Measuring, visualizing and diagnosing reference bias with biastools

A goal of recent alignment methods is to reduce reference bias, which occurs when reads containing non-reference alleles fail to align to their true point of origin. However, there is a lack of methods for systematically measuring, categorizing, and diagnosing reference bias. We present biastools, which analyzes and categorizes instances of reference bias. Biastools has different sets of functionality tailored to different scenarios, i.e. (a) when the donor genome is well-characterized and input reads are simulated, (b) when the donor is well-characterized and reads are real, and (c) when the donor is not well-characterized and reads are real. When possible, biastools divides instances of reference bias into categories according to their cause: bias due to loss, flux, or local misalignment. Biastools’s scan mode detects large-scale mapping artifacts due to structural variation and flaws in the reference representation. Our findings confirm that including more variants in a graph genome alignment method results in fewer reference biases. We also find that end-to-end alignment modes are effective in reducing bias at insertions and deletions, compared to local aligners that allow soft clipping. Finally, we use biastools to characterize the ways in which using the new telomere-to-telomere human reference can improve bias at a large scale. In short, biastools is a tool uniquely focused on reference bias, making it a valuable resource as the field continues to develop new aligners and pangenome representations to reduce bias.