Quantum Data Proposes DYRK1A Compounds Differing from Tabular Models

For the DYRK1A kinase, a new study reveals that scoring potential drug candidates with a quantum processor yields a distinctly different set of compounds for testing than traditional methods using tabular molecular descriptors. Researchers evaluated the same molecules using both approaches, a vector measured on a quantum processor and a standard representation, and found each model proposed unique compounds for experimental evaluation.

The work reports on which compounds each model identifies as active, and whether those compounds have been previously studied, noting that “what changes is which model marks them active.” This divergence means each approach would initiate a different chemical series for laboratory testing, highlighting the cost of predictive disagreement in early drug discovery.

Quantum Data Expands DYRK1A Chemical Series

The choice of which molecules to test experimentally dictates the progression of drug discovery, as compounds not assayed never enter the confirmation or analog series stages. This principle guided a recent investigation into the DYRK1A kinase, where researchers assessed the same set of molecules using both a traditional method, a tabular representation of molecular descriptors, and a quantum processor-derived vector, both fed into a support vector machine.

The study revealed a key divergence. The two models did not propose the same compounds for further evaluation. This discrepancy isn’t merely a matter of ranking; the quantum processor-based model identified already-labeled active compounds that the standard tabular representation overlooked, and conversely, the tabular representation highlighted actives the quantum model missed.

“That disagreement leaves hits out of the plate and chemical series unopened,” the study notes, highlighting the practical implications of differing predictions. The researchers tested the robustness of this finding by repeating the analysis with different data partitions, and the disagreement persisted, suggesting it isn’t simply a result of random variation or hardware noise.

Entangled Manifold Embedding Generates QPU-Derived Vectors

The technique of entangled manifold embedding (EME) defines a specific relationship between classical molecular descriptors and a quantum state, creating an injective map from descriptor space into a submanifold characterized by a controlled entanglement density. This approach, detailed in a recent investigation of the DYRK1A kinase, allows for the generation of molecular representations using a 71-qubit EME circuit, evaluated both on a matrix product state simulator and the IBM Quantum processor ibm_boston.

The EME protocol’s ability to define a bijection between real vectors and a manifold of Hilbert space with fixed entanglement is central to its function. By adjusting the entanglement density D, the mapping can be tuned, with D equaling one reducing the process to angle encoding.

This flexibility allows for exploration of the quantum state space in a way that circumvents limitations imposed by conventional data encoding strategies on noisy intermediate-scale quantum (NISQ) devices. The work reports that the choice of representation directly impacts which compounds enter confirmation studies and subsequent analog series, potentially opening avenues for novel chemical investigations.

QSAR Prioritizes Compounds for Experimental Evaluation

The choice of compounds for experimental evaluation diverges when using a quantum processor alongside traditional methods for identifying potential DYRK1A kinase inhibitors, according to a recent study. While both approaches operate within the same chemical space, the models prioritize different molecules for further investigation, impacting which chemical series are explored. This outcome highlights a critical consideration in drug discovery: the “cost of predictive disagreement.” The researchers emphasize that even if two models achieve similar overall performance metrics, they can still propose entirely different sets of compounds for testing, potentially leaving promising leads unexplored.

This means a different “chemical series” would be sent for experimental evaluation depending on the model used. The study explicitly states they are not claiming a quantum advantage, but rather focusing on the practical consequences of differing predictions.

DYRK1A Kinase as Target in Alzheimer’s Research

The selection of compounds for laboratory testing directly influences which potential drug candidates advance in Alzheimer’s research, as assay capacity inherently limits the scope of investigation. A study focused on the DYRK1A kinase demonstrated that scoring molecules with a traditional quantitative structure-activity relationship (QSAR) model, and a quantum processor-derived model, yielded differing proposals for experimental evaluation.

The research team evaluated 354 compounds, each labeled as either active or inactive against DYRK1A based on an IC50 value of 1 μM or less, using both the standard QSAR approach and a vector representation generated via a quantum processor. Analysis revealed that the quantum-derived model identified compounds missed by the traditional method, and vice versa, highlighting a potential blind spot in conventional screening. This discrepancy has practical consequences beyond predictive accuracy; a compound not assayed does not enter confirmation studies, effectively closing off an entire chemical series from further investigation.

Classical QSAR Uses Tabular Molecular Descriptors

The study of the DYRK1A kinase employed a standard approach to virtual screening using tables of molecular descriptors alongside a quantum processor-derived vector representation, revealing a divergence in proposed compounds for experimental evaluation. While both methods used the same RBF-SVM algorithm, the resulting lists of candidate molecules differed, impacting which chemical series would proceed to further investigation. The tabular representation of molecular descriptors consisted of 1080 columns, computed by the Mordred software and standardized across both training and test sets, without dimensionality reduction or structural recomputation.

This classical QSAR approach provides a baseline for comparison with the quantum-derived model, yet the two methods still propose distinct sets of compounds for testing. The accuracy of both models was assessed using a nearly balanced dataset, with active-class rates of 0.561 in the training set and 0.556 in the test set.

Quantum-Enhanced Vectors Differ in Hit Identification

The divergence between quantum-derived and traditional molecular representations extends to specific compounds proposed for experimental evaluation, with each model suggesting distinct chemical series for investigation. The process of measuring the quantum-derived vectors involved loading each sample into an EME block, followed by single-qubit X, Y, and Z readouts, ultimately generating 213-component vectors. These vectors, along with the tabular data, were then fed into the RBF-SVM classifier. The researchers emphasize that this discrepancy isn’t about one model being universally superior, but about the instances where they disagree, and the resulting impact on research direction.

Predictive Discrepancy Impacts Chemical Series Selection

This study demonstrates that even when two models achieve similar overall performance, they can diverge in identifying specific active compounds for the DYRK1A kinase, impacting which molecular structures progress further in the drug discovery process. This divergence isn’t simply a matter of ranking compounds differently; it’s about identifying entirely different sets of potential hits. The work clarifies that the models operate within a shared chemical space, but their decision boundaries differ, leading to these discrepancies in hit identification.

The researchers found no statistically significant difference in the overall ranking capacity of the two models, with a DeLong p-value of 0.117, supporting the idea that the disagreement stems from differing patterns of prediction rather than overall accuracy. This work focuses on the instances where the models disagree, specifically the wells and chemical series that are either gained or lost due to these predictive variations, and the potential impact on the drug discovery funnel.

Kernel SVM Classifies with Quantum-Derived States

The kernel SVM employed in this study reveals that even with identical metrics, differing molecular representations can propose distinct compounds for experimental evaluation. Researchers compared a classical descriptor representation with data obtained from a quantum processor, both fed into a shared RBF-SVM algorithm to assess potential DYRK1A kinase inhibitors. This comparison wasn’t focused on ranking performance, but on identifying outright disagreement in proposed compounds, a divergence that dictates which chemical series are pursued further.

The core question, as the team states, is whether the two representations propose different compounds for testing, given the same DYRK1A test set. The quantum approach uses an entangled manifold embedding with a local projected readout, a technique designed to mitigate the challenges of barren plateaus and hardware noise common in quantum machine learning.

This method projects the quantum system onto a classical space by extracting one-body reduced density matrices, allowing for a more robust kernel function. The kernel function itself evaluates inner products without explicitly computing the non-linear mapping, which is an important step in the QSVM process. The study’s methodology centers on the Lagrange dual formulation, maximizing alpha values subject to constraints, ultimately determining the decision function of the RBF-SVM.

A positive decision function signifies a proposed compound, meaning the signed distance to the hyperplane in kernel space is positive. The researchers investigated whether the observed disagreement between the models stemmed from the classifier family itself, testing whether alternative algorithms would highlight the same molecules.

More like this

Quantum Machine LearningIBM Quantum Flex and Premium plans add Rimay access

Machine LearningKipu Quantum adds Rimay machine learning tool to IBM Quantum Functions

Quantum HardwareIQM tests quantum machine learning on its Emerald and Garnet chips

CybersecurityQuantum data from ORCA Computing’s PT-2 chip refines malware detection