EpiMII is presented, a Graph Neural Network model for MHC-II epitope design, which learns from the structural features of epitopes to predict their sequences, potentially contributing to vaccine development.
Abstract
MHC-II neoantigens play a critical role in immunotherapy, either as direct effectors or through their influence on CD8+ T cells. However, only a small fraction of tumor DNA mutations qualify as functional neoantigens, and current prediction tools often lack accuracy, leading to the low immunogenicity of predicted neoantigens in vivo. Here, we present EpiMII, a Graph Neural Network model for MHC-II epitope design, which learns from the structural features of epitopes to predict their sequences. To train EpiMII, we constructed a reliable, large dataset containing 142,934 MHC-II epitope structures. This approach achieves a 4.2x improvement over ProteinMPNN, with a sequence recovery rate of 78.0% for known MHC-II epitopes in the Protein Data Bank. As a case study, we designed a neoantigen from hepatocellular carcinoma. All five designed epitopes significantly activated CD4+ T cells in vitro and induced secretion of IFN-γ and TNF-α. Notably, one epitope treatment significantly reduced tumor volume in mice in vivo. EpiMII offers a novel and efficient approach for identifying MHC-II epitopes/neoantigens, potentially contributing to vaccine development.
Abstract Motivation T-cell receptors (TCRs) recognize antigenic peptides presented by major histocompatibility complex (MHC) molecules and are central to adaptive immunity. Computational prediction of TCR–epitope binding (TEB) can accelerate immunotherapy development, yet remains hampered by limited labeled data, false-negative noise in unobserved pairs, and over-smoothing in graph-based models. Results We present DynaTCR, a dynamic graph ensemble learning framework for TEB prediction. DynaTCR encodes TCR and epitope sequences with protein language model embeddings and organizes them into a bipartite interaction graph. A graph regularization-variance-preserving aggregation (GR-VPA) encoder stabilizes message propagation and alleviates over-smoothing, while a global attention layer captures long-range dependencies. Multiple base learners are trained with iteratively updated hard-negative samples to reduce false-negative predictions. Under the StrictTCR evaluation protocol on four public datasets, DynaTCR achieves AUC improvements of 4.0–8.2 percentage points over the strongest existing method and up to 15.8 percentage points in AUPR. On the most stringently curated dataset, DynaTCR attains an AUC of 95.1%. Furthermore, on an independent structure-derived test set, DynaTCR achieves the highest AUC (72.6%) among all compared methods, demonstrating its robustness and effectiveness for TEB prediction and candidate prioritization. Availability Source code and data can be downloaded from: https://github.com/2014402680/TEB/.
The ImmunoFoundation Model (IFM), a multimodal deep learning system that integrates not only peptide sequences, 3D molecular structures, and biochemical properties but also TCR-MHC-peptide to achieve superior immunogenicity prediction and enable peptide optimization for therapeutic applications is developed.
Smita Krishnaswamy, J. Rocha, Hiren Madhu et al.· Journal of Immunology· 0 citations
Accurate identification of interactions between T-cell receptors (TCRs) and antigenic peptides presented by major histocompatibility complex (MHC) molecules is essential for advancing precision immunotherapy. However, existing approaches often exhibit limited generalization to unseen peptides and struggle to capture the complex interaction patterns underlying immune recognition. Here, we present TCR-IFNet, a biologically informed deep learning framework for interpretable TCR-peptide interaction prediction. The model integrates global contextual representations from protein language models with local motif refinement via a gated convolutional module. To model cross-sequence dependencies, we introduce a Fast Kolmogorov-Arnold Network (FastKAN)-based cross-attention mechanism for nonlinear interaction modeling, together with a bilinear attention network to aggregate residue-level features into compact interface representations. Evaluation across multiple settings indicates that TCR-IFNet achieves competitive performance compared with existing methods, with higher AUPRC observed on both antigen-specific and healthy-sourced datasets, as well as improved results on independent test sets. The model also shows consistent generalization to unseen peptides under different negative sampling strategies. In addition, TCR-IFNet provides biologically meaningful interpretability by identifying key residue-level interaction patterns consistent with structural binding interfaces. Collectively, these findings demonstrate that TCR-IFNet provides a robust and generalizable computational framework for characterizing TCR-peptide interactions.
Wen-Yu Xi, Ruheng Wang, Xiu-Cai Ye et al.· International Journal of Bio...· 0 citations
The adaptive immune system monitors cellular integrity by recognizing short peptides from intracellular proteins presented on major histocompatibility complex class I (MHC-I) molecules, collectively termed peptide–MHC complexes (pMHC), enabling detection of foreign or mutated proteins. With the rising importance of immunotherapies targeting cancer neoantigens, accurately predicting which peptides bind to MHC alleles is critical. Current computational methods for pMHC-I binding prediction fall into sequence-based methods, which rely heavily on large training datasets, and structure-based methods that leverage structural modeling and pMHC binding energetics. Although sequence-based methods are widely used, their performance depends on the size and quality of the training data. Structure-based approaches, by contrast, can generalize better across diverse MHC alleles, but they traditionally depend on identifying a single global minimum-energy conformation, an assumption that may be inadequate for the promiscuous binding of MHC-I molecules. To address these limitations, we developed STRUMP-I (STRUcture-based pMHC Prediction for class I), a novel pMHC-I binding prediction tool that directly leverages a broad set of force-field-derived energy terms as machine learning features. In the standard benchmark set, STRUMP-I achieved performance comparable to state-of-the-art sequence-based models overall and showed a clear advantage for alleles with limited or imbalanced representation. Furthermore, STRUMP-I complemented sequence-based methods by removing method-specific false positives and improving precision, with a more favorable precision–recall tradeoff than AF-FT. These evaluations reinforced the value of STRUMP-I as a structure-informed prioritization method, particularly for underrepresented alleles and as a high-precision post-prediction filter.
Adam Voshall, Jeongjun Chae, Honglan Li et al.· Computational and Structural...· 0 citations
Identifying which peptides bind major histocompatibility complex (MHC) molecules is central to vaccine design, neoantigen prioritization, and precision immunotherapy. Existing deep learning predictors largely encode amino acids as discrete symbols, thereby missing the residue-level chemistry driving molecular recognition. Performance also tends to degrade under class imbalance, for rare alleles, and on peptide– MHC combinations outside the training distribution. We developed CALFP-MHC, a framework that encodes each amino acid as a set of complementary cheminformatics fingerprints capturing functional groups, atomic connectivity, and substructural features, and combines positional encoding with supervised contrastive pre-training to organize the latent space by binding class before fine-tuning a binary classifier. Peptide–MHC interactions are modeled through a hybrid convolutional-transformer backbone. In a large-scale computational benchmark covering ∼18.7 million peptide–MHC pairs across 112 HLA class I and 53 class II alleles, CALFP-MHC achieved AUCs of 0.93-0.97 and PPVs of 0.66–0.94. Critically, performance remained above AUC 0.90 even at a 200:1 negative-to-positive ratio, where competing tools frequently collapsed toward chance. On independent experimental data containing 3,627 class I and 520 class II MS/MS-confirmed ligands and 570 validated neoantigens, the model maintained strong discrimination, correctly prioritizing immunogenic peptides and MHC-presented ligands. Attention and integrated-gradient analyses recovered established anchor positions (P2 and PΩ for class I, P1, P4, P6, and P9 for class II) and highlighted chemically interpretable functional groups consistent with known binding determinants. CALFP-MHC demonstrates that grounding residue representations in molecular chemistry, rather than sequence symbols alone, improves both robustness and interpretability in peptide–MHC binding prediction.
My-Diem Nguyen Pham, T. Ho, H. Nguyen et al.· bioRxiv· 0 citations
Adaptive immunity relies on T-cell receptor (TCR) recognition of non-self epitopes, short peptides presented by the Major Histocompatibility Complex (MHC) on the cell surface. Accurate computational prediction of TCR-epitope binding would unlock the development of targeted immunotherapies, such as cancer vaccines and TCR T cell therapies, while simultaneously deepening our fundamental understanding of self/nonself discrimination, pathogen recognition, and autoimmunity. We created an ensemble approach (enFoldX) that leverages structure prediction models such as AlphaFold3 to build sensitive binding predictors. enFoldX can distinguish T cell reactivity between peptides that differ by a single amino acid substitution, as needed for cancer neoantigens. enFoldX utilizes a customized highly parallelized workflow which allows us to produce ensembles of predicted protein structures at scale and train classifiers to infer reactivity based on distributions of engineered structure features and alignment confidence metrics. While state-of-the-art sequence-based approaches we evaluated could predict well for observed TCRs and epitopes close in sequence to training data, their applicability to novel sequences was limited. Conversely, our ensemble approach is the only model that showed true generalizability to novel datasets and even across species. Moreover, our ensemble approach outperforms the current co-folding methods which rely on predictions from the single top ranked structure. By leveraging the entire protein universe at scale, structure ensembles therefore enable classifiers that reflect physical free energies, providing a tractable path towards TCR T therapy design at the sensitivity required for cancer neoantigen discrimination and imparting lessons for a wide array of complex binding problems.
Olga Lyudovyk, Jonathan A. Levine, Melissa Pathil, Stephen Martis, Yuval Elhanati, Vinod P. Balachandran, Quaid Morris, Benjamin D. Greenbaum. enFoldX: AI classification of AlphaFold3-derived structural ensembles enables T cell specificity prediction [abstract]. In: Proceedings of AACR Drug Discovery and Development (AACR D3) Conference; 2026 Jul 21-24; Boston, MA. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(14_Suppl):Abstract nr PR004.
O. Lyudovyk, Jonathan A. Levine, M. Pathil et al.· Clinical Cancer Research· 0 citations
What does it take to trust AI-driven HVAC optimization? Our AI Model Factory combines agents, machine learning, reinforcement learning and deterministic checks in a governed workflow designed for messy, real-world building data. The post We built an AI factory for HVAC control appeared first on GPT-Lab.
MIT News · Artificial Intelligence· news.mit.eduAug 18, 2026
A new method for surgically removing training examples from a model reveals that as datasets grow, the link between what a model learns and what it produces dissolves.
Microsoft Research Blog· microsoft.comJul 30, 2026
LLMs do not get smarter just by remembering more. EvoLib turns experience into evolving knowledge, taking reusable skills and insights that help models learn and adapt across tasks long after deployment. The post EvoLib: Turning experience into evolving knowledge appeared first on Microsoft Research.
MIT News · Artificial Intelligence· news.mit.eduMay 20, 2026
