EPI Prediction

EPIXplorer integrates 9 robust algorithms to perform the prediction. To facilitate the selection of algorithms, we provide practical guidance from three aspects: By Model Type (BMT), By Input Type (BIT), and By Bio Sample (BBS). BMT divides 9 algorithms into supervised and unsupervised according to the type of prediction model. For the unsupervised algorithms, we recommend PreSTIGE. For the supervised algorithms, the performance of LoopPredictor is the best. BIT characterizes these algorithms by the input they support. If users have both 3D contact and multi-omics data, LoopPredictor can be a good option. If users only have epigenomic data, IM-PET can be chosen. If there is no input file prepared, the server provides “no upload” mode to convenient the usage. BBS classifies all the supported cell lines into 9 major types and lists the available algorithms for each cell line.

Selection Guidance and the corresponding algorithms for each category. The best algorithms are marked with red color.

By Model Type

The computational methods integrated in EPIXplorer utilize DNA sequence or multi-omics features (chromatin accessibility, transcription factor binding sites, histone modifications, etc.) to predict chromatin interactions. Based on the type of model, we categorize these methods into unsupervised- and supervised-based.

Unsupervised

This type of method predicts the E-P interaction with distance or correlations between regulatory elements, which considers the natural patterns within DNA sequence, and doesn’t require 3D contact profile. Based on the strategies used to link distal regulatory elements to target genes, the unsupervised methods could be further divided into distance-based, correlation-based, and decomposition-based.
PreSTIGE
Distance-based
Ernst et al.
Thurman et al.
Correlation-based
EpiTensor
Decomposition-based

Supervised

This type of method extracts features from sequence or multi-omics datasets, then trains the model with available 3D contact data (Hi-C, ChIA-PET, eQTL, etc.). The trained model will be used to predict the E-P interaction of different cell types. Based on the tasks of prediction, supervised methods can be divided into regression-based and classification-based.
IM-PET
TargetFinder
Classification
JEME
3DPredictor
Regression
LoopPredictor
Classification+Regression

By Input Type

To meet different requirement, EPIXplorer supports different types of input.

☑ 3D contact ☑ multi-omics data

If you have both 3D contact and multi-omics data, select the following methods according to your 3D contact data type.
TargetFinder 3DPredictor LoopPredictor
Hi-C
LoopPredictor
ChIA-PET
LoopPredictor
HiChIP
LoopPredictor
eQTL

☒ 3D contact ☑ multi-omics data

If you have only multi-omics data, select the following methods according to the file format of your data.
IM-PET
RNA expression required
EpiTensor
PreSTIGE
ChiP-seq data required
☒ 3D contact ☒ multi-omics data
no upload mode
If you have neither 3D contact nor multi-omics data, you need to prepare the genomic locus of Enhancer or target gene name as input.
EpiTensor Thurman et al.
PreSTIGE Ernst et al.
Enhancer locus / Target gene required

By Biosample

To save your running time, EPIXplorer provides pre-trained model for 97 common biosamples, you can simply choose the cell line, input the enhancer region, or choose target gene name to obtain the prediction results.

Copyright @ 2016-CSU-Bioinformatics Group l All Right Reserved l
鲁ICP备2021005642号-1