Robust Quantum Reservoir Computing for Molecular Property Prediction Preprint: JCIM ACS/123-QED

Daniel Beaulieu, Milan Kornjača, Zoran Krunic, Michael Stivaktakis, Thomas Ehmer, Sheng-Tao Wang, Anh Pham · 2024

Read Robust Quantum Reservoir Computing for Molecular Property Prediction Preprint: JCIM ACS/123-QED by Daniel Beaulieu, Milan Kornjača, Zoran Krunic, Michael Stivaktakis, Thomas Ehmer, Sheng-Tao Wang, Anh Pham (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-06758.txt.

Introduction

Daniel Beaulieu and co-authors test a quantum learning method for forecasting what drug-like molecules will do, in work from 2024. Standard quantum methods need slope checks on quantum gear, which can fade and stall training. The reservoir method skips that step, since the quantum part is not tuned directly.

The team fed molecule traits into the quantum reservoir and read the results with a simple layer. As data sets shrank, the quantum method held up better than standard tools, a trait that suits drug work where data are scarce. A mapping check also showed the quantum-touched data formed clearer low-dimension patterns. The authors see promise for small, costly data sets in drug hunting.

Catalog record

CREATORDaniel Beaulieu, Milan Kornjača, Zoran Krunic, Michael Stivaktakis, Thomas Ehmer, Sheng-Tao Wang, Anh Pham
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE35 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-06758

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Robust Quantum Reservoir Computing for Molecular Property Prediction Preprint: JCIM ACS/123-QED Daniel Beaulieu, Milan Kornjača, Zoran Krunic, Michael Stivaktakis, Thomas Ehmer, Sheng-Tao Wang, Anh Pham Abstract Machine learning has been increasingly utilized in the field of biomedical research to accelerate the drug discovery process. In recent years, the emergence of quantum computing has been followed by extensive exploration of quantum machine learning algorithms. Quantum variational machine learning algorithms are currently the most prevalent but face issues with trainability due to vanishing gradients. An emerging alternative is the quantum reservoir computing (QRC) approach, in which the quantum algorithm does not require gradient evaluation on quantum hardware. Motivated by the potential advantages of the QRC method, we apply it to predict the biological activity of potential drug molecules based on molecular descriptors. We observe more robust QRC performance as the size of the dataset decreases, compared to standard classical models, a quality of potential interest for pharmaceutical datasets of limited size. In addition, we leverage the uniform manifold approximation and projection technique to analyze structural changes as classical features are transformed through quantum dynamics and find that quantum reservoir embeddings appear to be more interpretable in lower dimensions. I Introduction Predicting drug activities is an important and difficult task in pharmaceutical discovery. This task has traditionally been performed using a labor-intensive process of repeated trials and errors within a laboratory setting Liu et al. 2018 ; Pantaleão et al. 2022 . Recently, machine learning techniques have been employed to accelerate the discovery process Vamathevan et al. 2019 ; Deng et al. 2023 ; Zeng et al. 2022 . Within a machine learning workflow, molecules can be mapped to features using different molecular descriptors, such as physiological, biochemical properties or molecular fingerprints, which can then be used to predict a target variable such as therapeutic efficacy Team a ; Comesana et al. 2022 ; Cereto-Massagué et al. 2015 ; Mellor et al. 2019 . However, an immediate challenge is how to extract the relevant molecular features to improve the overall performance of the model, since the structure-property relationship for molecules is complex Liu et al. 2018 ; Vamathevan et al. 2019 ; Vu et al. 2023 . Recently, the advent of quantum computing has given rise to a novel machine learning paradigm Caro et al. 2022 ; Schuld et al. 2014 ; Carleo et al. 2019 ; Biamonte et al. 2017 . Various quantum machine learning (QML) versions of classical ML models such as generative, support vector machines (SVMs), and neural networks have been explored Lloyd et al. 2020 ; Schuld 2021 ; Killoran et al. 2019 ; Huembeli et al. 2022 ; Lloyd et al. 2020 . Quantum kernel methods, in particular, have shown potential advantage on synthetic datasets of quantum origin Havlíček et al. 2019 ; Schuld and Killoran 2019 ; Liu et al. 2021 ; Huang et al. 2021 ; Kornjača et al. 2024 . As a result, there are early attempts to apply QML algorithms for predicting the properties of molecules and pharmacological activity Batra et al. 2021 ; Bhatia et al. 2023 ; Vakili et al. 2024 . In this manuscript, we explore the performance of quantum reservoir computing (QRC) for molecular property prediction. Reservoir computing Jaeger and Haas 2004 addresses the potential drawback of widely employed variational QML models. There, the need to estimate gradients on quantum hardware leads to fundamental trainability issues due to entanglement and noise-induced barren plateaus McClean et al. 2018 ; Marrero et al. 2021 ; Larocca et al. 2024 . QRC has been proposed as an alternative approach, in which the gradient estimation can be fully off-loaded to classical post-processing Fujii and Nakajima 2017 ; Martínez-Peña et al. 2021 ; Bravo et al. 2022 ; Senanian et al. 2023 ; Kornjača et al. 2024 . The QRC algorithm we employ in this study leverages entangled quantum dynamics of neutral atom arrays, a platform that has already shown promise for quantum simulation, optimization, and machine learning Scholl et al. 2021 ; Ebadi et al. 2021 ; Semeghini et al. 2021 ; Ebadi et al. 2022 ; Wurtz et al. 2023 ; Balewski et al. 2024 ; Bluvstein et al. 2023 ; Kornjača et al. 2024 . We applied the QRC algorithm to the dataset within the Merck Molecular Activity challenge (MMACD) Kulkarni 2012 , with the task of predicting pharmacological activity based on molecular fingerprints 1 1 1 The Competition was sponsored in 2012 by Merck & Co., Inc, located today at 26 E Lincoln Ave, Rahway, NJ 07065, US, USA. Merck KGaA and Merck & Co. are separate entities. Merck KGaA is headquartered in Germany, operates globally as EMD in North America, focusing on pharmaceuticals, life sciences, and performance materials, whereas Merck & Co., headquartered in the USA, operates as MSD outside North America, primarily focusing on pharmaceuticals, vaccines, and animal health. . We systematically report the performance of the classical and QRC models across different dataset sample sizes and feature choices generated by the quantum reservoir dynamics. We then compare classical features and embedded quantum reservoir features by employing the uniform manifold approximation and projection (UMAP) technique. Our results reveal a slower performance decay with decreasing training dataset size when QRC embeddings are used compared to classical features. The UMAP analysis reveals more structure for QRC-embedded data when projected to low dimensional spaces. Together, the results indicate that the QRC approach has the potential to build a more robust and interpretable model with less training data, which could be relevant for health-related applications. Figure 1 : Machine learning modeling workflow and analysis of the different embeddings . (1) A schematic workflow used to determine the optimal classical model, the feature selection process with the SHAP method to enable QRC simulation, and the subroutine to perform dataset subsampling. (2) An outline of the quantum reservoir computing algorithm. Both two-body observable embeddings and one-body observable embeddings are estimated and utilized to train the classical models. (3) UMAP analysis is used to interpret the difference between classical and quantum features. II Methods II.1 Molecular dataset To explore how quantum reservoir computing can be utilized to make predictions of molecular activities, we applied the QRC technique to predict the activity (ACT) values based on molecular descriptors as presented in the MMACD Kulkarni 2012 . The dataset contains 15 different biological activity prediction problems, each represented by a unique set of molecular descriptors. The MMACD has been used extensively in the machine learning community Ramsundar et al. 2015 ; Goh et al. 2017 ; Hechtlinger et al. 2017 . In this manuscript, we perform general modeling and the QRC pipeline to five of the smallest MMACD datasets, 4, 5, 9, 14, and 15, out of the entirety of the MMACD. For our UMAP projection and more detailed analysis, we focus our study on one of the 15 ACT datasets, MMACD 4. MMACD 4 was selected as the ideal candidate for probing the robustness of QRC due to having the smallest sample size. In addition, we chose MMACD 14 to perform our 25 subsample QRC modeling on analysis so as to reduce potential dataset bias for our 25 subsample results. II.2 Research Modeling Workflow Overview As shown

Read the complete text