Journal: Bioinformatics

Article Title: Deep5hmC: predicting genome-wide 5-hydroxymethylcytosine landscape via a multimodal deep learning model

doi: 10.1093/bioinformatics/btae528

Figure Lengend Snippet: Overview of Deep5hmC. ( A ) The training set of Deep5hmC can be derived from matched 5hmC-seq and other epigenetic data such as histone ChIP-seq or DNase-seq/ATAC-seq from one condition. Specifically, the 5hmC-seq data can be collected from tissue-specific human tissues, which include bladder, brain, breast, heart, kidney, liver, lung, marrow, ovary (female), pancreas, placenta (female), prostate (male), colon (sigmoid), colon (transverse), skin, stomach, and testis (male). The matched tissue-specific epigenetic data, such as histone ChIP-seq data profiling histone modification and DNase-seq/ATAC-seq profiling chromatin accessibility, can be collected from public consortiums such as Roadmap Epigenomics and ENCODE. In this context, Deep5hmC aims to predict genome-wide 5hmC modification in a single condition. ( B ) The training set of Deep5hmC can also be derived from matched 5hmC-seq and ChIP-seq from a case–control study (e.g. Alzheimer’s disease (AD) versus healthy control) for predicting differentially hydroxymethylated regions (DhMRs). ( C ) Deep5hmC is a multimodal deep learning model to improve the prediction of tissue/cell type-specific genome-wide 5hmC modification by leveraging both DNA sequence and epigenetic features such as histone modification and chromatin accessibility. Deep5hmC consists of four modules, including Deep5hmC-binary, Deep5hmC-cont, Deep5hmC-gene, and Deep5hmC-diff. Specifically, Deep5hmC-binary takes the labeled 5hmC peaks and non-peaks as the training set to identify the 5hmC-enriched regions. Deep5hmC-cont takes the normalized read counts in 5hmC peaks and aims to predict the continuous 5hmC modification genome-wide. By leveraging Deep5hmC-cont, Deep5hmC-gene aggregates the predictions of Deep5hmC-cont in the gene bodies as the surrogate for the predicted gene expression. Different from Deep5hmC-binary, Deep5hmC-diff takes the labeled DhMRs/non-DhMRs in a case–control design of 5hmC-seq as the training set to predict genome-wide DhMRs and may discover de novo DhMRs. ( D ) Model architecture of Deep5hmC. Deep5hmC consists of both sequence modality and epigenetic modality consisting of their own convolutional neural networks (CNNs) to derive separate feature representations, which will be joined later via the multi-modal factorized bilinear (MFB) pooling fusion layer. The output of the MFB fusion layer will further connect to fully connected layers and the output layer afterward.

Article Snippet: For “Human Tissues,” we carefully select aligned bed files of ChIP-seq data from Roadmap Epigenomics by ensuring a match between ChIP-seq data and 5hmC-seq data based on tissue type.

Techniques: Derivative Assay, ChIP-sequencing, Modification, Genome Wide, Control, Sequencing, Labeling, Expressing

Journal: Bioinformatics

Article Title: Deep5hmC: predicting genome-wide 5-hydroxymethylcytosine landscape via a multimodal deep learning model

doi: 10.1093/bioinformatics/btae528

Figure Lengend Snippet: Distribution pattern of histone modification around 5hmC peaks. EB 5hmC peaks are collected from “Forebrain Organoid” 5hmC-seq data and ChIP-seq data in “Brain Angular Gyrus” from seven histone marks are collected from Roadmap Epigenomics. Histone features are obtained and averaged in the neighborhood of all 5hmC peaks for the positive and negative sets, respectively. Specifically, histone features are created by segmenting an extended genomic region of 10 kb both upstream and downstream of each 5hmC peak into 41 1 kb windows with a sliding size of 500 bp and counting reads for each 1 kb windows. For each histone mark, the Kolmogorov–Smirnov test is performed to test the distribution difference of histone features between positive and negative 5hmC peaks and the P -value is reported.

Article Snippet: For “Human Tissues,” we carefully select aligned bed files of ChIP-seq data from Roadmap Epigenomics by ensuring a match between ChIP-seq data and 5hmC-seq data based on tissue type.

Techniques: Modification, ChIP-sequencing

Journal: Bioinformatics

Article Title: Deep5hmC: predicting genome-wide 5-hydroxymethylcytosine landscape via a multimodal deep learning model

doi: 10.1093/bioinformatics/btae528

Figure Lengend Snippet: Comparison of unimodal and multimodal Deep5hmC for predicting binary 5hmC modification sites. When using histone modification in the epigenetic modality, two unimodal models of Deep5hmC: Deep5hmC-Seq using only DNA sequence as the model input and Deep5hmC-His using only histone modification as the model input are compared to the default multimodal Deep5hmC-Seq+His using both DNA sequence and histone modification as the model input. 5hmC peaks from the EB stage “Forebrain Organoid” and two histone marks: H3K4me1 and H3K4me3 ChIP-seq data in all brain regions from Roadmap Epigenomics are used as the training set. ( A ) AUROC reported for three compared methods. ( B ) AUPRC reported three compared methods.

Article Snippet: For “Human Tissues,” we carefully select aligned bed files of ChIP-seq data from Roadmap Epigenomics by ensuring a match between ChIP-seq data and 5hmC-seq data based on tissue type.

Techniques: Comparison, Modification, Sequencing, ChIP-sequencing

Journal: bioRxiv

Article Title: Chrom-Sig: de-noising 1-dimensional genomic profiles by signal processing methods

doi: 10.1101/2025.08.12.670000

Figure Lengend Snippet: Paired End Example for GM12878 Chrom-Sig results for all paired-end datasets from GM12878 cell-line, visualized in the genome browser. CTCF Motif: CTCF binding sites with orientation. Original: Bedgraph file generated directly from input BAM/bed file. SICER peaks: Bed file result of running SICER algorithm on the original bedgraph file. Chrom-Sig FDR 0.1 pass: pass bedgraph generated from original bedgraph by Chrom-Sig (percentage refers to how many reads were retained by Chrom-Sig result from original bedgraph). SICER peaks (below Chrom-Sig FDR 0.1 pass): Bed file from SICER algorithm run on pass-pileup bed generated by Chrom-Sig. ChromHMM: Chromatin states.

Article Snippet: The analysis looks at GM12878 CTCF ChIP-seq ENCFF355CYX (36,269 peaks original, 24,872 peaks after Chrom-Sig) as well as GM12878 CTCF CUT&RUN replicates 4DNFI2G71DR4 (55,251 peaks original, 22,554 peaks after Chrom-Sig) and 4DNFI9U71IB4 (62,176 peaks original, 19,233 peaks after Chrom-Sig).

Techniques: Binding Assay, Generated

Journal: bioRxiv

Article Title: Chrom-Sig: de-noising 1-dimensional genomic profiles by signal processing methods

doi: 10.1101/2025.08.12.670000

Figure Lengend Snippet: Single End Example Chrom-Sig results for all single-end datasets (all single-end data is from GM12878 cell-line), visualized in the genome browser. CTCF Motif: CTCF binding sites with orientation. Original: Bedgraph file generated directly from input BAM/bed file. SICER peaks: Bed file result of running SICER algorithm on the original bedgraph file. Chrom-Sig FDR 0.1 pass: pass bedgraph generated from original bedgraph by Chrom-Sig (percentage refers to how many reads were retained by Chrom-Sig result from original bedgraph). SICER peaks (below Chrom-Sig FDR 0.1 pass): Bed file from SICER algorithm run on pass-pileup bed generated by Chrom-Sig. ChromHMM: Chromatin states.

Article Snippet: The analysis looks at GM12878 CTCF ChIP-seq ENCFF355CYX (36,269 peaks original, 24,872 peaks after Chrom-Sig) as well as GM12878 CTCF CUT&RUN replicates 4DNFI2G71DR4 (55,251 peaks original, 22,554 peaks after Chrom-Sig) and 4DNFI9U71IB4 (62,176 peaks original, 19,233 peaks after Chrom-Sig).

Techniques: Binding Assay, Generated

Journal: bioRxiv

Article Title: Chrom-Sig: de-noising 1-dimensional genomic profiles by signal processing methods

doi: 10.1101/2025.08.12.670000

Figure Lengend Snippet: CTCF Motif Analyses a) Top enriched motifs, E-value, and matching motifs from MEME-Chip for GM12878 CUT&RUN CTCF 4DNFI2G71DR4 before and after Chrom-Sig. b) Comparison of CTCF motif precision between original data and Chrom-Sig with FDR 0.1 and 5000 pseudo-reads for GM12878 ChIP-seq CTCF ENCFF355CYX, GM12878 CUT&RUN CTCF 4DNFI2G71DR4 and 4DNFI9U71IB4.

Article Snippet: The analysis looks at GM12878 CTCF ChIP-seq ENCFF355CYX (36,269 peaks original, 24,872 peaks after Chrom-Sig) as well as GM12878 CTCF CUT&RUN replicates 4DNFI2G71DR4 (55,251 peaks original, 22,554 peaks after Chrom-Sig) and 4DNFI9U71IB4 (62,176 peaks original, 19,233 peaks after Chrom-Sig).

Techniques: Comparison, ChIP-sequencing

Journal: bioRxiv

Article Title: Chrom-Sig: de-noising 1-dimensional genomic profiles by signal processing methods

doi: 10.1101/2025.08.12.670000

Figure Lengend Snippet: ChromHMM State Annotation Distribution Comparison of the distribution of chromHMM states between original data and Chrom-Sig with FDR 0.1 and 5000 pseudo-reads for K562 RNAPII ChIP-seq ENCFF480AJZ and ENCFF785OCU and GM12878 ATAC-seq ENCFF646NWY. The proportion of enhancer and promotor states increases when Chrom-Sig is applied to the data. Between K562 RNAPII ChIP-seq replicates there is an average of 12.3% higher distribution of enhancers and promotors (ENCFF480AJZ: 77% original vs 87.3% Chrom-Sig and ENCFF785OCU: 76.9% original vs 85.5% Chrom-Sig). In ATAC-seq data, the percentage of transcription and heterochromatin states drops from 28.9% to 12.6% after Chrom-Sig.

Article Snippet: The analysis looks at GM12878 CTCF ChIP-seq ENCFF355CYX (36,269 peaks original, 24,872 peaks after Chrom-Sig) as well as GM12878 CTCF CUT&RUN replicates 4DNFI2G71DR4 (55,251 peaks original, 22,554 peaks after Chrom-Sig) and 4DNFI9U71IB4 (62,176 peaks original, 19,233 peaks after Chrom-Sig).

Techniques: Comparison, ChIP-sequencing