Incorporating Quantum Annealing Methods in Ensemble Neural Networks for QSAR Problems

Prof. Nikitas J. Dimopoulos, University of Victoria,, 250-721-8902

We have employed quantum annealing techniques to the problem of NN ensemble learning. Quantum annealing can solve very large optimization problems, which require enormous computational time to solve using classical computing methods. Our approach contributed to methodologies developed by 1QBit and utilizing D-Wave’s system represented a purely Canadian solution to the quantum computation effort.

Our methodology includes heuristics that enhance the ability of the trained system to generalize.
The introduction of Quantum Annealing techniques in the NN training pipeline improved the generalization capabilities of the resulting ensembles [1].

Our methodology was tested both in simulation using 1 QBit’s environment and through the use of D-Wave’s D-Wave 2X system.
We have used this methodology to develop ensemble NN models that can accurately predict the biological activity of chemical compounds (QSAR), specifically PPAR -α and -γ (peroxisome proliferator activated receptors) agonists among others.

QSAR models are of interest in the pharmaceutical industry to assay drug candidates.

1. Ali Jooya, Babak Keshavarz, Nikitas Dimopoulos and Jaspreet Oberoi “Accelerating Neural Network Ensemble Learning Using Optimization and Quantum Annealing Techniques” 2nd International Workshop on Post Moore’s Era Supercomputing (PMES), Denver, November 13, 2017 doi: 10.1145/3149526.3149528