Here we provide a comparison between empirical base qualities and predicted base qualities output by DeepConsensus. The predicted base qualities should ideally reflect the true probability of the base being incorrect. In order to quantify this, we plot a calibration curve with predicted qualities, as output by the model, versus empirical qualities, as observed from the data (when labels are available).

• Empirical quality refers to the observed quality at any given predicted quality. Here, we aligned all the reads to CHM13 and counted how many times we observe correct and incorrect bases at a given predicted base quality. From this, we calculated the phred score to obtain the empirical quality.

• Predicted quality refers to the probability output of DeepConsensus or CCS associated with each base, which represents the model's confidence in the prediction. This is converted to a base quality score (e.g. probability of 0.99 corresponds to Q20).

This plot quantifies how well calibrated DeepConsensus is as measured on the m64062_190806_063919-chr19 dataset (CHM13, chr19 held out from training).

We observe that especially with --skip_windows_above=0, which runs DeepConsensus on all windows the model appears to be well-calibrated.