Benchmarks: ACNE: The complexity of the ACNE estimator is linear in the number of samples and probes

Author: Angel Rubio
Created on: 2010-05-06

ACNE is applied to each SNP separately. Thus, it never required more memory that the total amount of signals for one SNP across all samples. A high-level implementation of ACNE is also available for the aroma.affymetrix framework, which provides an automatic way of processing SNPs in chunks such that the memory usage is bounded/constant. Thus, ACNE can be done in aroma.affymetrix using only 1-1.5GB of RAM, regardless of the number of arrays processed. It is only the available disk space that limits the analysis.

The complexity is each of the steps in algorithm is linear with K (number of probes) or I (number of samples). We have run a simulation for a single SNP with I=10, 20, 50, 100, 200, 500, 1000, 2000, 5000 samples and a linear model fits perfectly the computing time (a standard Laptop with 2GB RAM). The y axis shows the total time per SNP. This simulation was performed for a 250K chip (K=20 probes per SNP). Total time in the summarization step can be obtained multiplying this time by the number of SNPs (in this case, about 80,000 seconds~=1 day for I=5,000 arrays).

ACNE is linear in K and I. The most complex process (computing the pseudinverse of H) is proportional to the number of samples.

Figure: The processing time of ACNE is proportional to the of the number of samples (I). The ACNE estimator was ran multiple times on simulated data sets with different number of samples. The presented data is the average times for many such runs.