Natural killer (NK) cells regulate tumor surveillance and antiviral immunity through the integration of activating and inhibitory receptor signals. Peptide loaded to HLA-E can bind to CD94-NKG2A/C heterodimer, with the two receptors delivering opposing signals. Despite these divergent functional outcomes, NKG2A and NKG2C differ by only four amino acids in their extracellular domains, rendering computational prediction of receptor preference a difficult task.
Using a publicly available dataset measuring NKG2A/C-mediated NK cell activation by HLA-E-bound peptides, we developed a quantum-inspired kernel classifier to discriminate receptor preference from peptide sequence alone. Nine physicochemical descriptors per amino acid position were encoded as quantum rotation angles, yielding a product-form kernel that weights position-wise compatibility and captures higher-order feature dependencies inaccessible to additive models. In cluster-disjoint cross-validation designed to minimize sequence leakage, the kernel achieved AUC = 0.87, above machine learning methods such as gradient boosting (0.83), convolutional neural network (0.81), and Boltz-1 inspired structural model(0.46). Importantly, in an independent dataset of 80 VL9-variant peptides withheld from training, the quantum kernel generalized mostß robustly (AUC = 0.80 versus 0.64–0.69 for all classical baselines), suggesting that the multiplicative kernel structure better captures the underlying biology. Kernel matrices computed on IBM quantum processing hardware showed high fidelity to classical simulation, confirming near-term hardware viability.
Permutation-based feature importance analysis identified positions 3, 6, and 8 as the principal individual determinants of receptor discrimination, with hydrophobicity as the dominant physicochemical contributor. Pairwise permutation analysis further revealed synergistic coupling between positions 2 and 8—residues flanking the canonical HLA-E anchor sites—whereby individual perturbation produced only marginal effects, but joint perturbation markedly degraded classification performance.
These results implicate conformational coupling between the peptide N- and C-termini in NKG2A/C selectivity, consistent with differential presentation of the peptide backbone to the receptor binding interface. By resolving cooperative positional dependencies that conventional sequence- and structure-based classifiers fail to capture, the quantum-inspired kernel framework addresses a previously intractable discrimination problem in NK cell biology. These findings open avenues for structure-guided peptide engineering in NK cell-directed immunotherapy and vaccine design.