PaperlayerView exact revision on arXiv

Source-linked AI summary

Hybrid quantum neural network for drug response prediction

Asel Sagingalieva, Mohammad Kordzanganeh, Nurbolat Kenbayev, Daria Kosichkina, Tatiana Tomashuk, Alexey Melnikov

arXiv:2211.05777v2quant-phcs.AI

TL;DR

Personalized drug-response prediction is important, but collecting the data needed by deep-learning models is challenging. The paper proposes a hybrid quantum neural network combining classical convolutional components with a deep quantum layer for IC50 prediction. The authors report more than 15% improvement in loss over a classical model with the same architecture.

  • Problem

    Deep learning benefits from large datasets, but collecting data for personalized-medicine tasks such as drug-response prediction is challenging.

    Source: §I · p. 1

  • Method

    The proposed HQNN combines graph-convolutional and convolutional networks with a deep quantum neural layer to predict IC50 values.

    Source: §I · p. 2

  • Results

    More than 15% improvement in loss is reported for the deep-quantum network compared with the classical model with the same architecture.

    Source: C. Training and results · p. 7

  • Takeaways & Limitations

    The authors report that the method achieves good performance on small-sample datasets, a setting relevant to personalized medicine where data collection may be challenging.

    Source: C. Training and results · p. 8

Abstract

from arXiv · show

Cancer is one of the leading causes of death worldwide. It is caused by a variety of genetic mutations, which makes every instance of the disease unique. Since chemotherapy can have extremely severe side effects, each patient requires a personalized treatment plan. Finding the dosages that maximize the beneficial effects of the drugs and minimize their adverse side effects is vital. Deep neural networks automate and improve drug selection. However, they require a lot of data to be trained on. Therefore, there is a need for machine-learning approaches that require less data. Hybrid quantum neural networks were shown to provide a potential advantage in problems where training data availability is limited. We propose a novel hybrid quantum neural network for drug response prediction, based on a combination of convolutional, graph convolutional, and deep quantum neural layers of 8 qubits with 363 layers. We test our model on the reduced Genomics of Drug Sensitivity in Cancer dataset and show that the hybrid quantum model outperforms its classical analog by 15% in predicting IC50 drug effectiveness values. The proposed hybrid quantum machine learning model is a step towards deep quantum data-efficient algorithms with thousands of quantum gates for solving problems in personalized medicine, where data collection is a challenge.

I. INTRODUCTION

Cancer’s genetic diversity calls for personalized treatment, while drug-response data are difficult to collect and deep learning generally benefits from large datasets. The paper proposes a hybrid quantum neural network (HQNN) for predicting drug-response IC50 values.

I. INTRODUCTION

  • Drug-response prediction matters because cancer cases differ through genetic mutations, making treatment selection a candidate for personalization.The introduction describes personalized oncology as seeking the most efficient and efficacious treatment.Source: §I · p. 1
  • Deep learning can support faster, less costly diagnosis and therapy selection, but collecting data for personalized-medicine tasks such as drug-response prediction is challenging.The passage notes that deep learning models generally benefit from large datasets.Source: §I · p. 1
  • Prior work describes hybrid quantum neural networks as potentially advantageous over classical networks on small datasets and identifies pharmacological prediction as a possible application.The cited examples include quantum-enhanced healthcare prediction and a hybrid quantum GAN for molecule discovery.Source: §I · p. 2
  • The proposed HQNN combines graph convolutional and convolutional networks with an 8-qubit Quantum Depth-Infused Neural Network Layer containing a 363-layer variational circuit.The layer encodes information, applies the variational circuit, and measures one qubit to produce a drug/cell-line response prediction.Source: §I · p. 2
  • The HQNN predicted IC50 values 15% better than its classical counterpart, which replaces the quantum layer with classical fully connected layers.The comparison is described as addressing a small-data drug-response prediction task.Source: §I · p. 2

II. MACHINE LEARNING FOR DRUG RESPONSE PREDICTION

This section reviews machine-learning approaches to drug sensitivity and frames the paper’s HQNN as a small-dataset approach to predicting IC50 values. The authors report that it predicts more accurately than its classical counterpart.

II. MACHINE LEARNING FOR DRUG RESPONSE PREDICTION

  • The GDSC-based approaches reviewed use drug descriptions, cancer-cell information, and calculated IC50 values.The section identifies GDSC as the shared dataset used in the cited model experiments.Source: §II · p. 2
  • IC50 represents the drug concentration needed to inhibit cells by 50%; higher values indicate lower drug effectiveness.The section describes it as a measure used to evaluate therapy effectiveness across cancer types.Source: §II · p. 2
  • Earlier approaches include elastic net biomarkers, random forests [38], and ridge regression, with the cited passages noting limits in feature coverage or generalizability.The elastic-net model used a relatively small feature set; the random-forest account notes room to improve robustness, and the ridge-regression account suggests more specific gene-structure features.Source: §II · p. 2
  • Neural approaches reviewed include an ensemble averaging predicted IC50 values [40], a multilayer model whose accuracy may improve with larger datasets, and the late-integration DL MOLI model.The approaches use, respectively, genomic fingerprints and selected mutation positions, drug-pathway analysis, and mutation, aberration-number, and omics data.Source: §II · p. 3
  • The authors characterize earlier classical networks as unable to account for all relevant factors, with small datasets and overfitting remaining concerns and results limited to special cases.This is the section’s stated motivation for pursuing a different approach.Source: §II · p. 3
  • The authors present an HQNN for IC50 prediction and report greater accuracy than its classical counterpart on a small dataset.They state that they had not found earlier studies using quantum or hybrid quantum neural networks for IC50 prediction.Source: §II · p. 3

III. RESULTS

The authors test their HQNN on a small portion of the GDSC dataset for drug-response prediction. They report lower IC50 prediction error than the classical analog, with the difference favoring the HQNN as training data decrease.

III. RESULTS

  • 15% lower prediction error is reported for the HQNN than for the classical analog on the tested small portion of GDSC.The passage also reports higher learning efficiency and fewer iterations for the HQNN.Source: §III · p. 4
  • On a small GDSC dataset subset, HQNN, combining graph convolutional, convolutional, and quantum layers, showed higher learning efficiency with fewer iterations and lower IC50 prediction error than its classical analog on 15%.Source: §III · p. 4
  • The authors state that the findings show good performance on small-sample datasets, a setting where data collection may be challenging.They identify personalized medicine as one such setting.Source: §III · p. 4

A. Description of dataset

The study uses version 6.0 of the Genomics of Drug Sensitivity in Cancer dataset to predict IC50 values for cell lines. The dataset combines drug-sensitivity measurements with genomic information.

A. Description of dataset

  • GDSC combines cell-line drug-sensitivity data with genomic datasets and genomic-feature analyses.Its compounds include approved drugs, clinical-development and trial drugs, and early-phase tool compounds.Source: A. Description of dataset · p. 4
  • The study uses IC50 as its primary outcome, measured in GDSC after 72 hours of drug treatment using fluorescence-based cell-viability assays.The dataset also includes dose-response-curve slope and area-under-the-curve values; the authors say the architecture could be adapted to AUC with possible loss-function modifications, but did not test it.Source: A. Description of dataset · p. 4

B. Hybrid quantum neural network architecture

The architecture processes drug chemistry and cancer-cell data in parallel, then combines their representations for quantum-layer analysis and IC50 prediction.

B. Hybrid quantum neural network architecture

1. Cell line representation

The cell line is represented as a binary mutation vector and encoded through a one-dimensional convolutional network into a 128-dimensional feature vector.

1. Cell line representation

  • A cell line is represented as a 735-dimensional binary vector, with 1 indicating a genomic mutation and 0 its absence.Source: §1 · p. 4
  • Three repetitions of convolutional and max-pooling layers encode the cell-line genomic features and reduce the feature map's spatial volume.Source: §1 · p. 4
  • A final fully connected layer flattens the output into a 128-dimensional vector.Source: §1 · p. 4

2. Drug representation

Drug molecules are represented as molecular graphs and processed by a graph convolutional network to form a 128-dimensional drug representation.

2. Drug representation

  • Each of the 223 drugs is represented by a molecular graph whose vertices are atoms and edges are chemical bonds.The graphs are derived from SMILES representations using PubChem and RDKit.Source: §2 · p. 4
  • The drug and cell-line representations each reach 128 neurons before being combined into a 256-neuron layer.The figure describes the drug chemistry branch as a graph convolutional network.Source: Figure 2 · p. 5
  • A graph convolutional network learns drug features from the molecular graphs.Its input includes an atom-feature matrix and an adjacency matrix representing atom connections.Source: §2 · p. 5
  • Three graph convolutional layers use ReLU activations, followed by global max pooling and fully connected layers producing a 128-dimensional vector.Source: §2 · p. 5

3. Quantum neural network

The quantum component encodes the combined 256 classical features on a lattice using eight qubits, interleaving feature embeddings with trainable variational layers before measurement.

3. Quantum neural network

  • Compared with the HQNN, the classical counterpart differs only in its last part, using two fully connected layers with 8 and 1 neurons; the HQNN has 1320 variational parameters versus 2056.Source: §3 · p. 6
  • Eight qubits encode the 256 features across a lattice of length 32 rather than assigning one qubit to each feature.The text cites qubit count and barren plateaus as reasons direct encoding of all 256 features is inaccessible or difficult to train.Source: §3 · p. 6
  • Angle embedding rotates each qubit around the Z-axis by an angle proportional to its corresponding input feature.Source: §3 · p. 6
  • Trainable rotations and CNOT operations form variational layers between feature-encoding layers, with five repetitions per lattice length and five additional layers before encoding.The quantum component has 1320 weights.Source: §3 · p. 6
  • The measurement stage propagates Z-measurement to all qubits and adjusts it with trainable multiplicative and additive parameters to predict IC50.Source: §3 · p. 6
  • At this circuit depth, the architecture is too challenging for current noisy quantum devices and is expected to remain on a simulator for the foreseeable future.Mathematically identical shallower circuits with more qubits are described as having greater potential for noisy-device execution.Source: §3 · p. 6

C. Training and results

The section evaluates the HQNN against a classical model on reduced, normalized GDSC drug/cell-line data using MSE. It reports a loss advantage for the HQNN and a larger relative gap with less training data.

C. Training and results

  • The reduced dataset was normalized and randomly split into 5,000 training and 1,000 testing samples, with training samples shuffled each epoch.IC50 values were transformed using a logistic-like function; testing samples remained in the same order.Source: C. Training and results · p. 7
  • Hyperparameters were selected using training performance and testing-set evaluation, without a separate dataset for hyperparameter optimization.The authors selected the settings that performed best on the testing set.Source: C. Training and results · p. 7
  • MSE measures the discrepancy between predicted and ground-truth IC50 values for each drug/cell-line pair, with lower values indicating better performance.The metric averages squared differences across the dataset’s drug/cell-line pairs.Source: C. Training and results · p. 7
  • Figure 3 reports the HQNN outperforms the classical counterpart by 15% in MSE loss across epochs, while the text says improvement exceeds 15% versus a same-architecture classical model.Source: Figure 3 · p. 7, C. Training and results · p. 7
  • HQNN MSE decreased monotonically during training, unlike the classical model’s volatile loss; gradients used backpropagation classically and the adjoint method quantumly.The authors describe the HQNN loss curve as more stable and efficient in minimizing loss.Source: C. Training and results · p. 7
  • As training samples decrease, the hybrid–classical MSE-loss difference grows in the HQNN’s favor across training sizes of 50, 200, and 5,000.Figure 4 presents the difference in MSE loss as a function of training size.Source: C. Training and results · p. 7, Figure 4 · p. 8

IV. CONCLUSION

The proposed HQNN combines graph and convolutional feature extraction with a deep quantum neural network for anticancer drug response prediction. Its circuit contains 1,320 trainable gates.

IV. CONCLUSION

  • The HQNN uses a deep quantum circuit with 1,320 trainable gates for anticancer drug response prediction.Source: §IV · p. 8
  • Drugs are represented as molecular graphs and cell lines as binary vectors, with graph convolutional and convolutional layers learning their features.Fully connected layers are also used for cell-line features.Source: §IV · p. 8
  • A newly designed Quantum Depth-Infused Neural Network Layer predicts response values for drug/cell-line pairs.Source: §IV · p. 8

Related papers

Loading 2211.05777v2…