Hybrid quantum-classical graph neural networks for tumor classification in digital pathology

Abstract

Advances in classical machine learning and single-cell technologies in therapeutic design have paved the way to understand interactions between disease cells, microenvironment, and therapeutics to accelerate cell-centric therapeutic approaches. However there remains challenges in these machine learning methods due to the scale and complexity of these tasks and the NP-hard combinatorial problems in spatial Biology in general. This paper aims to utilize quantum computing approaches which can potentially offer significant advantages in understanding these problems better. In this paper we create a hybrid quantum-classical graph neural network that takes in a learned representation from a classical graph neural network to classify three binary classification tasks under the breast cancer sub-typing task. We aim to pursue two main research questions. Firstly, can quantum models help in higher feature dimension where classical learning algorithms tend to overfit or underfit for low data regimes? Secondly, since current quantum devices have fewer qubits, which compression scheme is better for improving training of quantum neural network: classical compression or compression via quantum encodings? The results demonstrate that the hybrid quantum neural network (QNN) is at par with the state of art classical graph neural networks (GNN) in terms of weighted precision, recall and F1-score. We also show that by means of amplitude encoding, we can compress exponential information in logarithmic number of qubits and attain better performance than using classical compression which leads to dataloss while keeping the number of qubits in both regimes. This paper gives insights on how to train the QNN better with detailed ablation studies using different quantum ansatzes, optimizer routines, and by gradually training with different train data sizes on different breast cancer sub-typing tasks.