Application of Quantum Neural Networks in Drug Discovery: Predicting the Impact of Protein Mutations on Drug Binding

This article introduces an implementation of a quantum neural network based on the Farhi-Neven architecture, which is used to predict the impact of gene mutations on protein-drug binding interactions. The project reproduces the quantum machine learning component in the HypaCADD hybrid quantum-classical drug discovery workflow, demonstrating the practical application potential of quantum computing in biomedical classification tasks. The original author of the project is Behzad Jannati (Master of Computer Architecture at the University of Tehran), released on GitHub in May 2026. Original link: https://github.com/bhzadjnty7/QNN-for-Mutation-Impact-Prediction.

Traditional drug discovery methods are long and expensive (taking several years and billions of dollars in investment). The introduction of artificial intelligence has brought changes, but classical machine learning still faces challenges in complex molecular interaction problems. Quantum computing has attracted attention due to its potential advantages in handling specific optimization/classification problems. HypaCADD is a hybrid quantum-classical drug discovery framework that combines classical molecular docking, molecular dynamics simulation, feature extraction, and quantum neural networks to identify drug candidates resilient to gene mutations. This project focuses on the development of the mutation impact prediction module in this framework.

Architecture Design: Adopts the Farhi-Neven architecture (a variational quantum circuit dedicated to classification tasks). Qubit configuration: 4 input qubits (corresponding to 4 features) and 1 readout qubit, totaling 5 qubits (compatible with IBM5 qubit systems). Variational layers: 3 layers (RZX→RXX→RZX) containing 12 trainable parameters, updated iteratively via classical optimizers. Feature encoding: Uses RX rotation gates for encoding; features are processed with Min-Max normalization before encoding.

Feature Selection: 4 ligand-independent features: bind_site (whether the mutation is at the binding site), distance (distance between the mutation position and the ligand), polarity_change_index (amino acid polarity change index), volume_change_index (amino acid volume change index), balancing key impacts and generalization ability.

Dataset: Uses the GenoDock dataset from HypaCADD. Scale: 5142 samples in the training set, 5139 samples in the validation set. Class imbalance: non-disruptive mutations account for 93.5%, disruptive mutations account for 6.5%.

Training Strategy: Compares COBYLA and SPSA optimizers. COBYLA did not converge effectively; SPSA (stochastic approximation optimization, suitable for noisy environments) performed better. Final configuration: SPSA optimizer, 200 iterations, 0.05 learning rate, 0.05 perturbation amplitude.

Performance Metrics: Training set accuracy 91.52%, validation set accuracy 91.61%; initial loss 0.51, final loss 0.11.

Key Findings: 1. SPSA is significantly better than COBYLA; choosing the right classical optimizer is crucial. 2. QNN successfully learns biomedical classification patterns on severely imbalanced datasets. 3. The hybrid quantum-classical workflow is practically feasible. 4. QNN can be applied to real drug discovery tasks (not just proof of concept).

Current Limitations: 1. Dataset class imbalance (93.5% vs 6.5%). 2. 5-qubit limit on feature dimensions. 3. Use of quantum simulators (not run on real hardware).

Future Directions: 1. Training on real quantum hardware (to verify performance in noisy environments). 2. Introduce weighted loss functions to handle imbalanced data. 3. Try more complex variational circuits. 4. Build hybrid ensemble models (classical deep learning + QNN). 5. Conduct systematic performance comparisons with pure classical methods.