(V1.0 released) Provided function including:
1) filter the uploaded variants against public (1000Genome, dbSNPs, HapMap, ESP) and an in-house database to exclude the common variants;
2) process the remaining rare or de novo variants to a machine-learning based classifier to accurately distinguish "real" variants from "false" ones that arise due to sequencing bias;
3) annotate the function of variants (function region and damage to the protein). The accuracy can reach 90.48% with a specificity of 84.91% while testing on an experiment validated dataset.
(V1.2 released) Improvements including:
1)optimize the visualization of results page;
2) add a multi-threading module to accelerate the procress of feather extraction.
(V1.3 released) Improvements including:
1) to add more Sanger sequencing data to the training model for quality control;
2) to add rank function to estimating the pathogenicity of variants based on the probability of SVM model and
3) to add enrichment analysis and protein protein interaction network information to SNVfilter. In the revised and upgraded version of our web based tools, the training and testing of this model is carried out by using 1,860 SNVs validated by us through Sanger sequencing. The positive and negative dataset contain 1,300 and 560 SNVs respectively complementing the suggestion of the Editor. SNVfilter can also estimating the pathogenicity of variants based on a model constructed upon Support Vector Machine (SVM). The training and testing for this model is performed with 779,679 SNVs (the positive dataset contained 7,025 SNVs collected form OMIM, negative dataset contained 772,654 SNVs collected form HapMap and 1000 Genome with MAF > 5%).
(V1.4 demo released) Try use Deep Learning Framework - Caffe to perform pathogenicity prediction.
(V2.0 demo released) Improvements including:
1) Use TensorFlow to perform pathogenicity prediction;
2) new strategy to extract feather (4,980 feathers) for DNN;
3) remove the model for distinguish "real" variants from "false" ones.