Quantum-to-Classical Neural Network Transfer Learning Applied to Drug Toxicity Prediction

Area of Science:

• Computational chemistry
• Quantum machine learning
• Drug discovery

Background:

• Drug toxicity is a major obstacle in developing new therapeutics.
• Deep learning models for drug discovery face computational challenges due to large chemical spaces and matrix multiplications.
• Existing quantum machine learning approaches can be resource-intensive.

Purpose of the Study:

• To develop a hybrid quantum-classical neural network for predicting drug toxicity.
• To reduce the computational complexity of deep learning models in drug discovery.
• To leverage quantum computing's efficiency for enhanced drug candidate identification.

Main Methods:

• Designed a quantum circuit mimicking classical neural network behavior using matrix product calculations.
• Employed the Hadamard test for efficient inner product estimation, reducing qubit requirements.
• Implemented a hybrid approach allowing quantum-derived weights to be transferred to classical devices.

Main Results:

• The hybrid model achieved predictive accuracy comparable to its fully classical counterpart on the Tox21 dataset.
• Demonstrated seamless transfer of learning from the quantum to the classical component without performance disruption.
• Showcased a reduction in qubit requirements by half compared to conventional methods.

Conclusions:

• Hybrid quantum-classical models offer a promising avenue for scalable machine learning in drug discovery.
• Combining quantum advantages (reduced complexity) with classical advantages (noise-free computation) can overcome current limitations.
• This framework facilitates efficient drug toxicity prediction and accelerates the identification of safe and effective drug candidates.