Skip to content

SmartSeq2 Pipeline Timing and Cost

Jump to bottom
Jishu Xu edited this page Sep 27, 2017 · 33 revisions

Table of Contents

Google Cloud Computing Pricing

It could be very complicated to estimate the cost of cloud computing. For google cloud, the billing is on the reserved resource, such as number of cores, size of harddrive and RAM. You will be charged by an hourly rate over the time you use the reserved resource. The pricing also varies between machine type. For example, the hourly rate of using a high specs machine (more nodes and high RAM) is higher than low specs ones. The pricing page lists details.

scRNASeq Pipeline Timing and Cost

The cost of processing RNA-Seq on cloud computing depends one several factors. One is running time. Second is how many resources have been reserved during data processing. Third is machine type. It could be tricky to estimate the cost. For example, to run one RSEM job, we can either request min resource, such as 1-core with 4Gb RAM machine or we can request a 4-core and 15Gb machine. The hourly rate of the first type of machine is cheaper than the second one but it takes longer time to finish job.

Pipeline Runtime

In this task, we estimated the cost of processing scRNA-Seq data by using google cloud. The testing pipeline included the following modules/steps:

  • STAR alignment [request 8-core and 40Gb RAM]
  • RSEM to estimate gene counts [request 4-core and 4Gb RAM]
  • FeatureCount to calculate gene/exon/transcript counts [request 1-core and 4Gb RAM]
  • Picard to collect several sequencing and RNA-Seq specific metrics [request 1-core and 4Gb RAM]
  • Python script to parse pipeline output [request 1-core and 2Gb]

Timing and Cost

Then we tested this pipeline on a published scRNA-Seq dataset, which includes 864 single cells full length RNA-Seq data. And we summarize the total hours(in hour) and total cost(in dollar) in the following table.

StepName Program/software Machine Type Total Hours Total Cost AVG Hours AVG Cost
Star Star n1-highmem-8 292.133 140.115 0.338 0.1621
RSEM RSEM n1-standard-4 151.983 29.376 0.1759 0.034
CollectAlignemntMetrics Picard n1-standard-2 144.216 13.858 0.166 0.016
CollectRnaMetrics Picard n1-standard-2 144.555 13.890 0.167 0.0160
CollectInsertSizeMetrics Picard n1-standard-2 144.116 13.849 0.1668 0.016
CollectDuplicationMetrics Picard n1-standard-2 144.433 13.958 0.167 0.0161
FeatureCountsUniqueCounts FeatureCount n1-standard-2 144.266 14.179 0.167 0.0164
FeatureCountsMultiCounts FeatureCount n1-standard-2 144.483 14.200 0.1672 0.0164
CollectMetricsbySample python n1-standard-1 144.466 7.535 0.167 0.0087
Total 1454.65 260.9

Notice The Total Hours and Total Cost truly reflect the total amount of running one task in Google computing, which not only includes the actually programming runs but also includes the copying/localizing files into/from google cloud VM.

Correlation with scRNASeq Quality Metrics

First we calculated the correlation between overall pipeline running time/cost and RNASeq QC metrics. The overall pipeline running time and cost show correlation with several QC metrics, such as PF_BASES,INTERGENIC_BASES and PF_MISMATCH_RATE.

Next we calculated the correlation between each pipeline module's running time&cost and QC metrics STAR running hours and cost RSEM FeatureCounts Picard Collect Metrics

The timing and cost of STAR and RSEM have stronger correlation to RNASeq QC metrics compared to the timing and cost of Picard and FeatureCounts.

Then we examined the impact of metrics PF_BASES, PF_ALIGNED_BASES,INTRONIC_BASES, PF_MISMATCH_RATE,MEAN_READ_LENGTH on pipeline running time and cost. In general, PF_BASES,PF_ALIGNED_BASES and INTRONIC_BASES tell us the size of sequencing data and PF_MISMATCH_RATE tell us the quality of alignment.

Impact on STAR

And RSEM images/costs/total_hours_cost_reads_RsemExpression_BASES_facet_plot.png

Clone this wiki locally