TL;DR
This paper introduces MultiTAP, a novel multi-modal deep learning framework that integrates both sequence and structural information of TCR-peptide-HLA complexes to predict tumor antigen immunogenicity. By leveraging this rich data, MultiTAP achieves superior performance compared to existing methods, offering a promising tool for advancing cancer immunotherapies.
What Problem Does This Paper Address?
Identifying immunogenic tumor antigens is crucial for developing effective cancer immunotherapies. However, this process is challenging due to the complex interactions between T cell receptors (TCRs), tumor-derived peptides, and human leukocyte antigens (HLAs), as well as the inherent variability and evolution of tumor antigens. Current computational models often struggle to accurately predict which tumor antigens will elicit a strong immune response, leading to a need for more sophisticated prediction tools.
Methods
The core of this work is the MultiTAP (Multi-modal Tumor Antigen Predictor) framework. MultiTAP employs a multi-modal deep learning architecture that integrates:
- Sequence Data: TCR CDR3 sequences, peptide sequences, and HLA sequences.
- Structural Data: Predicted structures of peptide-HLA (pHLA) complexes.
Key methodological aspects include:
- Feature Extraction: Using pre-trained protein language models (like ESM-1b) for TCR sequences, one-hot encoding for peptide sequences, PSSM alignments for HLA sequences, and Laplacian/adjacency matrices for pHLA structures.
- Multi-modal Fusion: An attention mechanism is employed to effectively combine features from different modalities (TCR, peptide, and HLA), allowing the model to learn complex interactions. This mechanism helps to simulate the "virtual binding" state of the epitope.
- Hierarchical Processing: The model uses a multi-layer GCNN and attention mechanisms to process and integrate these diverse features.
- Interpretability: The attention mechanism allows for analysis at the residue level, revealing which parts of the TCR-peptide-HLA complex are most influential in predicting immunogenicity.
Datasets
- TPHD (TCR-peptide-HLA Dataset): A newly constructed multi-modal dataset created by aggregating data from three public databases: IEDB, McPAS-TCR, and VDJdb. This dataset includes TCR-peptide-HLA binding data, HLA allele information, and importantly, predicted pHLA structural data. It comprises approximately 17,000 data points.
- HLA-A:11:01 Dataset: A dataset specifically designed to test generalization capabilities, focusing on the HLA-A:11:01 phenotype, which is prevalent in East Asian populations. This dataset was derived from literature on hepatitis B patients and contains 4701 TCR-pHLA entries.
- Benchmark Datasets: Datasets used by existing tools DeepAIR and TEIM were also utilized for comparative evaluation.
Key Insights and Ideas
- The Power of Multi-modality: Integrating sequence and structural information of TCR-peptide-HLA complexes significantly boosts the accuracy of tumor antigen immunogenicity prediction.
- HLA Information is Crucial: The inclusion of HLA sequence and, more importantly, structural information is critical for accurate prediction, especially for diverse HLA phenotypes.
- Attention for Interpretability: The attention mechanism not only improves performance but also provides insights into which residues and interactions within the TCR-peptide-HLA complex are most important for determining immunogenicity.
- Indirect Learning of HLA Information: MultiTAP appears to learn HLA-related information indirectly through its interactions with epitopes and TCRs, rather than directly encoding it from HLA sequences alone.
- Robustness to Variability: The model demonstrates strong generalization capabilities, performing well even on unseen HLA alleles and novel antigen sequences, which is vital for real-world applications.
Key Results
- Superior Performance: MultiTAP significantly outperformed baseline models (NetTCR-2.0 and ERGO-I) on the TPHD dataset across multiple metrics (AUC, PRC, Precision, Recall, F1). For instance, on the independent TPHD test set, MultiTAP achieved higher AUC values compared to baselines.
- Out-of-Distribution (OOD) Robustness: MultiTAP maintained superior predictive performance in OOD tests, demonstrating robustness in classifying antigens across diverse HLA phenotypes and epitope sequences, unlike baseline models which showed significant performance degradation.
- HLA-A:11:01 Performance: On the HLA-A:11:01 dataset, MultiTAP achieved an average prediction AUC exceeding 0.6, with one antigen reaching an AUC close to 0.7, indicating enhanced predictive capability in this specific context.
- Feature Importance: Feature ablation studies revealed that the combination of peptide, HLA sequence, and pHLA structure yielded the most accurate predictions, underscoring the importance of all these modalities.
Limitations and Future Work
- Structural Data Limitations: The authors acknowledge that due to the limited availability of crystal structure data, they relied on predicted structures from xTrimo-Multimer. The accuracy and constraints of this prediction model might have limited a deeper understanding of TCR-peptide-HLA complex binding patterns.
- Future Work: The authors hope to access higher-quality crystal structure data in the future. They also plan to further investigate the binding patterns of TCR-peptide-HLA complexes to provide better suggestions for cancer immunotherapies.
Summary updated