[Paper Review] Non-Hemolytic Peptide Classification Using A Quantum Support Vector Machine

Shengxin Zhuang, John J. Tanner+8 more|arXiv (Cornell University)|Feb 6, 2024

Machine Learning in BioinformaticsBiochemistry, Genetics and Molecular Biology3 citations

TL;DR

This study applies a quantum support vector machine (QSVM) to classify peptides as hemolytic or non-hemolytic using three experimental peptide datasets. The QSVM outperforms classical SVMs and the best prior classical machine learning results, achieving up to 3.1% higher accuracy on HemoPI-1 and demonstrating a verifiable quantum advantage in computational biology.

ABSTRACT

Quantum machine learning (QML) is one of the most promising applications of quantum computation. However, it is still unclear whether quantum advantages exist when the data is of a classical nature and the search for practical, real-world applications of QML remains active. In this work, we apply the well-studied quantum support vector machine (QSVM), a powerful QML model, to a binary classification task which classifies peptides as either hemolytic or non-hemolytic. Using three peptide datasets, we apply and contrast the performance of the QSVM, numerous classical SVMs, and the best published results on the same peptide classification task, out of which the QSVM performs best. The contributions of this work include (i) the first application of the QSVM to this specific peptide classification task, (ii) an explicit demonstration of QSVMs outperforming the best published results attained with classical machine learning models on this classification task and (iii) empirical results showing that the QSVM is capable of outperforming many (and possibly all) classical SVMs on this classification task. This foundational work paves the way to verifiable quantum advantages in the field of computational biology and facilitates safer therapeutic development.

Motivation & Objective

• To explore whether quantum machine learning (QML) models like QSVM can outperform classical machine learning on real-world biological classification tasks.
• To evaluate the performance of QSVM on classifying hemolytic vs. non-hemolytic peptides using three wet-lab-verified datasets.
• To determine if quantum kernels can achieve better generalization than classical kernels on this specific bioinformatics problem.
• To establish a baseline for quantum kernel design using random Pauli string feature maps on classical data.
• To assess the potential for verifiable quantum advantage in computational biology using near-term quantum algorithms.

Proposed method

• The QSVM uses a data encoding unitary that maps 40 physicochemical peptide descriptors into a 6-qubit quantum state using random Pauli string feature maps.
• Quantum kernel methods are employed to compute inner products in a high-dimensional Hilbert space, enabling non-linear classification without explicit state preparation.
• The quantum kernel is evaluated via numerical simulations on a classical computer, emulating the quantum feature map and kernel computation.
• Classical SVMs with RBF, polynomial, and linear kernels are trained and compared using the same 40-descriptor input representation.
• Hyperparameters (e.g., $t$, $s$) for the quantum kernel are optimized using grid search over training/validation splits.
• Performance is evaluated using standard metrics: accuracy, precision, recall, and F1-score on test sets across three datasets: HemoPI-1, HemoPI-2, and HemoPI-3.

Experimental results

Research questions

• RQ1 Can a quantum support vector machine outperform classical SVMs on the non-hemolytic peptide classification task?
• RQ2 Does the QSVM achieve higher accuracy than the best published classical machine learning results on the same peptide datasets?
• RQ3 How robust is the QSVM performance to variations in the choice of quantum feature map (e.g., random Pauli strings)?
• RQ4 Can quantum kernels with random Pauli encodings match or exceed classical kernel performance on classical data?
• RQ5 What is the potential for verifiable quantum advantage in computational biology using near-term quantum algorithms?

Key findings

• The QSVM achieved a test accuracy of 96.8% on HemoPI-2 and 94.5% on HemoPI-3, outperforming classical SVMs with RBF and polynomial kernels.
• On HemoPI-1, the QSVM achieved 96.8% accuracy, surpassing the best classical kernel (RBF) by 3.1 percentage points.
• The QSVM improved upon the previous state-of-the-art results by Plisson et al. by 3.1%, 2.0%, and 2.8% on HemoPI-1, HemoPI-2, and HemoPI-3 respectively.
• Random Pauli string feature maps yielded consistent performance across 30 sampled configurations, with testing accuracies matching or exceeding classical kernel results.
• The QSVM demonstrated a 2% and 0.9% accuracy gain over classical SVMs on HemoPI-2 and HemoPI-3, respectively, despite using identical input descriptors.
• The results suggest that quantum kernels can provide a practical advantage even when classical data is used, supporting the feasibility of verifiable quantum advantage in bioinformatics.