QuantSpin & DIVBIO Study Predicts GLP-1R Drug Activity with Quantum Circuits

A collaboration between Shenzhen-based quantum computing company SpinQ (QuantSpin Technology) and drug developer DIVBIO (DeepBio) is exploring how simulated quantum circuits can refine predictions of drug activity. The joint study incorporates quantum circuits into traditional neural networks to better estimate the concentration needed for a half maximal effect on the GLP-1R receptor, a key target for type 2 diabetes and weight management drugs. The work is published as a preprint on arXiv.

Quentir Medicine Monitor, a platform tracking quantum pharmaceutical advancements, has assessed the research as Technology Readiness Level 2, indicating it has moved beyond basic principle observation. Quentir’s official website reported on research conducted by QuantSpin, noting the study’s entry into international independent observation.

QuantSpin & DIVBIO Combine Quantum Circuits with Neural Networks for GLP-1R Efficacy Prediction

The research team achieved improved efficacy prediction for GLP-1R receptor activation by incorporating a five-qubit quantum circuit residual head into existing neural network architectures. This structure functions as a corrective term, layered onto predictions generated by conventional linear models, and was tested across multiple encoder types and data splitting strategies. The team’s work, a collaboration between QuantSpin Technology and DeepBio, demonstrates a performance advantage over both baseline models and the widely used Chemprop molecular property prediction tool, the company says.

A key element of the approach involves a classical encoder that transforms molecular structures into numerical representations. A multilayer perceptron processes molecular fingerprints while a graph neural network analyzes atomic bonds. This encoded data then feeds into both a linear prediction head, establishing a baseline and the newly developed quantum head.

The quantum head encodes the representation into five qubits, applies trainable rotation gates, and uses a CNOT gate to connect them, ultimately refining the initial prediction. The entire five-qubit circuit was simulated on traditional computing hardware using PyTorch and state vector simulation, enabling accurate calculations without requiring access to a quantum computer.

The quantum head, containing 36 trainable parameters, consistently outperformed a traditional neural network residual head with 69 parameters across all four encoder and data split settings tested. Results, reported as the average of five training runs, indicate a measurable improvement in predicting the concentration required for half-maximal effect in experimental testing.

ChEMBL v37 Dataset Details: GLP-1R, ADRB2, and MC4R Receptor Compounds

The ChEMBL version 37 dataset provided the foundation for this research, specifically data pertaining to the human GLP-1R receptor, using cAMP assays and reporting precise EC50 values as a measure of receptor activation. Researchers from SpinQ (QuantSpin Technology) and DIVBIO (DeepBio) curated a dataset of 1,052 compounds for this receptor, classifying them as either 472 small molecules or 580 peptides based on the presence of amide bonds.

This detailed molecular characterization allowed for a focused analysis of structure-activity relationships critical for predicting drug efficacy. Complementing the GLP-1R data, the team also incorporated datasets for the β2 adrenergic receptor (ADRB2) containing 829 compounds and the melanocortin-4 receptor (MC4R) with 767 compounds, serving as control sets for comparative analysis. The study’s methodology involved weighting and summing measurement results before adding them to baseline prediction values, a process designed to refine the accuracy of efficacy predictions.

While the study demonstrates potential gains in prediction accuracy through the integration of quantum-inspired circuits, the authors acknowledge that classical models with similar mathematical structures can achieve comparable results, suggesting that the performance improvements are not solely attributable to quantum effects.

They report that “the research results also show that classical models using the same mathematical structure may achieve most of the performance gains without using quantum hardware.” The team’s work, detailed in their paper numbered 2610. 13037 published October 9, 2026, remains distant from clinical application, as the model currently predicts in vitro efficacy under specific experimental conditions.

Five-Qubit Quantum Head Architecture & Simulation via PyTorch State Vector

The five-qubit quantum head employed in this research calculates a constrained quadratic function of the encoder output, a mathematical property the team rigorously demonstrated alongside their simulations. This finding allows for the construction of comparable classical control models, essential for establishing a baseline against which to measure any quantum advantage. Researchers constructed a 36-parameter classical model, mirroring the quantum head’s complexity and subject to real orthogonal constraints, then trained it using identical procedures.

This approach enabled a direct comparison of performance, moving beyond simply demonstrating a functional quantum circuit to assessing its practical benefit, according to QuantSpin Technology. Under random data splitting, the performance of the quantum head and the classical control model were virtually indistinguishable, indicating the quantum component did not offer a substantial improvement in that scenario. However, when employing a more challenging structural cluster split, grouping similar compounds to avoid artificially inflated accuracy, subtle differences emerged.

The quantum head reduced the mean absolute error of a model based on molecular fingerprints from 0.677 to 0.639 log units, and a model based on molecular graphs from 0.680 to 0.641 log units. While the advantage was small, the quantum head showed a 0.020 log unit improvement on the fingerprint encoder and 0.008 log units on the graph encoder under these stricter conditions.

The team’s simulations, performed using PyTorch via state vector simulation, circumvented the need for access to actual quantum computing hardware, enabling detailed analysis and validation, the company says. This allowed for accurate calculation of the five-qubit circuit’s behavior without the limitations of current quantum processor availability or fidelity.

Structural Cluster Splitting Reveals 6% Error Reduction with Quantum Residual Head

Structural cluster splitting, a rigorous validation technique, revealed a 6% reduction in error when paired with a quantum residual head for predicting GLP-1R receptor activity, a key target in type 2 diabetes and obesity therapeutics. This method groups structurally similar compounds together during model training and testing, ensuring a more challenging and realistic evaluation than random splitting, where similar molecules might artificially inflate performance metrics.

While the average error reduction was approximately 0.04 log units, the researchers also constructed a classical control model with an equivalent number of trainable parameters, 36, and the same real orthogonal constraints as the quantum head.

Under random splitting conditions, the performance of the quantum and classical models was comparable. These findings suggest that the observed improvement isn’t solely attributable to the quantum circuit’s architecture, but requires further investigation to determine the true potential of quantum machine learning in this domain.

The study also explored the impact of noise, simulating conditions expected in near-term quantum hardware through five progressively increasing levels. At lower noise levels, the quantum head largely maintained its performance advantage; however, at the highest noise level, its performance fell below that of the baseline model without a residual head. The team acknowledges that validation on independent chemical series is still needed to confirm the robustness of these findings, and that the research does not yet assess whether the improvement translates to better ranking of candidate compounds or ultimately influences compound selection for synthesis.

More like this

Quantum Machine LearningPennyLane’s flexibility suits quantum chemistry research

Quantum ApplicationsShionogi wins quantum drug-design prize with D-Wave’s help

Quantum Error CorrectionState-space model cuts decoder complexity, aiding quantum computers

Quantum AlgorithmsQuantum graph networks find larger cliques with less training