TuringQ and Shanghai Jiao Tong University Designed Working Defensins and Xylanases With a Photonic Quantum Training Head: arXiv 2611.08434, 5 November 2026
How TuringQ and Shanghai Jiao Tong University put a photonic quantum head inside protein design training
Quentir Medicine Monitor
Evidence-based insights for quantum medicine.
A team from the Shanghai photonics company TuringQ and Shanghai Jiao Tong University has published a protein design model whose training borrows a photonic quantum circuit as a helper, and has taken the proteins it generated into the laboratory. The preprint, posted to arXiv on 5 November 2026, reports antifungal activity for three designed plant defensins and measured enzyme activity for designed xylanases and short-chain dehydrogenases.
The quantum component sits in one specific place: the training loop. A photonic quantum neural network, or a quantum version of a Kolmogorov-Arnold network, serves as an auxiliary head that predicts how soluble a sequence will be, and its prediction error helps shape the shared representation the design model learns from. Once training finishes, every sequence is generated by an ordinary classical process on ordinary computers.
The author list spans a company and a university physics department. Ke-Xin Tang, Zhen-Qiang Weng, Kai Yan, Yao-Yao Lu, Yu-Ze Zhu, Jun-Jie He, Jie-Cheng Yang, Xiang Zhao and Lin Yang are at TuringQ; Hang Song and Xian-Min Jin are at the School of Physics and Astronomy and the State Key Laboratory of Photonics and Communications at Shanghai Jiao Tong University, with further affiliations at the Chip Hub for Integrated Photonics Xplore in Wuxi, Atomology Co. and Hefei National Laboratory. The work is funded by the National Key R&D Program of China and the national Innovation Program for Quantum Science and Technology.
Protein design asks for an amino-acid sequence that folds the required way and survives contact with water and with a production host. The model the team calls DiffSequence-Q starts from a noisy sequence and repeatedly replaces amino acids until a clean sequence remains, a procedure known as multinomial diffusion. Two conditions ride along at every step: the required secondary structure for each individual residue, written in the nine-state DSSP vocabulary, and a single continuous number for the solubility the designer wants.
Solubility is where the quantum part enters. During joint training, a small auxiliary head reads the shared Transformer representation of a sequence and predicts a solubility score, and the error in that prediction is optimized together with the design objective. In the classical reference version the head is a multilayer perceptron. In the two quantum versions it is a photonic quantum neural network, whose optical modes carry the computation, or a quantum Kolmogorov-Arnold network, which replaces fixed network weights with learnable one-variable functions built from a Fourier series of a chosen order.
On the standard CATH v4.2 test split introduced by Ingraham and colleagues in 2019, the pretrained design model gave the highest secondary-structure recovery of the four models compared when the whole sequence is regenerated, ahead of the strongest baseline by 8.96 percentage points on the strict nine-state measure and 4.77 percentage points on the coarse three-state measure. Its three-state recovery was 60.57 percent with the whole sequence masked, 66.61 percent with the first half masked and 63.11 percent with the last half masked. The authors also report that one of the baselines, the multimodal protein language model DPLM-2, produced structures with the highest mean confidence score in all three settings, so the ranking depends on which quality is measured.