0.4:

• Features:
• For missing values using genome mean of feature for sequence and conservaton features. 1 for p-value. All other features have zero as missing value
• Updating DGVCount to 02/25/2020 on hg19/hg38
• Update dbVARCount to 10/20/2021 on hg19/hg38
• Update ISCApath to 11/03/2021 on hg19/hg38
• Replace tfbsConsSites with UCSC table encRegTfbsClustered on hg19/hg38
• Software:
• Using parSMURF for training
• Complete retraining of hg19 and hg38 builds (hg19: AUROC=0.993; AUPRC=0.394; hg38: AUROC=0.996; AUPRC=0.610)

0.3.1.post1:

• New hg38 release. Completely retrained on the new genome build.
• Training data
• Liftover positives. No change in size.
• Negatives used from CADD v1.4 GRCh38 (human derived), filtered as described in the original paper (https://doi.org/10.1016/j.ajhg.2016.07.005) Size slightly different (hg38: 13,902,234; hg19: 14,755,199).
• Features
• Same size as in hg19: 26 features.
• We tried to use the same features as in hg19. Sometimes new versions of data have to be used (e.g. DGV, ISCA, dbVAR).
• Training was done with the parSMURF implementation of hyperSMURF.
• Same hg19 parameters are used.
• Metrics via 10-fold cytoband cross-validation (same cytoband to fold map).
• Area under the ROC curve: 0.996 (hg19: 0.989, see https://doi.org/10.1016/j.ajhg.2016.07.005).
• Area under the precision recall curve: 0.548 (hg19: 0.441,, see see https://doi.org/10.1016/j.ajhg.2016.07.005).
• Scores for hg19 in this release are the same as version 0.3.1. Only the files have been renamed.

0.3.1:

• Bugfix of region chr17:79759050-81195210. Region is missing in older versions.

0.3:

• First official public version.
• Values for positions in training data are computed by cytoband-aware 10 fold cross-validation.
• Other position scores are compted by a generalized model of all training data.
• This version was used in the Genomiser publication (Smeley et.al. A Whole-Genome Analysis Framework for Effective Identification of Pathogenic Regulatory Variants in Mendelian Disease. AHJG. 2016).