Test bowtie2 alignment vs STAR

[tacazares]

We have generated our ATAC-seq data using the STAR alignment method. We want to test how models trained on bowtie2 alignments will perform. Bowtie2 is used by the ENCODE consortium and is used by other methods such as BiChrom.

We have already generated the STAR ATAC-seq signal data. We will use a subset of this data for our testing. We will use 11 TFs in 6 cell types and test in GM12878.

To Do:

• Align the ATAC-seq signal using bowtie2
• Average replicates
• Min-max normalize signal tracks
• Train models
• Select best models
• Predict in GM12878
• Benchmark predictions
• Compare AUPR of the bowtie2 aligned samples to our current STAR results

[tacazares]

I trained models using data that was generated using bowtie2. There was almost no difference from the data that was generated using STAR with the default parameters. Both methods were normalized using min-max normalization to the 99th percentile max. There is a slight change in 1 TF, but most are within an acceptable range of difference.

Based on these results we can continue with the current data that have already generated or we could perform re-alignment of the data with bowtie2. We might get pushback from reviewers over not optimizing our alignment method, but it seems not to matter as much. We can test across sample performance, but based off of these results and previous comparisons we are probably going to have similar performance.