Probabilistic module

[jasper1918]

Need a probabilistic module to assign significance. Metrics available currently:

• Unique vs all reads
• Number of junctions reads supporting breakpoint broken down by left, right and all orientations
• Number of discordant spanning reads broken down by all orientations
• Alignment scores for supporting reads
• Number of mismatches on supporting reads
• Non-clipped sequence lengths for overhangs, junction pairs, discordant pairs
• Base Qualities of all reads
• Assembly coverage, length, and status on crossing predicted junction
• Distance between breakpoints
• homology score(min, max, median, uq?)
• anchor and overhangs metrics including lengths, diversity, mismatch rate
• Expression Values
• Multimapping

[jasper1918]

Tried a few models(logistic regression, decision trees, neural networks) but in the end could get better AUC with hard thresholds due to the fact that boundary conditions have a huge influence. Used >6k events and achieved 95% sensitivity and 98% specificity. The FP/FN most come from homology artifacts on the mapping side. Marking closed until more data become available to train more robustly.