Quantum Kernels Sharpen Prediction of CAR T-Cell Killing Power

Chimeric antigen receptor T-cell therapy has transformed the treatment of certain blood cancers, but designing the best receptor remains a combinatorial nightmare. A CAR’s extracellular domain determines which antigen the cell recognizes, while the intracellular signaling domains dictate how the engineered cell behaves once it meets its target. Researchers can shuffle a menu of co-stimulatory motifs across several positions inside the receptor, and the number of possible designs grows exponentially with each added motif and position. Exhaustive experimental screening is impossible, so the field has turned to machine learning to guide exploration — with mixed results, because the datasets available are small relative to the vast design space.

A team from IBM Quantum, IBM Research, and the Center for Cellular Construction now reports one of the largest experimental deployments to date of quantum machine learning in the life sciences, using a 61-qubit quantum circuit to help classify the killing power of CAR T-cell designs. Writing in the journal Quantum Machine Intelligence, Filippo Utro, Meltem Tolunay, and colleagues describe how a hybrid quantum-classical technique called the projected quantum kernel, or PQK, allowed a conventional support vector machine to extract biologically meaningful patterns from a sparsely sampled library of receptor designs. The work suggests that near-term quantum processors, even without error correction, may earn their keep in data-starved biological problems.

The underlying dataset came from earlier work by Kyle Daniels and colleagues at Stanford, who built a combinatorial library of 13 intracellular signaling motifs, labeled M1 through M13, that could be placed in up to three positions within the CAR’s intracellular domain, with every construct ending in a terminal M14 motif. The team measured cytotoxicity and stemness for 246 experimentally produced constructs. Cytotoxicity, assessed by how well the engineered cells killed Nalm6 leukemia cells, showed a bimodal distribution with a natural dividing line at a survival value of 0.62, allowing the researchers to frame the problem as binary classification: high cytotoxicity versus low. Stemness, by contrast, showed a single asymmetric peak and was not amenable to clean binary labeling, so the study focused on cytotoxicity.

Encoding the designs for a quantum computer required care. Each of the 14 motifs, plus an empty placeholder for unoccupied positions, was one-hot encoded across four positions, producing a 60-dimensional binary feature vector — hence a 60-qubit circuit for the first embedding. The researchers tested two quantum embeddings drawn from the theoretical literature on quantum kernels. The first, E1, is the ZZ Feature Map, which loads each binary feature onto a qubit as a rotation angle of either pi or pi/2 and entangles qubits through controlled phase gates. The second, E2, is a Trotterized simulation of the one-dimensional Heisenberg model, which required an additional ancilla qubit for a total of 61.

The PQK trick is that the quantum computer never tries to learn anything itself. Instead, the classical data is loaded onto the circuit, where it lives in a high-dimensional Hilbert space, and is then projected back into a classical representation by measuring every qubit in the X, Y, and Z bases. These measurements yield single-qubit reduced density matrices whose expectation values become the new features. A standard classical kernel method — here a support vector classifier — then does the actual classification. This design sidesteps the training instabilities and noise sensitivity that plague variational quantum algorithms on today’s pre-fault-tolerant hardware.

The experiments ran on IBM Heron R2 superconducting processors, ibm_marrakesh and ibm_kingston, devices with fixed-frequency transmon qubits linked by tunable couplers and median two-qubit gate errors around 0.3 percent. Each expectation value was estimated with 10,000 shots distributed across 64 Pauli-twirled circuit instances, with readout error mitigated using the TREX technique and qubit layouts selected for best performance. The full run on ibm_marrakesh consumed about 100 minutes of quantum processor time, roughly 22 seconds of quantum processing per data point. Before committing to hardware, the team applied the geometric framework of Huang and colleagues to check whether the dataset even held promise for quantum advantage: the geometric separation g of 15.78 was comparable to the square root of the training set size, 13.12, and the quantum model complexity of 1.53 was well below the classical value of 6.09 — conditions consistent with a potential predictive edge.

On hardware, the ZZ embedding with a pi/2 rotation angle and eight feature-map repetitions delivered the best results, reaching median and maximum F1 scores of 0.75 and 0.81 across ten random 70/30 train-test splits, compared with 0.73 and 0.77 for an optimized support vector machine on the original features. Performance dropped at twelve repetitions, consistent with accumulated hardware noise in deeper circuits. The Heisenberg embedding never beat the classical baseline, its deeper circuits falling victim to noise. Reordering features so that correlated ones sat near each other on the chip changed nothing meaningful, suggesting the circuit depth already generated sufficient entanglement. The team also benchmarked against approximate classical simulations — operator backpropagation and matrix product state methods — which were six to ten times slower and still slightly less accurate than the quantum processor results.

The most striking result came from validation. On an independent cohort of 16 previously synthesized constructs — a severely imbalanced set with only one low-cytotoxicity sample — the PQK-based model achieved a median Matthews correlation coefficient of 0.415, versus minus 0.067 for the classical model, a statistically significant difference. The quantum-enhanced classifier caught the single minority-class sample in 8 of 10 splits, a recall of 0.8, while the classical SVM managed it in only 1 split. Precision remained low for both models, and the authors caution that the tiny, imbalanced validation set makes these findings preliminary rather than definitive proof of generalization.

Perhaps more intriguing than the headline metrics is where the quantum kernel helped most. The PQK-based classifier consistently predicted specific motif-position combinations that the classical model could not, and never the reverse: across all ten splits there was no case in which the original-data SVM significantly outperformed PQK on a given motif and position. The advantage grew as information thinned out. At the third motif position — the least sampled, most often empty or terminal — the number of source proteins benefiting from PQK rose to eight, compared with four at the first position. Motifs derived from LAT, CD40, and LAIR1 were predicted better by the quantum model at all three positions, and LAT-derived motifs are clinically relevant, having been linked to improved CAR T-cell responses in antigen-low acute lymphoblastic leukemia.

The authors frame the work as an early demonstration that hybrid quantum-classical learning can navigate underdetermined biological design spaces, and as one of the first large-scale applications of quantum machine learning to cell therapy optimization. The overall gains are modest, and the quantum hardware offered favorable runtime and predictive utility relative to the best approximate classical simulations rather than a dramatic leap. But the pattern — advantages concentrated exactly where data is sparsest — is the pattern quantum machine learning theorists have long predicted. As CAR T-cell applications expand beyond oncology and combinatorial receptor libraries keep outpacing experimental throughput, quantum-enhanced feature maps may become a practical tool for steering synthetic immune receptor design toward regions of the design space no classical model would think to explore.

Subject of Research: Hybrid quantum-classical kernel methods for predicting CAR T-cell cytotoxicity from combinatorial signaling domain designs

Article Title: Enhanced prediction of CAR T-cell cytotoxicity with quantum-kernel methods

Article References: Utro, F., Tolunay, M., Rhrissorrakrai, K., Gujarati, T. P., Shi, J., Capponi, S., Amico, M., Earnest-Noble, N., & Parida, L. (2026). Enhanced prediction of CAR T-cell cytotoxicity with quantum-kernel methods. Quantum Machine Intelligence, 8(2), Article 105. https://doi.org/10.1007/s42484-026-00446-w

Image Credits: AI Generated

DOI: 10.1007/s42484-026-00446-w

Keywords: CAR T-cell therapy, quantum machine learning, projected quantum kernel, chimeric antigen receptor, support vector machine, synthetic immunology, IBM Heron QPU, cytotoxicity prediction, combinatorial library, cell therapy, quantum computing, signaling motifs