF2-FT (58 out of 72 correct predictions). This suggests that NetMHCIIpan already possesses a strong capability for identifying the correct core region in pMHC-II complexes. However, while its core-binding predictions are generally highly accurate, the method falls short when it comes to predicting whether a peptide will bind at all. This gap shows the need for improvement in its binding prediction capabilities, even though its core identification abilities are already quite advanced. It is also important to note that all the methods examined in this study are relatively recent, and it remains unclear how much information was already included in their training sets. This uncertainty may impact the evaluation of their performance, especially if some peptides or MHC molecules in our test set were part of their training data, potentially inflating the predictive accuracy. Therefore, future benchmarking studies should consider the training set overlap to ensure a more unbiased assessment. The consensus approach proposed in this study, which combines the strengths of both structure- and sequence-based methods, provided a synergistic solution to improve overall prediction accuracy. By leveraging the high sensitivity of AF3 and AF2-FT for binder prediction and the superior ability of NetMHCIIpan to filter non-binders, we were able to achieve more reliable predictions of pMHC-II binding core regions. This integration of methodologies emphasizes the importance of using a combined approach to balance the strengths of each method. A key challenge for predictive models lies in their ability to generalize to unseen MHC-II alleles. Unlike conventional binding affinity benchmarks, prediction of the binding core region requires experimentally determined pMHC-II complex structures, which remain extremely limited in number. As a result, constructing an independent and statistically meaningful external test set is currently impractical. To approximate model generalizability within these constraints, we employed chimeric pMHC-II complexes, hybrid molecules incorporating domains from different species, as a surrogate test case. These chimeric structures represent biologically relevant but inherently difficult prediction scenarios, as their allelic configurations differ from those typically used in model training. Our analysis revealed that sequence-based models, particularly NetMHCIIpan and DeepMHCII, performed poorly on these chimeric complexes. It reflects their reliance on allele-specific motifs and their limited capacity to extrapolate to novel allelic contexts. In contrast, the AlphaFold-based methods maintained relatively stable performance. Although large-scale external validation is presently constrained by data availability, the inclusion of chimeric MHCs provides a realistic and biologically meaningful proxy for assessing model generalizability. Looking ahead, future research should focus on addressing the weaknesses of structure-based methods, particularly their tendency to overestimate binding by misclassifying unbound peptides. Because accurate binding classification is a necessary step for identifying the correct binding core, improving the ability of structure-based models to filter out non-binders will be crucial for improving overall accuracy. Meanwhile, sequence-based methods like NetMHCIIpan demonstrate strong performance in distinguishing binders from non-binders but are relatively less accurate in pinpointing the precise binding core. Therefore, balancing the strengths of both approaches, combining the robust binding discrimination of sequence-based models with the structural precision of AlphaFold-based predictions, will be essential for advancing binding core prediction in pMHC-II complexes. Achieving this balance between precision and recall in core identification represents a key direction for future model development. Conclusion This study demonstrates that sequence-based and AlphaFold-based methods provide complementary strengths for pMHC-II binding core prediction. Structure-based approaches excel in identifying binding cores among true binders, whereas NetMHCIIpan effectively filters non-binders. By integrating these methods, the proposed consensus strategy substantially improves overall prediction accuracy. Future work should focus on enhancing non-binder discrimination in structure-based models and improving binding prediction in sequence-based approaches. Supplementary Information Additional file 1: Dataset S1. Non-redundant pMHC-II complex dataset used in this study. Additional file 2: Figure S1. Precision–Recall (PR) and ROC curves for sequence-based prediction models. Figure S2. Representative mispredicted pMHC-II complexes illustrating structural causes of core misclassification across all prediction methods. Acknowledgements The authors acknowledge Arontier Co. for providing supercomputing resources, made available through funding secured by J.K. About this supplement This article has been published as part of BMC Medical Genomics Volume 19 Supplement 1, 2026: Proceedings of the 18th International Workshop on Data and Text Mining in Biomedical Informatics. The full contents of the supplement are available online at https://link.springer.com/journal/12920/volumes-and-issues/19-1/supplement. Abbreviations AlphaFold2 fine-tuned for peptide binding AlphaFold3 Half maximal inhibitory concentration Immune Epitope Database International ImMunoGeneTics information system Dissociation constant Major histocompatibility complex Major histocompatibility complex class II Predicted aligned error Peptide–major histocompatibility complex class II Precision–recall Receiver operating characteristic T-cell receptor Authors’ contributions The study was conceived and designed by SK and YC. SK led the data curation and analysis, wrote the manuscript, and contributed to funding acquisition. YC supervised the overall study, contributed to data curation and study design, and secured funding. WS contributed to funding acquisition and manuscript writing. HK and HL performed data analysis. JK contributed to funding acquisition. All authors reviewed and approved the final manuscript. Funding This work was supported in part by the Ministry of Health & Welfare, Republic of Korea (RS-2025-25459531) and by the Ministry of Science and ICT (MSIT), Republic of Korea, through the National Research Foundation of Korea (NRF) grants (RS-2025-02213506, RS-2026-25505636 and RS-2026-25505288); by the Korea Health Industry Development Institute (KHIDI) (NRF-2024M3A9J4006525); by the Institute for Information & Communications Technology Planning & Evaluation (IITP), funded by MSIT (IITP-2024-RS-2022–00156439 and IITP-2024-RS-2024–00438263); and by the Ministry of Health and Welfare, Republic of Korea (RS-2025–25459531). Data availability The datasets supporting the conclusions of this article are included within the article and its additional files. Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Additional file 1: Dataset S1. Non-redundant pMHC-II complex dataset used in this study. Additional file 2: Figure S1. Precision–Recall (PR) and ROC curves for sequence-based prediction models. Figure S2. Representative mispredicted pMHC-II complexes illustrating structural causes of core misclassification across all prediction methods. Data Availability Statement The datasets supporting the conclusions of this article are included within the article and its additional files.