MolQAE: Quantum Autoencoder for Molecular Representation Learning

Yi Pan 1,* , Hanqi Jiang 1,* , Wei Ruan 1 , Dajiang Zhu 2 , Xiang Li 3 , Yohannes Abate 4 , Yingfeng Wang 5,† , Tianming Liu 1,† · 2025

Read MolQAE: Quantum Autoencoder for Molecular Representation Learning by Yi Pan 1,* , Hanqi Jiang 1,* , Wei Ruan 1 , Dajiang Zhu 2 , Xiang Li 3 , Yohannes Abate 4 , Yingfeng Wang 5,† , Tianming Liu 1,† (2025) 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-2505-01875.txt.

Introduction

Pan and co-authors build the first quantum autoencoder that works from a molecule's full structure. An autoencoder squeezes data into a compact code, then rebuilds it; the quantum version stores that code in quantum states. Their system, MolQAE, turns the standard text form of a molecule directly into quantum states using adjustable rotation gates.

Training balances two aims: rebuild the molecule faithfully, and keep the leftover trash states tidy. In tests, the system captured molecular traits well and rebuilt molecules with fidelity close to perfect, even after heavy shrinking of the code. The design fits today's noisy small quantum machines. The authors see it as a first quantum path into chemistry computing, with drug and materials work ahead.

Catalog record

CREATORYi Pan 1,* , Hanqi Jiang 1,* , Wei Ruan 1 , Dajiang Zhu 2 , Xiang Li 3 , Yohannes Abate 4 , Yingfeng Wang 5,† , Tianming Liu 1,†
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-01875

Browse the full library · About the Tower of Babel

Full text

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

MolQAE: Quantum Autoencoder for Molecular Representation Learning Thanks: These authors contributed equally to this work. Thanks: † Corresponding authors: Yingfeng Wang ([email protected]) and Tianming Liu ([email protected]). Yi Pan 1,* , Hanqi Jiang 1,* , Wei Ruan 1 , Dajiang Zhu 2 , Xiang Li 3 , Yohannes Abate 4 , Yingfeng Wang 5,† , Tianming Liu 1,† Abstract We introduce Quantum Molecular Autoencoder (MolQAE), the first quantum autoencoder to leverage the complete molecular structures. MolQAE uniquely maps SMILES strings directly to quantum states using parameterized rotation gates, preserving vital structural information. Its quantum encoder-decoder framework enables latent space compression and reconstruction. A dual-objective strategy optimizes fidelity and minimizes trash state deviation. Our evaluations demonstrate effective capture of molecular characteristics and a remarkable preservation of fidelity, approaching robust molecular reconstruction even with substantial dimensionality reduction. Our model establishes a quantum pathway in cheminformatics by being the first to process complete molecular structural information with a dedicated quantum architecture considering the Noisy Intermediate-Scale Quantum (NISQ)-era development and promising significant advances in drug and materials discovery. I Introduction Molecular representation learning has emerged as a fundamental challenge in computational chemistry and drug discovery, requiring methods that can encode complex molecular structures into compact, information-rich representations suitable for downstream machine learning tasks [ 1 ] . While classical approaches based on handcrafted descriptors and graph neural networks have achieved remarkable success [ 2 ] , they face inherent limitations when processing the exponentially large conformational and electronic state spaces that characterize molecular systems [ 3 ] . These limitations become particularly acute when attempting to capture quantum mechanical properties such as electron correlation and delocalization, which are crucial for accurate prediction of molecular behavior yet difficult to represent in classical feature spaces. The fundamental quantum nature of molecular systems suggests that quantum computing may offer natural advantages for molecular representation learning [ 4 ] . Quantum computers can efficiently represent superposition states and entanglement patterns that directly correspond to molecular electronic structures, potentially capturing chemical information that classical representations struggle to encode. Recent advances in quantum machine learning have demonstrated promising results in classification [ 5 ] , regression [ 6 ] , and generative modeling [ 7 ] , with quantum embeddings showing particular promise in capturing complex structural features that elude classical approaches [ 8 , 9 ] . However, the application of quantum computing to molecular representation learning remains largely unexplored, despite the natural alignment between quantum mechanical properties of molecules and the computational capabilities of quantum systems. Quantum autoencoders (QAEs) present a particularly compelling approach for molecular representation learning [ 10 , 11 ] . Originally developed for quantum data compression, QAEs utilize parameterized quantum circuits to encode input states into lower-dimensional latent representations while preserving essential information. Unlike classical autoencoders that operate on vector representations in Euclidean space, QAEs manipulate quantum states directly in Hilbert space, enabling them to leverage quantum phenomena such as superposition and entanglement for more efficient information encoding. This quantum mechanical foundation makes QAEs especially suitable for molecular systems, where electronic wavefunctions and molecular orbitals naturally exist as quantum states. Despite these theoretical advantages, significant challenges have prevented the practical application of quantum methods to molecular representation learning. The primary obstacle lies in effectively bridging classical molecular descriptions—such as SMILES strings or molecular graphs—with quantum state representations [ 12 ] . Previous attempts have typically relied on simplistic encoding schemes that discard crucial structural information, such as atomic connectivity and bond type. Additionally, training quantum circuits presents unique difficulties including measurement noise, limited circuit depth on near-term devices, and the presence of barren plateaus in the parameter landscape that can trap optimization in suboptimal regions [ 13 ] . These challenges necessitate novel approaches that can preserve molecular information through the classical-to-quantum encoding process while maintaining trainability on current quantum hardware. In this work, we present the Quantum Molecular Autoencoder (MolQAE), a framework that addresses these challenges through a carefully designed quantum circuit architecture for molecular representation learning. Our approach introduces three key innovations that distinguish it from previous quantum molecular methods. First, we develop a direct quantum encoding scheme that maps SMILES token sequences to parameterized quantum states, preserving sequential and structural information that frequency-based or amplitude encoding methods typically discard. Each token’s positional and chemical information is encoded through quantum gate parameters, enabling the representation of complex molecular patterns including rings, branches, and stereochemistry. Second, we design a hierarchical encoder-decoder architecture with specialized processing for latent and ancillary qubits, allowing flexible compression ratios while maintaining high reconstruction fidelity. Third, we introduce a dual-objective training strategy that balances reconstruction accuracy with effective ancillary state compression, addressing the unique optimization challenges of quantum circuits through hybrid quantum-classical methods. The key contributions of this paper are: • We introduce the first parameterized quantum state preparation method that directly maps SMILES token sequences to quantum states, preserving structural and sequential information critical for molecular characterization. • We design the first quantum autoencoder circuit with specialized latent and ancillary qubit processing, enabling flexible compression ratios while maintaining quantum coherence properties essential for molecular representation. • We develop a training approach that combines quantum circuit execution with classical parameter optimization, addressing barren plateaus and measurement noise through careful loss function design and gradient estimation techniques. • We demonstrate through systematic experiments that quantum autoencoders can achieve significant dimensionality reduction while preserving molecular properties, establishing a foundation for future quantum molecular applications. This work represents an important step toward practical quantum advantage in molecular sciences. By demonstrating that quantum circuits can effectively encode and compress molecular information, we establish a foundation for future applications in drug discovery, materials design, and chemical reaction prediction. As quantum hardware capabilities expand and circuit depths increase, frameworks like MolQAE may enable the exploration of larger molecular systems and more complex chemical phenomena that remain intractable for classical methods. II Related Work II-A Molecular Representation Learning Traditional molecular representation learning methods, such as graph neural networks (GNNs) or SMILES-based models, have achieved strong performance in molecular property prediction by encoding structural features of molecules [ 14 , 15 ] . However, these approaches often neglect important quantum-level properties and suffer from limitations like poor generalization in out-of-distribution scenarios [ 16 , 17 ] . For example, current GNNs struggle to incorporate multiscale chemical features or dynamic molecular conformations, and typically depend on large annotated datasets [ 18 , 19 ] . II-B Quantum Embedding Quantum embedding has emerged as a powerful framework for representing classical data in quantum-enhanced feature spaces. Recent studies have validated its effectiveness in real-world scenarios. Albrecht et al. [ 9 ] demonstrated that quantum feature maps implemented on neutral atom quantum processors can distinguish graph structures and perform molecular classification tasks with accuracy comparable to classical kernels. Ghosh et al. [ 8 ] proposed a quantum feature mapping strategy based on the dynamics of the XY spin model, which enables effective quantum reservoir computation with enhanced expressive capacity. II-C Quantum Autoencoder Quantum autoencoders have shown promise in compressing quantum information and dimensionality reduction for molecular dynamics [ 10 , 20 ] . While recent QAE explorations have tackled representations derived from parts of molecular information, such as 3D atomic coordinates (QVAE-Mole)

Read the complete text