Abstract
By comparing various k-mer length combinations across seven different classification algorithms, we found that B-cell receptor-based cell subset classification outperforms T-cell receptor-based classification, achieving an average AUC of over 96%. This finding offers a new perspective on the classification of immune cells. We also observed that 11 distinct cell subpopulations exhibited differences in cell proportions, inflammatory factor expression, cell communication, and metabolic pathways, with notable activity in metabolic pathways. These variations may reflect the adaptive changes of cell subpopulations in response to different disease states. This study aims to uncover the potential biological significance of immune prediction, target antigens, and effective evaluation by analyzing the immune characteristics of specific cell subsets at the cellular level. These findings will not only enhance our understanding of immune system functions but also offer new directions for the development and optimization of immunotherapy.
Introduction
The human immune system relies on the immense diversity of B-cell receptors (BCRs) and T-cell receptors (TCRs) to recognize and neutralize pathogens. While single-cell sequencing has revolutionized our ability to identify cell types, linking the "immune repertoire"—the collection of these receptor sequences—to specific cellular functions and disease states remains a significant computational challenge. Existing methodologies often prioritize gene usage (VDJ rearrangement) but frequently overlook the fine-grained structural motifs within the receptors that dictate specificity. Furthermore, there is a critical need to integrate repertoire data with functional "omics" to understand how immune subpopulations adapt during infections like COVID-19.
The KDDC Framework
To address these challenges, the researchers developed KDDC, a framework integrating four primary computational stages:
- K-mers Clipping: Highly variable CDR3 amino acid sequences are decomposed into overlapping segments (k-mers) of lengths $k=2, 3, \text{ and } 4$. This process converts categorical sequence data into a numerical count matrix suitable for machine learning.
- Dataset Filtering: To mitigate the high dimensionality and sparsity inherent in k-mer data, threshold filtering is applied to remove outliers and sparse features.
- Dimension Reduction: Analysis of Variance (ANOVA) is employed to retain only those k-mers showing statistically significant differences across groups ($p < 0.05$).
- Classification Algorithms: The refined data is processed through seven machine learning models, including Random Forest (RF), Support Vector Machines (SVM), and eXtreme Gradient Boosting (XGBoost), to classify cell subsets and patient infection status.
Key Insights and Results
The application of KDDC to COVID-19 patient data yielded several significant findings:
- Superiority of BCR Predictivity: A major discovery was that BCR sequences are significantly more effective than TCR sequences for classifying cell subsets and infection status. B-cell-based models consistently achieved an average AUC of over 96%, while the patient status prediction model reached an exceptional 98.8% AUC using 3-mer sequences.
- Identification of Structural Motifs: The study pinpointed specific amino acid motifs (e.g., DLW and FDL) that were highly influential in identifying specific B-cell subsets, suggesting that certain structural "words" are intrinsically tied to cell identity.
- Metabolic Rewiring: Beyond sequence classification, the research integrated metabolic pathway scoring (via KEGG). They observed that COVID-19 induces significant metabolic shifts; for instance, the
B_c1_TCL1Asubpopulation showed upregulated oxidative phosphorylation during mid-stage infection. - Immune Coordination Disruptions: Post-infection analysis indicated a weakened binding affinity between specific CD4+ and CD8+ T-cell subsets, providing a potential mechanistic explanation for immune dysregulation in severe cases.
Conclusion and Future Directions
The KDDC framework demonstrates high accuracy in classifying immune cell heterogeneity and disease states by focusing on receptor structural motifs. While currently validated on COVID-19 datasets, its potential extends to other infectious diseases and oncology. Future work will focus on automating cell annotation and applying the framework to optimize immunotherapy, enabling personalized medical interventions based on an individual's unique immune repertoire.
Summary updated