A calculation pipeline of expectation maximization for admixture application

Abstract

This thesis presents a calculation method of Admixture. Admixture is a tool to stratify the genotype data to infer the mixing ratio of its ancestry groups. Likelihood calculation was proposed to represent the admixture calculation. Expectation Maximization method was proposed to solve the likelihood problem. Expectation Maximization is based on an iterative mehod that takes very long time. Moreover, the calculation time is too slow for large genetic data. This dissertation proposed a design of FPGA for likelihood calculation. The FPGA has high floating-point operation performance and the calculation is suitable for FPGA. The number of floating point circuit that can be instantiated is more than the floating point units in a general processor. Systolic architecture was applied to the design. The systolic architecute and FPGA can directly stream the input data to the calculation unit. The streaming of input data reduces the memory overhead. The memory overhead is a main bottleneck of the calculation. Furthermore, the design of FPGA connection can reduce the communication cost of each calculation unit. The FPGA operated at 25 MHz gains speed up 4.5x when compared to a single core processor operated at 2.4GHz.