PaperlayerView exact revision on arXiv

Source-linked AI summary

Quantum Generative Models for Small Molecule Drug Discovery

Junde Li, Rasit Topaloglu, Swaroop Ghosh

arXiv:2101.03438v1cs.ETcs.LGquant-ph

TL;DR

Classical generative models cannot explore all regions of chemical space, while a full quantum GAN for QM9-like molecules requires more than 90 qubits. The paper proposes QGAN-HG, a qubit-efficient hybrid quantum generator with a classical discriminator; it reports molecule-distribution learning at par with the original GAN and faster training with patched circuits.

  • Problem

    Classical generative models cannot generate all possible distributions, and a full quantum GAN requires more than 90 qubits for QM9-like molecules.

    Source: §I · p. 1

  • Method

    QGAN-HG combines a parameterized quantum circuit with a classical neural network generator and discriminator, and the paper also evaluates patched circuits.

    Source: §I · p. 2, A. Quantum GAN Flavors · p. 3, B. Quantum GAN with Hybrid Generator · p. 3

  • Results

    15 extra quantum gate parameters let QGAN-HG learn molecule distributions at par with the original GAN, while patched quantum GANs considerably accelerate training.

    Source: §VI · p. 6

  • Takeaways & Limitations

    The model is intended to generate a library of novel, valid molecules for screening by a downstream docking engine.

    Source: §I · p. 2

    Real IBM quantum computers were used for QGAN-HG inference, not training, because per-iteration waiting time was long.

    Source: §I · p. 2

Abstract

from arXiv · show

Existing drug discovery pipelines take 5-10 years and cost billions of dollars. Computational approaches aim to sample from regions of the whole molecular and solid-state compounds called chemical space which could be on the order of 1060 . Deep generative models can model the underlying probability distribution of both the physical structures and property of drugs and relate them nonlinearly. By exploiting patterns in massive datasets, these models can distill salient features that characterize the molecules. Generative Adversarial Networks (GANs) discover drug candidates by generating molecular structures that obey chemical and physical properties and show affinity towards binding with the receptor for a target disease. However, classical GANs cannot explore certain regions of the chemical space and suffer from curse-of-dimensionality. A full quantum GAN may require more than 90 qubits even to generate QM9-like small molecules. We propose a qubit-efficient quantum GAN with a hybrid generator (QGAN-HG) to learn richer representation of molecules via searching exponentially large chemical space with few qubits more efficiently than classical GAN. The QGANHG model is composed of a hybrid quantum generator that supports various number of qubits and quantum circuit layers, and, a classical discriminator. QGAN-HG with only 14.93% retained parameters can learn molecular distribution as efficiently as classical counterpart. The QGAN-HG variation with patched circuits considerably accelerates our standard QGANHG training process and avoids potential gradient vanishing issue of deep neural networks. Code is available on GitHub https://github.com/jundeli/quantum-gan.

I. INTRODUCTION

Drug discovery searches a vast chemical space, but classical generative models cannot explore every region and full quantum GANs face severe qubit demands. The paper proposes a qubit-efficient hybrid quantum GAN for molecular generation and examines its patched-circuit variant.

I. INTRODUCTION

  • Drug development spans target discovery, molecular design, preclinical studies, and clinical trials, making marketable drugs expensive and time-consuming to develop [1].Small molecules are a major class of approved drugs, and de novo design navigates the space of organic molecules,.Source: §I · p. 1
  • Classical generative models cannot generate all possible distributions, in part because adding molecular fragments entails exponentially many choices.Graph-based molecular generation can automate navigation toward chemical regions abundant in desired molecules.Source: §I · p. 1
  • Quantum GANs are presented as offering few-qubit searches over exponentially large chemical spaces, while existing demonstrations remain constrained to simple or dimension-reduced tasks [17], [18].The cited examples address two digits or reduce MNIST images to two dimensions.Source: §I · p. 1
  • A full quantum GAN would require more than 90 qubits to reconstruct QM9-like molecules, exceeding the capacity of commercially available quantum computers described in the paper.The qubit estimate follows from the dataset’s five atom and five bond types.Source: §I · p. 1
  • The authors propose a qubit-efficient quantum GAN with a hybrid generator and classical discriminator to learn molecule distributions, and compare patched circuits with a single large circuit.The model is based on MolGAN and aims to generate novel, valid molecules for downstream docking.Source: §I · p. 2
  • The paper presents quantum machine learning for drug discovery as a first and tests patched QGAN-HG, molecular-property outcomes, and inference on IBM quantum computers.The IBM hardware run covers inference, not training, because per-iteration waiting time is long.Source: §I · p. 2

A. Computational Drug Discovery

The section describes GAN-based molecular generation as a competition between a generator and discriminator, with molecular graphs and chemical evaluations defining the task.

B. Computational Drug Discovery

  • The generator maps random noise to synthetic chemical data, while the discriminator predicts whether a sample comes from real data or the generator.The discriminator maximizes correct classification, and the generator minimizes the fake-sample loss.Source: A. Computational Drug Discovery · p. 2
  • Molecules are represented as graphs whose nodes encode atoms and edges encode bonds; a generated molecule is valid when chemically stable and highly affine to receptor binding sites.The model is trained until its synthetic chemical distribution approaches real chemical space.Source: A. Computational Drug Discovery · p. 2
  • Fréchet distance and RDKit chemical-property scores are used to assess GAN output quality.RDKit evaluates chemical properties of generated molecules.Source: A. Computational Drug Discovery · p. 2

B. Quantum Machine Learning

Quantum neural networks use parameter-dependent unitary transformations, with their architecture shaped by qubits, circuit layers, and gates. The section introduces circuit gates, measurement, and noise as relevant concepts.

B. Quantum Machine Learning

  • A quantum neural network is a variational quantum circuit trained through parameter-dependent unitary transformations, with architecture determined by qubit count, circuit layers, and applied gates.Quantum systems are described as producing patterns classical computers cannot efficiently produce.Source: B. Quantum Machine Learning · p. 2
  • Figure 2 presents a sample molecular graph from QM9 alongside its corresponding atom vector A and bond matrix B, and also displays all quantum gates used in the study.Source: Figure 2 · p. 3
  • Single-qubit rotation gates Ry and Rz and the two-qubit CNOT gate modulate qubit states; measurement yields expectation values after a number of shots.Gate pulses rotate states along axes in the Bloch sphere.Source: B. Quantum Machine Learning · p. 3
  • The paper explores quantum noise during QGAN-HG inference on IBM hardware rather than mitigating it, noting that systematic noise may benefit neural-network generalization.Listed noise sources include relaxation, dephasing, gate errors, and readout error.Source: B. Quantum Machine Learning · p. 3

III. QUANTUM GENERATIVE ADVERSARIAL NETWORKS

Full quantum generators are impractical for QM9-like molecules because they require more than 90 qubits, motivating a hybrid generator with a classical discriminator. QGAN-HG converts quantum-circuit measurements into features for classical molecular-graph generation, while patched circuits trade expressive capacity for reduced quantum resources.

A. Quantum GAN Flavors

  • A full quantum generator for QM9-like molecules requires more than 90 qubits, so the proposed approach pairs a quantum generator with a classical discriminator.The hybrid design is intended to learn molecule distributions with fewer qubits.Source: A. Quantum GAN Flavors · p. 3

B. Quantum GAN with Hybrid Generator

  • QGAN-HG uses a parameterized quantum circuit to produce a feature vector that a classical neural network maps to atom vectors and bond matrices.The circuit includes initialization, repeatable parameterized layers, and measurement; the measured expectation values form the features.Source: B. Quantum GAN with Hybrid Generator · p. 3, Figure 3 · p. 3
  • The classical generator stage uses separate atom and bond output layers and Gumbel-Softmax reparameterization to produce discrete molecular graphs.The authors crop 85.07% and 98.03% of the original MolGAN generator parameters in the two reduced configurations.Source: B. Quantum GAN with Hybrid Generator · p. 3
  • Patched QGAN-HG concatenates quantum sub-circuits, reducing quantum-resource requirements and enabling more efficient simulation, but reducing the quantum state dimension and expressive power.The paper also describes the classical discriminator and separate, alternating optimizer updates for discriminator versus generator parameters.Source: B. Quantum GAN with Hybrid Generator · pp. 3–4

A. Dataset and Metrics

Experiments use QM9 and evaluate generated-molecule distribution similarity with Fréchet distance, alongside drug-property metrics. The authors use the atom-and-bond distance with batches of 128 molecules for subsequent experiments.

A. Dataset and Metrics

  • QM9 contains 133,885 molecules with up to 9 heavy atoms, drawn from carbon, nitrogen, oxygen, and fluorine.Source: A. Dataset and Metrics · p. 4
  • Fréchet distance compares generated and real molecule distributions using 50 sampled batch-points from each distribution.The study evaluates atom-and-bond (A+B) and bond-only (B) distances at batch sizes of 128 and 512 molecules.Source: A. Dataset and Metrics · p. 4
  • The A+B distance means are 12.3342 and 12.6387, with variations of 0.7057 and 0.7849, for batch sizes of 128 and 512, respectively.Subsequent experiments use A+B distance with batches of 128 molecules.Source: A. Dataset and Metrics · p. 4
  • Drug-quality evaluation includes QED for druglikeness, logP for solubility, and SA for synthetic accessibility.The properties are measured using RDKit.Source: A. Dataset and Metrics · p. 4

B. Implementation Details

Training uses a noiseless simulator rather than a real quantum machine, while inference compares simulator and IBM Q16 Melbourne execution. The paper also specifies its training setup and reports comparisons across GAN variants.

B. Implementation Details

  • Training uses a noiseless simulator; the authors do not train on real quantum machines because of long waiting times on IBM Q machines.The simulator supports customized noise settings, but noiseless conditions are used in this paper.Source: B. Implementation Details · p. 4
  • Training uses Adam, a single RTX 2080 Ti for the classical component, and PennyLane’s default qubit plugin for the quantum stage.The minibatch is 128, the initial learning rate is 0.0001 for both generator and discriminator, and training is set to 5,000 epochs with Fréchet-distance early stopping if collapse occurs.Source: B. Implementation Details · p. 4
  • Inference runs the hybrid generator on both the PennyLane simulator and IBM Q16 Melbourne, and GAN variants are compared using 1,000 generated molecules.The training-comparison figure includes Fréchet-distance curves, lowest-distance epochs, and elapsed training times across GAN flavors.Source: B. Implementation Details · pp. 4–5, Figure 5 · p. 5

V. EVALUATION RESULTS

QGAN-HG retains strong learning behavior under substantial generator parameter reduction, including a configuration with only 1.97% of MolGAN’s parameters. Increasing circuit layers or qubits does not substantially accelerate learning in the reported variants.

A. QGAN-HG Results

  • At 14.93% of original MolGAN generator parameters, QGAN-HG and MolGAN reach slightly lower Fréchet distances than moderately reduced MolGAN.The 14.93% configuration is compared with original and moderately reduced MolGAN; the comparison reaches a reasonably good training point within 5,000 epochs.Source: A. QGAN-HG Results · p. 5
  • At 1.97% of original MolGAN generator parameters, highly reduced MolGAN can hardly be learned, while QGAN-HG still shows a sharp downward learning curve.The QGAN-HG circuit uses only 15 gate parameters.Source: A. QGAN-HG Results · p. 5
  • GAN model collapse is identified when Fréchet distance begins increasing after a training point.Early stopping uses the lowest reached Fréchet distance.Source: A. QGAN-HG Results · p. 5
  • Adding circuit layers or qubits does not substantially accelerate learning relative to QGAN-HG in the reported comparison.The tested variants include L = 2 and N = 10, and the authors use early stopping at the lowest reached Fréchet distance to somewhat prevent training instability after collapse.Source: A. QGAN-HG Results · p. 5

B. P-QGAN-HG Results

Patched QGAN-HG variants achieved learning quality comparable to an integral circuit while using fewer gate parameters and reducing simulation time. Drug-property results were generally consistent with Fréchet-distance results, with reported differences between simulator and real-quantum runs.

B. P-QGAN-HG Results

  • P2-QGAN and P4-QGAN had learning quality comparable to an integral circuit, while smaller circuits and early convergence significantly reduced simulation time.P2-QGAN used two 4-qubit subcircuits; P4-QGAN used four 2-qubit subcircuits.Source: B. P-QGAN-HG Results · p. 5

C. Drug Properties

  • GAN training was evaluated by Fréchet distance, while generated-molecule quality was evaluated with property metrics reported for 1,000 molecules across all variations.Table I marks the best results in bold and uses “-” for metrics not successfully evaluated by RDKit.Source: C. Drug Properties · p. 5, Table I · p. 6
  • Drug-property results were generally consistent with Fréchet-distance results; simulator and real-quantum results were consistent except for higher unique and synthesizability scores on the real machine.Diversity and novelty scores were high across models, while synthetic accessibility was low relative to the benchmark.Source: C. Drug Properties · p. 6
  • For inference, each GAN variation used the sample with the lowest Fréchet distance within 5,000 epochs.Source: C. Drug Properties · p. 6

VI. CONCLUSION

The conclusion presents QGAN-HG as a parameter-efficient approach to learning molecule distributions. Its patched variants achieved comparable drug-property accuracy and accelerated training relative to the integral quantum circuit.

  • With 15 extra quantum gate parameters, QGAN-HG learned the molecular task at par with the original GAN, while a classical GAN with 85.07% reduced parameters could not properly learn the molecule distribution.Source: §VI · p. 6
  • QGAN-HG with 98.03% reduced parameters was significantly more efficient than the highly reduced classical counterpart.Source: §VI · p. 6
  • Patched QGAN achieved comparable drug-property accuracy with 12 extra quantum gate parameters and considerably accelerated the integral-circuit QGAN-HG.The authors describe it as a possible alternative to classical GAN because of less training time and avoidance of a possible gradient-vanishing problem in deep neural networks.Source: §VI · p. 6

Related papers

Loading 2101.03438v1…