Hybrid Quantum-Classical Graph Neural Networks

Understanding how tumor cells self-organize and interact within the tumor microenvironment (TME) is a long standing question in cancer Biology. In This project, we modelled TME as a graph structure to comprehensively encode the tissue composition in terms of biologically meaningful entities, such as cells, tissues, and their interactions. As a result, GNNs can elegantly integrate cellular information with tumor morphology, topology, and interactions among cells and/or tissue structures. We created a hybrid classical-quantum network which combines a GNN with a Variational Quantum Classifier (VQC).

Implementation of hybrid GNN-VQC model.

We followed two approaches for training our Hybrid Network with different encoding schemes and feature dimensions:

(i) with a pretrained GNN (having frozen weights)

(ii) with trainable GNN and VQC parameters to backpropogate simultaneously.

Classical compression vs Quantum compression.

We showed that end-to-end training is significantly better than serially training such models and demonstrate results on a real-world breast-cancer subtyping task. In detailed ablation studies we observed that quantum compression can be significantly better to qubit requirements without information loss unlike lossy classical compression. Future directions could be to explore how other such classical networks can be combined with quantum circuits to enhance their trainability and improve generalization.

Source code:

End to end trainable GNN-VQC

This paper presents a novel approach in the intersection of quantum computing and machine learning, specifically focusing on the application of a hybrid quantum-classical graph neural network (GNN) for tumor classification in digital pathology. The proposed methodology integrates a Variational Quantum Classifier (VQC) with GNN, exploring two variants: one with fixed pretrained GNN parameters and another with end-to-end training of GNN+VQC. The study reveals that the hybrid model achieves comparable performance to state-of-the-art classical GNNs in weighted precision, recall, and F1-score. Additionally, it demonstrates the efficacy of amplitude encoding in compressing information with a logarithmic number of qubits, outperforming classical compression techniques. The research also highlights the advantages of end-to-end training in enhancing the model’s performance.

Hybrid Quantum-Classical Graph Neural Networks

References

2023