Poster: GPU-accelerated artificial neural network for QSAR modeling

A GPU-accelerated OpenCL implementation of a back-propagation artificial neural network for the creation of QSAR models for drug discovery and virtual high-throughput screening is presented. A QSAR model for HSD achieved an enrichment of 5.9 and area under the curve of 0.83 on an independent data set which signifies sufficient predictive ability for virtual high-throughput screening efforts. The speed-up demonstrated on this data set allows for the complete cross-validated feature optimization of QSAR models based on ANNs within 24 hours on a workstation equipped with 4 consumer GPUs achieving performance equal to that of ~340 cores. This GPU-accelerated ANN framework for the creation of optimized QSAR models from biological data will be available free of charge for academic users at http://www.meilerlab.org through a server interface.