Paper deep dive
TransNRank: Towards Accurate Neoantigen Ranking with Transformer
Zhiyin An, Yuenan Hou, Shumeng Duan, Yiming Zhou, Yuanting Zheng, Leming Shi
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Personalized neoantigen prediction is challenging due to the scarcity of positive samples, the noise of the experimental data, the severe class imbalance trait and the complex of immunogenicity features. Prior arts, such as linear regression and XGBoost fail to model long-range dependencies and contextual relationships within peptide features, therefore the performance of neoantigen positive recall rate is limited. In this paper, we present a novel deep learning framework based on Transformer, coined as TransNRank. By leveraging the self-attention mechanism, our model captures both local and global feature contexts, enabling more accurate recognition of immunogenic neoantigens. A positive-aware training objective is utilized to handle the class imbalance problem, assigning more weights to those few positive samples. Extensive experiments are performed on NCI, TESLA and HiTIDE datasets. Notably, our TransNRank can push the upper bound top 20 recall rate of neoantigen prediction from 46.9% (45 from 96) to 53.1% (51 from 96), while reducing the training epochs from 200 epochs to 20 epochs. Furthermore, we analyze the features contribution based on TransNRank and find that the mutation at anchor and TCGA expression level play an unexpected important role in neoantigen prediction, and removing insignificant features to reduce the input dimensionality of peptides does not drastically impair the overall performance of the model. Our paradigm not only streamlines the prediction pipeline but also sets a new state-of-the-art for neoantigen discovery, with broad implications for accurate immuno-oncology.
Tags
Links
- Source: https://arxiv.org/abs/2608.01924v1
- Canonical: https://arxiv.org/abs/2608.01924v1
Trouble viewing inline? Open PDF directly →
Full Text
44,772 characters extracted from source content.
Expand or collapse full text
TransNRank: Towards Accurate Neoantigen Ranking with Transformer Zhiyin An, Yuenan Hou, Shumeng Duan, Yiming Zhou, Yuanting Zheng*, Leming Shi* Abstract Personalized neoantigen prediction is challenging due to the scarcity of positive samples, the noise of the experimental data, the severe class imbalance trait and the complex of immunogenicity features. Prior arts, such as linear regression and XGBoost fail to model long-range dependencies and contextual relationships within peptide features, therefore the performance of neoantigen positive recall rate is limited. In this paper, we present a novel deep learning framework based on Transformer, coined as TransNRank. By leveraging the self-attention mechanism, our model captures both local and global feature contexts, enabling more accurate recognition of immunogenic neoantigens. A positive-aware training objective is utilized to handle the class imbalance problem, assigning more weights to those few positive samples. Extensive experiments are performed on NCI, TESLA and HiTIDE datasets. Notably, our TransNRank can push the upper bound top 20 recall rate of neoantigen prediction from 46.9% (45 from 96) to 53.1% (51 from 96), while reducing the training epochs from 200 epochs to 20 epochs. Furthermore, we analyze the features contribution based on TransNRank and find that the mutation at anchor and TCGA expression level play an unexpected important role in neoantigen prediction, and removing insignificant features to reduce the input dimensionality of peptides does not drastically impair the overall performance of the model. Our paradigm not only streamlines the prediction pipeline but also sets a new state-of-the- art for neoantigen discovery, with broad implications for accurate immuno-oncology. 1 Introduction Neoantigens derived from tumor somatic mutations are presented to T cells via human leukocyte antigen (HLA) molecules, and these neoantigens are recognized as non-self antigens by T cells, which further activate tumor-infiltrating T cells and eliminate tumor cells through endogenous immune responses (1, 2). Therefore, neoantigens are regarded as ideal targets for tumor immunotherapy (3). Nevertheless, due to the high heterogeneity of individual HLA genotypes, tumor somatic mutation profiles and tumor immune microenvironments, it is difficult to identify universal shared tumor neoantigens applicable to all cancer patients (4, 5). Personalized tumor neoantigens have ushered in a new era of individualized medication in precision oncology (6-11). Based on the principles of tumor neoantigen generation and function, personalized tumor neoantigen prediction primarily involves two main steps: variant peptide detection and immunogenicity prediction of variant peptides (6, 12, 13). The process of variant peptide detection is relatively mature and standardized, while the immunogenicity prediction is still challenging in this field (14- 16). Although numerous somatic mutations occur in individual tumor patients, only a limited number of mutated loci can generate immunogenic peptides. Early studies predicted tumor neoantigens merely based on the ranking of a single feature, namely MHC- peptide binding affinity (17, 18). Subsequent studies have confirmed that integrating multiple features can dramatically improve the accuracy of neoantigen prediction (19). This is mainly attributed to the fact that multiple factors jointly determine the immunogenicity of neoantigens, including their binding affinity to HLA, peptide expression level, TCR recognition capability, and others (20, 21). For each influencing factor, such as HLA-peptide binding affinity, feature values can be obtained using various software or prediction tools (22-25). Therefore, theoretically, a single peptide can have hundreds of immunogenicity-related features. Previous studies have found that single- feature prediction of tumor neoantigens is less accurate than multi-feature prediction (19). Multi-feature prediction replaced single-feature methods and become the mainstream approach for neoantigen immunogenicity prediction (26, 27). The multi-feature integration and prioritization scoring of candidate personalized neoantigens still largely rely on conventional machine learning algorithms, such as random forest, logistic regression (LR) and XGBoost (26, 28). To our knowledge, few efforts have been made to construct scoring frameworks for candidate neoantigens directly based on neural networks. The self-attention mechanism of the Transformer has given it unique advantages in in predicting HLA-peptide binding affinity (HLApollo) (29, 30). However, its potential in predicting personalized tumor neoantigen immunogenicity based on multiple feature prediction values has not yet been explored. In this study, we adopted the Transformer architecture for neoantigen immunogenicity prediction for the first time (30). Our results demonstrate that Transformer-based models, termed as TransNRank, achieve a higher predictive ceiling than conventional algorithms for neoantigen discovery. Using our developed Transformer framework, we performed feature ablation experiments. We also prove that the majority of input features are dispensable. Notably, mutations at anchor positions and TCGA-derived expression levels, rather than expression values from patient tumor samples, exert unexpectedly prominent effects on neoantigen prediction. 2 Results 2.1 Transformer improves personalized neoantigen prediction In this study, we used a standardized and publicly available dataset called NeoRanking. This harmonized dataset comes from three independent datasets, NCI (19), TESLA (31) and HiTIDE (26). Each mutant peptide with 31 immunogenic features and true label derived from the tumor patients are organized by Markus et al for benchmarking (26). Briefly, the NCI data were split into training and testing sets, while the TESLA and HiTIDE datasets were used exclusively as test data. The training set comprised 57 patients with 82 immunogenic peptides, and the test set comprised 42 patients with 96 immunogenic peptides (HCI-test: 23 patients, 21 immunogenic peptides; TESLA: 8 patients, 34 immunogenic peptides; HiTIDE: 11 patients, 41 immunogenic peptides) (Figure 1A, B). The dataset is highly imbalanced, with non-immunogenic peptides vastly outnumbering immunogenic ones (Figure 1C). We proposed that the Transformer, a self-attention-based neural network, has the potential to outperform traditional models reported in previous studies (30). Compared with conventional approaches, the Transformer contains a larger number of learnable hyper-parameters, most of which are continuous-valued (32). We therefore first explored the optimal settings for discrete hyper-parameters to identify the best training configuration. Given the limited size of the training set, we found that models with fewer layers achieved superior predictive performance. We next investigated the impact of continuous-valued hyper-parameters on model performance and observed that learning rate and weight decay exerted the greatest influence (Figure S1). For industrial manufacturing, the majority of RNA vaccines are capable of loading up to 20 immunopeptides (33). We therefore adopted top-20 recall as the primary screening criterion after model training. As a benchmark, we downloaded the pre-trained models provided by Markus et al. and additionally retrained 10 models using their publicly released code, selecting the best-performing ones for subsequent evaluation (Figure S2A–D). With the positive class weight fixed at 10,000, we trained 20 models and chose the optimal one for evaluation (Figure S2E). We also incorporated a previously established Voting strategy based on probability-weighted allocation, as prior work has demonstrated that voting across multiple models can further improve prediction accuracy. We first compared the top-N recall of the four models. 20 peptides are considered as the maximum peptides that most mRNA vaccines encode (33). 50 peptides are considered as the most adenovirus vaccine can load (34, 35). The top 100 are recommended by TESLA to further verify their immunogenity (31). We find that the Transformer exhibited a significant advantage over LR, XGBoost, and their voting ensemble on the NCI test set, though no marked improvement was observed on the TESLA or HiTIDE datasets (Figure 1D–F). Overall, the Transformer showed superior recall for immunogenic peptides within the top-20 and top-50 lists, whereas its recall within the top-100 list was inferior to that of LR (Figure 1G). To comprehensively evaluate the Transformer model in personalized neoantigen prediction, we compared it against all 25 teams participating in the Tumor Neoantigen Selection Alliance (TESLA), as well as the previously published LR and XGBoost voting classifier, denoted here as LRXGB-Voting. We also adopted a voting strategy combining the optimal Transformer and LR models, referred to as TransLR-Voting. All models were assessed using three unified metrics employed by Wells et al.: fraction ranked (FR), which is equivalent to top-100 recall; top-20 immunogenic fraction (TTIF); and area under the precision-recall curve (AUPRC) (31). Encouragingly, our trained Transformer model achieved the best performance among all teams in terms of both FR and TTIF (Figure 1H– I). Although it did not rank first in AUPRC, it still outperformed the LRXGB-Voting model trained on the same input and output data (Figure 1J). Since full precision-recall curves tend to assume an "L" shape, which compresses visualization of fine-grained performance differences, only the segments corresponding to the top-50 and top-100 retrievals are shown in the supplementary materials (Figure S3). Notably, although previous studies suggest that weighted classifier ensembles generally outperform individual models, our TransLR-Voting approach did not demonstrate a clear advantage over the Transformer-only counterpart (Figure 1H–J). 2.2 Differential analysis of three best models Given the improved performance of the Transformer model, we sought to further dissect the detailed differences among the three classifiers. In this section, we present an in-depth comparative analysis of the LR, Transformer, and XGBoost models. We first compared the predicted probability scores generated by each model. Notably, both LR and Transformer adopted relatively conservative strategies, as their highest predicted immunogenicity probability scores remained below 0.5, whereas the scores predicted by XGBoost were distributed across the full 0–1 range (Figure S4). We then analyzed the pairwise overlap of prediction results among the three best- performing models using the Jaccard index. Our results showed that peptides with higher predicted rankings exhibited lower pairwise redundancy across models. Specifically, Transformer and LR shared the highest overlap for top-10 predictions, while XGBoost and LR showed the highest overlap for top-20, top-50, and top-100 predictions (Figure 2A). We also compared recall overlap across different datasets and observed that the classifiers achieved the highest overlap on the HiTIDE dataset and the lowest on the TESLA dataset (Figure S5). Next, we evaluated the consistency of the peptides recalled within the Top-N predictions of all three models. In the top-10 prediction sets, only 43.2% of peptides were simultaneously recalled by all three models. The proportions of peptides recalled by exactly two models were 13.6% (Transformer and XGBoost), 9.09% (Transformer and LR), and 2.27% (XGBoost and LR), respectively (Figure 2B). The Jaccard indices for the different datasets are provided in the supplementary materials (Figure S6). As expected, as Top-N increased, the enrichment of immunogenic peptides within the predicted sets gradually approached saturation, and the proportion of immunogenic peptides concurrently retrieved by all three high-performance models rose correspondingly (Figure 2B). For most given patients, the presence of immunogenic peptides within the top-20 predicted candidates is a prerequisite for eliciting a specific immune response to the personalized mRNA vaccine and, consequently, for the potential efficacy of the therapy in that patient. To this end, we identified patients who had immunogenic peptides in the top- 20 predictions from each of the three models and performed a detailed comparative analysis across the three patient groups. Our findings revealed that when the top-20 predictions were used for vaccine design, Transformer was projected to benefit two more patients than either XGBoost or LR. Moreover, in three patients, immunogenic peptides were exclusively recalled by the Transformer algorithm (Figure 2C). The detailed prediction results for all patients by the three classifiers are provided in the supplementary materials (Figures S7– S9). 2.3 Overlap filtering impact on the prediction performance In the study by Wells et al., it was generally assumed that shared peptides predicted by independent pipelines carry higher confidence in terms of immunogenicity, and thus the peptides most frequently ranked within the top 50 across all participants were selected for experimental testing (7). To investigate whether intersection filtering could serve as a viable ensemble strategy to further improve the Top-N prediction recall of individual machine learning models, we constructed overlapping candidate sets by taking the intersection of the Top 100 peptides predicted by different models. We first performed pairwise intersection analysis between every two classifiers (LR, Transformer, and XGBoost) and quantified the number of shared candidate peptides retained within the Top 100 prediction scope for each patient. The results showed that after pairwise intersection filtering, the number of overlapping peptides retained in the Top 100 pool for each patient remained no fewer than 50 across all three datasets (Figure S10 A– C). We then extended this approach to integrate predictions from all three classifiers simultaneously. Under this three-model intersection strategy, the number of consensus peptides within the Top 100 dropped substantially, with four patients yielding fewer than 50 overlapping candidates (Figure S10 D). To further evaluate the performance of intersection-based filtering, we tested multiple ranking schemes in which peptides within the intersection sets were re-ranked according to the predicted binding probabilities generated by each respective classifier. Across all pairwise and three-way intersection workflows tested, we consistently observed that intersection filtering failed to improve the Top-N recall of any single base classifier (Figure 3 A–C). Taken together, these findings indicate that simply intersecting high-ranking peptide candidates from multiple models does not enhance the recall performance of individual prediction algorithms in neoantigen screening. Nevertheless, intersection filtering did increase the proportion of immunopeptides among the retained candidates, most likely by reducing false negatives (Figure 3 D–E). Therefore, although overlap filtering can be useful for excluding negative peptides under budget-constrained conditions, it should not be relied upon as a strategy for improving Top-N recall. 2.4 The impact of feature elimination on the prediction results In real-world applications, extracting a larger set of features typically entails prohibitively labor-intensive work. To reduce computational costs and experimental workload without compromising model performance, we sought to identify and eliminate unimportant features. To this end, we employed two complementary strategies to rank feature importance: (1) directly extracting the feature weights from the input layer of the best- performing model, and (2) applying SHAP (SHapley Additive exPlanations) to assess feature contributions to the same model. Interestingly, the input-layer weights of the best Transformer model showed only a modest correlation (r = 0.29) with the SHAP-derived importance values, indicating that the two measures are not interchangeable (Figure 4A). The most notable discrepancies were observed for "NetStab Rank" and "Mutation at Anchor," both of which ranked highly in SHAP importance but considerably lower in input-layer weights (Figure S10). Given this divergence, and uncertainty as to which metric more faithfully reflects true feature contributions, we initially extracted the top 15 features from each ranking method and retrained the model separately (Figure 4B). The results revealed that the model trained on the SHAP-selected top 15 features outperformed its counterpart trained on the input-layer- weight-selected top 15 features (Figure 4C). Encouragingly, we found that using these reduced feature sets for model training and evaluation did not significantly impair overall performance (Figure 4D). Notably, even with the sparse SHAP-selected top 15 features, the best model still surpassed both logistic regression and XGBoost models trained on the full 31-feature set, demonstrating the feasibility of sparse feature representation within the Transformer framework (Figure 4D). Given that models trained on the SHAP top 15 features achieved both higher upper bounds and higher mean values in top‑20 recall, we proceeded with this subset for the final elimination analysis. In this analysis, we observed a sharp drop in top‑20 recall between the top 7 and top 6 features (Figure 4D), strongly suggesting that the first 7 features represent the most critical determinants for immunogenicity prediction. These seven core features are: NetMHCpan Rank, MixMHCpred Rank, PRIME Rank, Mutation at Anchor, NetStab Rank, MixMHCpred Score DAI, and TCGA Cancer Expression. Of particular note, Mutation at Anchor emerged as an unexpectedly influential factor in neoantigen ranking. Moreover, TCGA-derived cancer expression level exhibited greater importance than expression values calculated from tumor RNA‑seq data, underscoring the value of publicly available reference expression resources in this context. 3 Materials and Methods 3.1 A Quick Glimpse of Transformers Transformers have served as a pivotal role in many pioneering algorithms in natural language processing, computer vision, bionformatics, etc. To capture the complex interplay relationship between peptide features and peptide immunogenicity, we employed a standard transformer architecture as the core learnable component of our prediction paradigm. As opposed to the previously used traditional LR and XGBoost, transformers can capture long-range dependencies and sequence-context interactions via the self- attention mechanism, which is critical for recognizing immunogenic patterns of patient- derived variant peptides. 3.2 Datasets for Model Training and Testing We used the harmonized datasets published in previous paper containing indenpendent NCI, TESLA and HiTIDE datasets, which is available at site https://figshare.com/s/147e67dde683fb769908 (26). The initial 31 features were used as input for neoantigen prediction. There are a total of 660121,142546, 49378 samples for three datasets, respectively. Each sample represent one peptide. All the peptides are among 8-12 amino acids length which is restrictive to HLA class I, which containing the immunogenic information, where the negative peptides are much more than positive ones. For peptides derived from the projects of NCI is the largest compilation, which contains 660818 negative peptides and 103 positive peptides, we use part of them labeled as NCI- training for model training (487670) and the rest of them (172451) for model testing. Another two independent datasets TESLA and HiTIDE are also used for model testing to verify the generalization of prediction. Overall, the dataset for testing include 364375 samples from 42 patients with 96 positive samples. The dataset for model training (NCI-training) was further processed using 5-fold cross-validation. Specifically, the entire NCI-training set was randomly split into five mutually exclusive folds; in each fold iteration, four folds (80%) were adopted as the training subset and the remaining one fold (20%) served as the validation subset. To avoid data leakage, none of the independent test sets were exposed during the entire model training and cross-validation phase. Normalization & Filtering. Since there are some dirty data and empty values in the raw input features, we remove these sequences based on the self-designed rules and normalize the filtered sequences following NeoRanking (26). We do not perform additional data augmentation since we empirically find that more data augmentations lead to inferior performance. Multi-Process and GPU Acceleration. In the experiment, we adopt the multi- process framework in PyTorch to accelerate the training phase. We utilize GPU to accelerate the training and inference of our model. 3.3 Hyper-parameter Setting and Optimization At the core of the hyper-parameter optimization is the definition of the scoring function which evaluates the quality of the searched hyper-parameters. We first conduct a comprehensive test of different hyper-parameter settings. After the exploration of hyper- parameters optimization, we fix the dimension of models as 48, the number of attention heads as 2, the number of layers as 2, the patience of early stopping as 15, the number of training epochs as 20, and narrowed down the learning rate and weight decay within certain ranges (learning rate: 1e-6 ~ 5e-4; weight decay: 1e-3 ~ 5e-2). Model scoring function. Following NeoRanking, we select the summation of the ranking of immunogenic peptides as the scoring function. 푆푐표푟푒 = ∑ 푒 !"∙(% ! !&) ()& , α = 0.05 Score is the scoring result of the model used for hyper-parameters optimization. α is a contant. r i is the ranked value of immunogenic peptides. 3.4 Realization of Voting Strategy The score for each peptide was recauculated by distributing 0.5 weights to TransNRank and 0.5 weights to LR, respectively. 3.5 Handling of unbalanced data At the training phase, negative neo-peptides was downsampled to 100000 when running the Hyperopt optimization, while all the immunogenic 82 neo-peptides are used for training. The weight of positive neo-epitopes in the binary cross-entropy (BCE) loss was set as the ratio of negative to positive peptides (1314), or higher (10000). 푙표푠 = −푤 * ∙푦 +,%- ∙푙표푔(푝) − (1−푦 +,%- )∙푙표푔(1−푝) This is the formula to calculate loss in our model, where w p represents the weight of positive neo-epitopes, y true means the true label of peptides, p is the predicted value of peptides. When the positive peptides (y true = 1) have lower p near 0, the loss value would be much higher augmented by w p . 3.6 Model Interpretation with Shapley Analysis and input layer weights We used two complementary approaches - SHapley Additive exPlanations (SHAP) and input layer weight inspection to interpret feature contributions (30, 36). 3.7 Ablation Experiment After retaining only the top 15 features ranked by SHAP values, we performed ablation experiments by sequentially eliminating the lowest-ranked features one by one. Each trimmed feature subset was adopted for subsequent model training and testing, with the feature set strictly consistent between the training and testing phases. For each feature combination, we trained 30 independent models, computed the average top-20 recall, and visualized the maximum, mean, and minimum predictive performance in the final results. 3.7 Data Processing and Visualization All the data processing and visualization use Python or R. 3.8 Statistical Analysis The Wilcoxon rank-sum (Mann-Whitney U) test was used for pairwise comparisons of non-parametric data among independent groups. A two-tailed P < 0.05 was defined as statistically significant. Discussion In this work, we proposed TransNRank, the first multi-feature integrated Transformer framework for personalized neoantigen immunogenicity prediction. Benchmark results verified that Transformer architecture outperforms traditional machine learning algorithms, offering guidance for feature screening in clinical translation. TransNRank’s advantages over LR and XGBoost stem from self-attention, which models complicated nonlinear correlations between immunogenic features. Its performance gain is most obvious on top-20 and top-50 recall, two core evaluation indicators closely linked to personalized vaccine peptide screening. Distinct prediction patterns exist among the three models. Low Jaccard coefficients, especially for high-ranked candidates, prove their complementary effects; merely 43.2% of top-10 peptides are commonly predicted by all three models, meaning each model captures unique immunogenic signatures. Though ensemble voting generally boosts model performance in many tasks, our TransLR-Voting failed to surpass standalone TransNRank. It indicates the Transformer has extracted nearly all effective predictive information, while extra base models only introduce redundant noise. Intersection filtering elevates the proportion of immunogenic peptides in retained candidates yet fails to lift top-N recall. Strict intersection filtering forms a consensus bottleneck and discards true positives with inconsistent predictions across models, raising false negative rates. Nevertheless, candidate intersection is suitable for scenarios with limited experimental validation capacity. We thus recommend intersection filtering as a negative screening method to eliminate low-confidence peptides, rather than a positive selection strategy to enrich immunogenic candidates. Feature ablation experiments deliver instructive clinical findings. Top-20 recall declines sharply when feature quantity is reduced from seven to six, demonstrating that only a small feature set dominates prediction performance. Seven core features (NetMHCpan Rank, MixMHCpred Rank, PRIME Rank, Mutation at Anchor, NetStab Rank, MixMHCpred Score DAI, TCGA Cancer Expression) suffice for accurate prediction and greatly cut computational costs. Two findings deserve attention. First, Mutation at Anchor is far more influential than previously recognized, as anchor-site mutations disrupt peptide-HLA binding stability and further affect T cell recognition, which highlights the importance of previous findings. Second, pan-cancer TCGA expression outperforms individual patient RNA-seq data, because large-scale TCGA cohorts provide stable expression signals free from single- sample technical bias. This finding suggests public reference expression databases can replace patient-specific RNA-seq, lowering clinical detection costs. Several limitations remain in our research. First, the limited training set (57 patients, 82 positive peptides) restricts the construction of deeper Transformer structures. Second, our model only targets 8–12 amino acid HLA-I peptides and is not applicable to HLA-I neoantigens with distinct presentation rules. Third, all validation labels originate from in vitro assays instead of real clinical vaccine responses, while the correlation between in vitro activity and in vivo efficacy remains unclear. Fourth, TCR-peptide binding features are not integrated into the current framework. We plan to incorporate mature TCR prediction and structural modeling modules in follow-up upgrades (37-40). Overall, TransNRank validates the great potential of Transformer models in neoantigen prediction. It achieves higher prediction performance and rational feature simplification, highlighting the vital roles of anchor mutations and pan-cancer reference expression. With the rapid clinical development of neoantigen vaccines, TransNRank can facilitate candidate screening and benefit precision tumor immunotherapy. 1. Castle JC, Kreiter S, Diekmann J, Löwer M, van de Roemer N, de Graaf J, et al. Exploiting the mutanome for tumor vaccination. Cancer Res. 2012;72(5):1081-91. 2. Sahin U, Derhovanessian E, Miller M, Kloke BP, Simon P, Löwer M, et al. Personalized RNA mutanome vaccines mobilize poly-specific therapeutic immunity against cancer. Nature. 2017;547(7662):222-6. 3. Xie N, Shen G, Gao W, Huang Z, Huang C, Fu L. Neoantigens: promising targets for cancer therapy. Signal Transduct Target Ther. 2023;8(1):9. 4. Ragone C, Cavalluzzo B, Mauriello A, Tagliamonte M, Buonaguro L. Lack of shared neoantigens in prevalent mutations in cancer. Journal of Translational Medicine. 2024;22(1):344. 5. Platten M, Bunse L, Wick A, Bunse T, Le Cornet L, Harting I, et al. A vaccine targeting mutant IDH1 in newly diagnosed glioma. Nature. 2021;592(7854):463-8. 6. Lang F, Schrörs B, Löwer M, Türeci Ö, Sahin U. Identification of neoantigens for individualized therapeutic cancer vaccines. Nat Rev Drug Discov. 2022;21(4):261-82. 7. Ye L, Zhao G, Ma J, Gan Q, Luo N, Pan D, et al. Personalized cancer vaccines: bridging immune-oncology and precision medicine for advanced therapeutics. Signal Transduct Target Ther. 2026;11(1). 8. Sahin U, Schmidt M, Derhovanessian E, Cortini A, Vogler I, Omokoko T, et al. Individualized mRNA vaccines evoke durable T cell immunity in adjuvant TNBC. Nature. 2026;651(8107):1088-96. 9. Rojas LA, Sethna Z, Soares KC, Olcese C, Pang N, Patterson E, et al. Personalized RNA neoantigen vaccines stimulate T cells in pancreatic cancer. Nature. 2023;618(7963):144-50. 10. Lausen M, Petersen NV, Long GV, Khattak MA, Ascierto PA, Queirolo P, et al. Immunogenicity of an AI-designed personalized neoantigen vaccine, EVX-01, in combination with anti-PD-1 therapy in patients with metastatic melanoma. Journal of Clinical Oncology. 2024;42(16_suppl):9561-. 11. Yarchoan M, Gane EJ, Marron TU, Perales-Linares R, Yan J, Cooch N, et al. Personalized neoantigen vaccine and pembrolizumab in advanced hepatocellular carcinoma: a phase 1/2 trial. Nat Med. 2024;30(4):1044-53. 12. Rubinsteyn A, Kodysh J, Hodes I, Mondet S, Aksoy BA, Finnigan JP, et al. Computational Pipeline for the PGV-001 Neoantigen Vaccine Trial. Front Immunol. 2017;8:1807. 13. Tang S, Madhavan S. neoantigenR: An annotation based pipeline for tumor neoantigen identification from sequencing data. bioRxiv. 2017:171843. 14. Xiao W, Ren L, Chen Z, Fang LT, Zhao Y, Lack J, et al. Toward best practice in cancer mutation detection with whole-genome and whole-exome sequencing. Nature Biotechnology. 2021;39(9):1141-50. 15. Katsikis PD, Ishii KJ, Schliehe C. Challenges in developing personalized neoantigen cancer vaccines. Nat Rev Immunol. 2024;24(3):213-27. 16. Ott PA. The promises and challenges of neoantigen cancer vaccines. Nat Biotechnol. 2026. 17. Reynisson B, Alvarez B, Paul S, Peters B, Nielsen M. NetMHCpan-4.1 and NetMHCIIpan-4.0: improved predictions of MHC antigen presentation by concurrent motif deconvolution and integration of MS MHC eluted ligand data. Nucleic Acids Res. 2020;48(W1):W449-w54. 18. Hundal J, Kiwala S, McMichael J, Miller CA, Xia H, Wollam AT, et al. pVACtools: A Computational Toolkit to Identify and Visualize Cancer Neoantigens. Cancer Immunol Res. 2020;8(3):409-20. 19. Gartner J, Parkhurst MR, Gros A, Tran E, Jafferji MS, Copeland A, et al. A machine learning model for ranking candidate HLA class I neoantigens based on known neoepitopes from multiple human tumor types. Nat Cancer. 2021;2(5):563-74. 20. Pishesha N, Harmand TJ, Ploegh HL. A guide to antigen processing and presentation. Nat Rev Immunol. 2022;22(12):751-64. 21. Poran A, Scherer J, Bushway ME, Besada R, Balogh KN, Wanamaker A, et al. Combined TCR Repertoire Profiles and Blood Cell Phenotypes Predict Melanoma Patient Response to Personalized Neoantigen Therapy plus Anti-PD-1. Cell Rep Med. 2020;1(8):100141. 22. Jurtz V, Paul S, Andreatta M, Marcatili P, Peters B, Nielsen M. NetMHCpan-4.0: Improved Peptide-MHC Class I Interaction Predictions Integrating Eluted Ligand and Peptide Binding Affinity Data. J Immunol. 2017;199(9):3360-8. 23. Tadros DM, Racle J, Gfeller D. Predicting MHC-I ligands across alleles and species: how far can we go? Genome Med. 2025;17(1):25. 24. O'Donnell TJ, Rubinsteyn A, Laserson U. MHCflurry 2.0: Improved Pan-Allele Prediction of MHC Class I-Presented Peptides by Incorporating Antigen Processing. Cell Syst. 2020;11(1):42-8.e7. 25. Gfeller D, Schmidt J, Croce G, Guillaume P, Bobisse S, Genolet R, et al. Improved predictions of antigen presentation and TCR recognition with MixMHCpred2.2 and PRIME2.0 reveal potent SARS-CoV-2 CD8(+) T-cell epitopes. Cell Syst. 2023;14(1):72- 83.e5. 26. Müller M, Huber F, Arnaud M, Kraemer AI, Altimiras ER, Michaux J, et al. Machine learning methods and harmonized datasets improve immunogenic neoantigen prediction. Immunity. 2023;56(11):2650-63.e6. 27. Zhao X, Wei L, Xie Z, Zhang X. NeoGuider: neoepitope prediction using advanced feature engineering. Genome Med. 2025;18(1):13. 28. Borch A, Carri I, Reynisson B, Alvarez HMG, Munk K, Montemurro A, et al. IMPROVE: a feature model to predict neoepitope immunogenicity through broad-scale validation of T-cell recognition. Front Immunol. 2024;15:1360281. 29. Thrift WJ, Lounsbury NW, Broadwell Q, Heidersbach A, Freund E, Abdolazimi Y, et al. Towards designing improved cancer immunotherapy targets with a peptide-MHC-I presentation model, HLApollo. Nat Commun. 2024;15(1):10752. 30. Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, et al. Attention is all you need. Advances in neural information processing systems. 2017;30. 31. Wells DK, van Buuren M, Dang K, Hubbard-Lucey VM, Sheehan KCF, Campbell KM, et al. Key Parameters of Tumor Epitope Immunogenicity Revealed Through a Consortium Approach Improve Neoantigen Prediction. Cell. 2020;183(3):818- 34.e13. 32. PyTorch v2.13.0 Transformer [Available from: https://docs.pytorch.org/docs/2.13/generated/torch.n.Transformer.html. 33. Ott PA, Hu Z, Keskin DB, Shukla SA, Sun J, Bozym DJ, et al. An immunogenic personal neoantigen vaccine for patients with melanoma. Nature. 2017;547(7662):217- 21. 34. Bartolomeo R, Troise F, Allocca S, Sdruscia G, Vitale R, Bignone V, et al. Development of a Potency Assay for Nous-209, a Multivalent Neoantigens-Based Genetic Cancer Vaccine. Vaccines (Basel). 2024;12(3). 35. D’Alise AM, Willis J, Duzagac F, Hall MJ, Cruz-Correa M, Idos GE, et al. Nous-209 neoantigen vaccine for cancer prevention in Lynch syndrome carriers: a phase 1b/2 trial. Nature Medicine. 2026;32(3):1002-11. 36. Lundberg SM, Lee S-I. A unified approach to interpreting model predictions. Advances in neural information processing systems. 2017;30. 37. Wu F, Zhao Y, Xiao Y, Qin C, Wang F, Wu Z, et al. Fast and accurate modeling of TCR-peptide-MHC complexes using tFold-TCR. bioRxiv. 2025:2025.01.12.632367. 38. Gao Y, Gao Y, Wu S, Li D, Zhou C, Meng F, et al. Weakly supervised peptide-TCR binding prediction facilitates neoantigen identification. Cell Syst. 2025;16(10):101403. 39. Jiang M, Yu Z, Lan X. VitTCR: A deep learning method for peptide recognition prediction. iScience. 2024;27(5):109770. 40. Pham MN, Nguyen TN, Tran LS, Nguyen QB, Nguyen TH, Pham TMQ, et al. epiTCR: a highly sensitive predictor for TCR-peptide binding. Bioinformatics. 2023;39(5). Figure 1. Dataset composition and model performance benchmarking (A–C) Summary statistics of patient and peptide counts across datasets. (A) Patient numbers (teal) and immunogenic peptide counts (maroon) for training and held-out test partitions. (B) Patient and immunogenic peptide counts for the independent benchmark datasets NCI-test, TESLA, and HiTIDE. (C) Number of patients (maroon, immunogenic peptides) and total candidate peptides (cyan) within each independent benchmark dataset. (D–G) Immunogenic peptide recall count achieved by each prediction model (Transformer, LR, XGBoost, LRXGB-Voting) across benchmark cohorts: (D) NCI-test, (E) TESLA, (F) HiTIDE, and (G) aggregated total cohort. Colored bars denote recall thresholds for Top20, Top50, and Top100 ranked neoantigens; horizontal red line marks the recall level of the Transformer baseline. (H–J) Benchmark performance of all participating teams across evaluation metrics. Horizontal boxplots illustrate the distribution of team scores for (H) TTIF, (I) FR, and (J) AUPRC. Teams are vertically sorted for consistent comparison across panels. Figure 2. Correlation and overlap analysis of LR, Transformer and XGBoost neoantigen prediction models. (A) Pairwise Jaccard index heatmaps showing the similarity of ranked neoantigen candidate sets (Top 10, Top 20, Top 50, Top 100) predicted by LR, Transformer and XGBoost. Color intensity corresponds to Jaccard index values. (B) Venn diagrams illustrating the overlap of immunogenic peptides retrieved within Top 10, Top 20, and Top 50 candidate neoantigens across the three models. Absolute counts and percentage fractions are labelled within each region. Numbers in parentheses denote the total immunogenic peptides captured by each individual model. (C) Venn diagram showing overlapping patients with detected immunogenic peptides (imm-peps) when each model selects the Top 20 neoantigen candidates. Values indicate absolute patient counts and corresponding proportions. Figure 3. Immunogenic peptide recall and positive neoantigen enrichment of multi-model intersected candidate peptides. (A–C) Aggregated immunogenic peptide recall across all patients using intersecting Top100 candidate peptides. Recall values are stratified by ranking cutoffs (Top10, Top20, Top50) and grouped by different intersection strategies: peptides unique to each single model and shared candidates between model pairs. Intersecting peptides excluded (“Intersection not taken”) serve as the reference baseline. (D-E) Patient-sorted positive label ratio (immunogenic fraction) of three-model overlapping peptides (red) versus Transformer-only peptides (gray) for top100 (D) and top50 (E) intersected peptides, with annotated group mean values. ***p < 0.001, two-group comparison. Figure 4. Feature importance analysis and ablation experiments guided by SHAP values. (A) Scatter plot of absolute input layer weight versus absolute SHAP importance for all features. Dashed lines indicate thresholds for Top15 feature selection under each metric. (B) Pie charts showing category distribution of Top15 features ranked by input weight (left) and SHAP importance (right). (C) Top20 immunogenic peptide recall for feature subsets defined by input weight ranking, SHAP ranking, and their overlapping features. (D) Iterative feature elimination experiment sorted by SHAP importance. Top20 recall is tracked upon sequential removal of low-ranked features, with shaded region indicating the range between minimum and maximum recall. Supplementary Materials Supplementary Figure 1 Figure S1. Top-20 recall performance against four hyperparameters: batch size, dropout rate, learning rate and weight decay. Scatter points represent the measured top-20 recall values from individual hyperparameter trials. Coloured smoothed curves illustrate the general trend of top-20 recall as each hyperparameter varies. The red dashed horizontal line denotes the baseline recall achieved by the logistic regression (LR) model. Supplementary Figure 2 Figure S2. Top-20 recall across repeated model runs for different architectures. (A) Published logistic regression (LR) model, 10 independent training runs. (B) Retrained logistic regression (LR) model, 10 independent training runs. Red box highlights the runs achieving peak recall performance. (C) Published XGBoost model, 10 independent training runs. (D) Retrained XGBoost model, 10 independent training runs. Red box marks the run with maximum top-20 recall. (E) Transformer model, 20 independent training runs. Red box denotes the best- performing run. Bar colour gradient corresponds to the Top-20 recall value of each individual run. Mean, minimum and maximum recall values are annotated for each panel. Supplementary Figure 3 Figure S3. Precision-recall curves for logistic regression (LR), XGBoost, and Transformer ranking models. (A) Precision-recall curves generated under the Top-100 candidate ranking setting. (B) Precision-recall curves generated under the Top-50 candidate ranking setting. Green, blue, and red lines correspond to the LR, XGBoost, and Transformer models, respectively. Light pink shaded regions mark the practical recall range prioritized for downstream candidate screening. Supplementary Figure 4 Figure S4. Distribution of model-predicted probabilities stratified by true binary labels. Panels from left to right correspond to the logistic regression (LR), Transformer, and XGBoost models. The x-axis denotes ground-truth labels (0 = negative samples, 1 = positive samples), and the y-axis shows predicted probability scores. Each violin plot embeds boxplots and overlaid scatter points representing individual samples. Asterisks indicate statistically significant differences between predicted probability distributions of positive and negative groups. Supplementary Figure 5 Figure S5. Pairwise Jaccard index heatmaps quantifying overlaps of top- ranked candidate peptides predicted by LR, Transformer and XGBoost. Model similarities are evaluated at four ranking thresholds: Top 10, Top 20, Top 50, and Top 100 peptides. (A) NCI dataset; (B) TESLA dataset; (C) HiTIDE dataset. Each cell value represents the Jaccard index between candidate peptide sets from two models, and colour gradient indicates the magnitude of set overlap. Supplementary Figure 6 Figure S6. Triple Venn diagrams illustrating overlaps of top-ranked candidate peptides predicted by LR, Transformer and XGBoost. Comparisons are shown for Top 10, Top 20, Top 50, and Top 100 peptide cutoffs. (A) NCI dataset; (B) TESLA dataset; (C) HiTIDE dataset. Numbers inside each region denote the count of unique peptides, and percentages represent the proportion relative to the union of peptides from the three models. Parenthetical values beside each model label indicate the total number of peptides retrieved by that model at the corresponding ranking threshold. Supplementary Figure 7 Figure S7. Patient-level violin plots of predicted probabilities output by the Transformer model. Each subplot corresponds to an individual patient. The x-axis represents ground- truth labels (0 = negative peptide samples, 1 = positive peptide samples), and the y-axis denotes model-predicted binding probabilities. Violin distributions are overlaid with individual scatter points for each peptide. P-values indicate statistical differences between predicted probability distributions of positive and negative samples; NA denotes non-significant comparisons. Asterisks mark significance levels, and “ns” indicates no statistically significant difference. Supplementary Figure 8 Figure S8. Patient-level violin plots of predicted probabilities generated by the logistic regression (LR) model. Each panel corresponds to an individual patient. The x-axis indicates ground-truth labels (0 = negative peptide samples, 1 = positive peptide samples), and the y-axis represents model-predicted binding probabilities. Violin distributions are overlaid with scatter points for individual peptides. P-values assess statistical differences between predicted probability distributions of positive and negative samples; NA denotes non- significant comparisons. Asterisks denote significance levels. Supplementary Figure 9 Figure S9. Patient-level violin plots of predicted probabilities generated by the XGBoost model. Each subplot corresponds to an individual patient. The x-axis represents ground- truth labels (0 = negative peptide samples, 1 = positive peptide samples), and the y-axis denotes model-predicted binding probabilities. Violin distributions are overlaid with scatter points for individual peptides. P-values indicate statistical differences between predicted probability distributions of positive and negative samples; NA denotes non- significant comparisons. Asterisks mark significance levels. Supplementary Figure 10 Figure S10. Intersection quantity of the top 100 prioritized candidate neoepitope peptides predicted by four distinct ensemble machine learning frameworks in individual patient samples. (A) Linear regression (LR) combined with XGBoost prediction framework; (B) Transformer combined with XGBoost prediction framework; (C) Transformer combined with linear regression (LR) prediction framework; (D) Triple integrated framework combining Transformer, LR and XGBoost. The y-axis denotes the number of overlapping peptides retained in the top-100 high-scoring peptide list for each patient. Blue dots represent the HITIDE patient cohort, green dots represent the NCI-test cohort, and red dots represent the TESLA cohort. Patients are sequentially arranged along the x-axis.