Accelerated Spiking Convolutional Neural Networks for Scalable Population Genomics

Population genomics studies the genetic composition of populations to explain the evolutionary mechanisms that have contributed to species' adaptation to their environment. With continuous advances in DNA sequencing technologies, dataset sizes have grown, leading to an increased demand for processing power. Typical acceleration efforts for bioinformatics map established analytical models, including Convolutional Neural Networks (CNNs), to accelerator hardware like Graphical Processing Units (GPUs) and Field-Programmable Gate Arrays (FPGAs). However, these methods are not explicitly designed for high throughput or efficient mapping to these platforms, limiting their performance potential. Here, we address two major computational problems in population genomics: detecting selective sweeps and recombination hotspots. These evolutionary processes, crucial for species adaptation and survival, find applications in plant breeding, human genetics, and drug design. We propose a scalable solution that anticipates larger sample sizes, and, for the first time, we introduce a co-design approach that explores SCNNs for scalable population genomics and their implementation on FPGA technology. We choose SCNNs because these can be efficiently implemented on FPGA hardware due to their massive parallelism and binary communication, resulting in lower resource utilization and high-Throughput performance. Our findings show that when using SCNNs, it is only necessary to process a portion of the sample size while maintaining a classification accuracy comparable to conventional CNNs. We demonstrate the performance of our system in FPGA hardware for several genomic tasks, achieving up to 3X throughput and 100X less power consumption. This paves the way for a high-performance, future-proof acceleration solution that addresses the computational challenges of increasing sample sizes and longer chromosome lengths.