Skip to main content
. 2022 Nov 9;50(20):11442–11454. doi: 10.1093/nar/gkac990

Figure 5.

Figure 5.

Incremental learning for designing future MPRA experiments. (A, B) MpraNet performance on the validation set increases with the training set size, with MPRA positives and background fixed at a 1:10 ratio. This suggests incorporation of results from future MPRA experiments can further improve MpraNet. (C) MpraNet gives a continuous score from 0 to 1. Choosing a threshold to select top scoring variants gives varying recall on the validation set. In comparison with other benchmark scores, MpraNet achieves higher recall at any threshold. (D) For any fixed threshold and corresponding recall, MpraNet precision is estimated using Bayes’ rule, assuming a value of MPRA function prevalence. Here we estimate precision for four thresholds corresponding to testing the 25 000 to 100000 top scoring variants from the 1000 Genomes Project. For example 100 000 (pink) corresponds to the top 0.011 percentile.