An integrated view of the structure and function of the human 4D nucleome

Abstract

The dynamic three-dimensional (3D) organization of the human genome (the 4D nucleome) is linked to genome function. Here we describe efforts by the 4D Nucleome Project¹ to map and analyse the 4D nucleome in widely used H1 human embryonic stem cells and immortalized fibroblasts (HFFc6). We produced and integrated diverse genomic datasets of the 4D nucleome, each contributing unique observations, which enabled us to assemble extensive catalogues of more than 140,000 looping interactions per cell type, to generate detailed classifications and annotations of chromosomal domain types and their subnuclear positions, and to obtain single-cell 3D models of the nuclear environment of all genes including their long-range interactions with distal elements. Through extensive benchmarking, we describe the unique strengths of different genomic assays for studying the 4D nucleome, providing guidelines for future studies. Three-dimensional models of population-based and individual cell-to-cell variation in genome structure showed connections between chromosome folding, nuclear organization, chromatin looping, gene transcription and DNA replication. Finally, we demonstrate the use of computational methods to predict genome folding from DNA sequence, which will facilitate the discovery of potential effects of genetic variants, including variants associated with disease, on genome structure and function.

Fig. 1. Overview of the approach to generate and integrate genomic data on the 4D nucleome.

Top left, schematic of the two types of complementary genomic assay for mapping 3D genome folding and the relative distances of genomic loci to nuclear bodies in H1 and HFFc6 cells (top left). Top right, different chromatin interaction mapping methods were compared and benchmarked to assess their ability to identify and quantify 3D genome features at scales ranging from chromatin compartments (megabase) to focally enriched chromatin interactions (kilobase). Bottom left, additional multimodal datasets generated or used to facilitate integrative analyses (see below). HIPmap, high-throughput imaging position mapping. Bottom middle, multiple integrative modelling and analysis approaches were conducted to reveal the spatial features of chromatin loci by combining 3D genome features and various multimodal datasets. The connections between different input data and integrative analyses is illustrated through colour-coded flow paths. Bottom right, an illustrative cartoon summarizes the overarching aim of the project, which aims to provide insights into structure–function relationships by connecting variable 3D genome features (represented on the x axis) derived from multimodal datasets (y axis) with key cellular functions, such as transcription and replication (z axis). Our models pave the way for identifying the sequence determinants of genome folding and predicting how different variants might influence this folding process. Hi-C contact map examples were drawn using ORCA.

Fig. 2. Methods for chromatin interaction mapping differ in quantitative detection of compartmentalization.

a, Contact maps (100-kb bins, chromosome 19: 0–20 Mb) generated using the indicated methods were obtained with H1 cells. The plots below the heat maps show EV1 (compartments: red, A compartment; blue, B compartment). Bottom, magnified contact maps (corresponding to the blue squares in the heat maps in the top panel; 25-kb bins, chromosome 19: 17.5–20 Mb). The plots below the bottom heat maps show insulation profiles. b, Contact maps (100-kb bins, chromosome 2: 0–70 Mb) were generated using the indicated methods obtained with HFFc6 cells (top). The plots below the heat maps show EV1. Bottom, magnified contact maps (corresponding to the blue squares in the heat maps in top panel; 25-kb bins, chromosome 2: 12–16 Mb). The plots below the bottom heat maps show insulation profiles. c, Spearman correlation of compartment profiles determined by Eigenvector decomposition (Methods). d, Compartment strength quantified using eigenvectors from contact data obtained using the corresponding 3D methods. e, Pearson correlation of genome-wide insulation scores for all methods. f, Aggregated insulation scores at strong boundaries detected in multiple datasets (Methods, Supplementary Note) for all methods. g, Preferential interactions quantified in Hi-C, Micro-C, ChIA-PET, PLAC Seq, SPRITE and GAM, using DamID-seq for lamin B1, early and late replication timing (E/L RT) using RepliSeq and TSA–seq for SON to rank loci. The fold enrichment indicates the preference of loci with similar associations with speckles (SON), nucleoli (POLR1E/NIFK), lamina (lamin B), or that display early or late replication, to interact with each other, as detected by the indicated assays.

Fig. 3. Cross-platform loop comparisons in the H1 cell line.

a, Schematic of the construction of the feature matrix used for UMAP projection of chromatin loops. b, UMAP projection and clustering of 141,365 union loops in H1 cells, based on the ChromHMM state composition at their loop anchors. Cluster IDs and the number of loops per cluster are labelled on the UMAP. c, Fold enrichment of state pairs in each cluster. ChromHMM States: AP, active promoter; WP, weak promoter; TE, transcriptional elongation; TT, transcriptional transition; SE, strong enhancer; PP, poised promoter; repressed, polycomb repressed. d, The size distribution (genomic distance between loop anchors) of chromatin loops in each cluster. Within each violin, the black dot represents the median, and the vertical line represents 1.5 × the interquartile range (IQR). The number of loops (n) in each cluster is indicated on the UMAP in b. e, The cluster composition of loops detected by each platform. f, Example illustrating differences among platforms in detecting insulator-related loops. Contact maps are plotted at 5-kb resolution, with loops detected by each platform marked by blue circles. g, Example illustrating differences among platforms in detecting transcription-related loops. Contact maps are plotted at 1-kb resolution; loops detected by each platform are marked by blue circles. Loops linking the SOX2 gene to distal enhancers are indicated by black arrows, and the interacting enhancers are highlighted with yellow shading.

Fig. 4. 3D structure–function relationship and its cell-to-cell variability.

a, SPIN states define spatial genome compartments. The heat maps show enrichment of histone marks, Repli-seq and caRNAs (columns) across SPIN states (rows) in H1 cells. The fold change shows the ratio of observed signals over the genome-wide expectation. b, The average radial positions for a chromosome 1 segment in H1 (top) and HFFc6 (bottom) cells. c, Representative single-cell 3D structures of chromosome 1 in H1 (left) and HFFc6 (right) cells. The yellow circles mark POU3F1 loci; red and blue shading denotes chromatin in speckle- and lamina-associated SPIN states; spheres indicate predicted speckles. d, POU3F1 expression (RNA-seq) in H1 and HFFc6 cells (left). Right, the joint distribution of the nearest speckle/lamina distances of POU3F1 across 1,000 structures. The box plot shows the median (centre line), IQR (box) and the whiskers extend to 1.5 × IQR. n = 7 (H1) and 5 (HFF). e, The same as in d but for THBS1. f, t-Distributed stochastic neighbour embedding (t-SNE) of transcription start site (TSS)-containing regions for the top 25% expressed genes, based on 3D-structure features, separated by genome-wide top (left) and bottom (right) quartiles of SAF. g, The log₂-transformed fold enrichment of 14 structural features for highly expressed class I and II genes (top): radial position (RAD, 1 − norm); chromatin decondensation (RG, ±500 kb); distance to nearest speckle/nucleolus (SpD/NuD); interior localization probability (ILF); speckle, lamina and nucleolus association frequencies (SAF, lamina association frequency (LAF) and nucleolus association frequency (NAF)); interchromosomal interaction probability (ICP), trans A/B ratio (TransAB), cell-to-cell variability (δRAD, δRG, δSpd and δNuD). Bottom, log₂ enrichment of genomic features (within 200-kb bin) and histone marks (±10 kb from TSS). h, Spatial enhancer count within 350 nm from TSS for class I and II genes: intrachromosomal (left), ultra-long-range >1 Mb (middle) and interchromosomal (right). The box plots are as described in d. n = 2,275 (class I) and 659 (class II). P values were calculated using two-sided Mann–Whitney U-tests with no adjustment for multiple comparisons. i, Pile-up Hi-C contact frequencies (10-kb bins) centred on TSS of class I and II genes showing mean contact decay with sequence distance.

Fig. 5. Cell-to-cell variability in 3D genome features.

a, Merged scHi-C contact maps imputed by Higashi or predicted by the SBS model, as compared to bulk Hi-C and raw contact maps from scHi-C without imputation (top). Bottom left, insulation scores from bulk Hi-C, calculated after Higashi imputation, and SBS modelling. Bottom right, Spearman correlation coefficients between these contact maps. b, 3D models, raw scHi-C contact map, imputed maps from three similar cells between Higashi imputation and SBS model maps. The Higashi–SBS model contact map pairs have distance-stratified similarity scores of 0.69, 0.64 and 0.77 (left to right). c, The average normalized intensity of chromatin loops across 188 cells was calculated and compared by dividing loops on the basis of their position within TADs and A/B compartments (comp.). Left, the difference between loops in the same TAD (n = 181 cells) and loops spanning multiple TADs (n = 7 cells). Right, the difference between loops in the same A/B compartment (n = 157 cells) and loops spanning different compartments (n = 31 cells). A chromatin loop near RABGAP1L is highlighted in the right plot. The original distribution of the normalized intensity of this loop in each cell is shown in the right plots. Loops are stratified into groups depending on whether they locate within one TAD (n = 181 cells) or span TADs (n = 7 cells), or the A/B compartment state of loop anchors in each single cell (n = 15 (AA), n = 142 (BB) and n = 31 (AB) cells). The box plots show the median (centre line), IQR (box) and the whiskers extend to 1.5 × IQR. P values were calculated using two-sample two-sided t-tests. d, Aggregated contact map from single-cell Hi-C data at RABGAP1L (for cells with z score > 1.96). The circle with a dashed line indicates the 450-kb loop identified by SnapHi-C. e, Knight-Ruiz-normalized bulk Hi-C map from WTC-11 at RABGAP1L.

Fig. 6. Associations between enhancer–promoter loops and gene regulation.

a, Gene transcription levels versus the number of interacting enhancers in H1 cells. In each box plot, the centre line indicates the median, the box limits represent the upper and lower quartiles, and the whiskers extend to 1.5 × IQR. The number of genes for each group (from left to right) is 5,328, 1,696, 1,540, 1,506, 1,342, 1,195, 1,981, 2,004 and 3,056, respectively. b, Expression breadth (that is, the number of tissues in which a gene is expressed) for genes with different numbers of interacting enhancers in H1 cells. c, The percentage of housekeeping genes with different numbers of interacting enhancers (HRT Atlas v1.0). d, Genome browser view of the region surrounding the housekeeping gene EIF1. The blue arcs represent chromatin loops linking the EIF1 promoter with distal enhancers. e, Dynamics of chromatin loops linking housekeeping gene promoters and distal enhancers between H1 cells and HFFc6. f, Genome browser view of the CMAS locus in H1 cells. g, Lamin-B1 DamID-seq signals surrounding lamina-associated genes and their interacting enhancers in H1 cells. TES, transcription end site.

Fig. 7. A/B compartments and SPIN states represent subnuclear regions of distinct replication timing and gene expression.

a, Schematic of human genome folding into A/B compartments, SPIN states, TADs, subTADs and loops integrated with early/late replication timing and IZs. b, SPIN states were classified as either fully embedded within A/B compartments (within), co-registering A/B compartments (co-register) or partially overlapping (other). c, The fraction of each SPIN state co-registered or nested within A/B compartments in H1 cells. d, The average chromatin landscape at IZs in H1 cells. IZs have been grouped depending on their replication timing (RT). The tracks represent the high-resolution replication timing, chromatin compartments, expression and histone marks. e, We computed right-tailed, one-tailed empirical P values using a resampling test with size and A/B compartment-matched null IZs for the intersection of early and late S phase IZs with dot boundaries, dotless boundaries and no boundaries. f, An example of chromatin profiles around IZs (portion of chromosome 2 from 20 Mb to 58 Mb). The tracks represent the chromatin contacts, four groups of IZs depending on their replication timing, the high-resolution replication timing, chromatin compartments, the SPIN states, expression (minus and plus strands), H3K27Ac, H3K4me3 and H2AX.

Extended Data Fig. 1. Chromatin interaction assays quantitatively differ in detection of small compartment domains.

a. Eigenvector 1 obtained from SPRITE data derived from a range of cluster sizes along a typical genomic loci, showing that small compartments are not detected when data from larger SPRITE clusters is used. b. Examples of Eigenvector 1 profiles obtained from data generated with the different genomic assays indicated. c. Compartmentalization strength calculated with interaction data obtained with different SPRITE clusters, for H1 and HFFc6. e,f. Cumulative distributions of compartment sizes, as detected with Hi-C, Micro-C, ChIA PET, PLAC Seq, and SPRITE (cluster size 2–100) for H1 and HFFc6 cells.

Extended Data Fig. 2. Aggregate peak analysis of different chromatin loop sets across platforms.

Contact maps at 25-kb resolution were used. a. High-confidence loops detected by at least two platforms in each cell line. b. Pol2 ChIA-PET loops in each cell line. c. Loops uniquely detected by Pol2 ChIA-PET. d. Random control loops for Pol2 ChIA-PET-unique loops. For each Pol2 ChIA-PET-unique loop, control loops are defined by considering interactions between each anchor and loci positioned at the same genomic distance on the opposite side.

Extended Data Fig. 3. Cross-platform comparisons of chromatin loops.

a. Number of chromatin loops detected by each platform in each cell line. b. Cross-platform comparison of loop anchors in each cell line. For each cell line, the top UpSet plot summarizes the overlap of loop anchors detected by different platforms, and the bottom heatmap shows the fold enrichment of ChromHMM states for each anchor category defined in the UpSet plot. c. UMAP projection and clustering of 146,140 union loops in HFFc6 based on the ChromHMM state composition at loop anchors. Cluster IDs and loop counts are labelled on the UMAP. d. Fold enrichment of the top three chromatin state pairs for each cluster in HFFc6. e. Violin plots showing the loop size distribution across clusters in HFFc6. Within each violin, the black dot represents the median, and the vertical line represents 1.5 times the interquartile range. The number of loops (n) in each cluster is indicated on the UMAP in panel c. f. UMAP projection of chromatin loops, with loops detected by each indicated platform highlighted in red in separate columns. g. Cluster composition of loops detected by each platform in HFFc6. Colours match those used in panel e.

Extended Data Fig. 4. Transcription factor (TF) binding signatures of different loop clusters in H1 cells.

a. ChIP-Seq binding profiles of selected TFs around both anchors of loops from different clusters. Each row represents one loop. b. Fraction of loop anchors bound versus fold enrichment for 62 TFs. c,d. Chromatin state and TF binding differences between CTCF-bound and CTCF-unbound loops in clusters 3–6. c, Fold enrichment of different ChromHMM states at CTCF-bound versus CTCF-unbound loop anchors. d, Fold enrichment of the top 12 TFs (excluding CTCF) with the highest enrichment scores at CTCF-bound versus CTCF-unbound loop anchors. Values above each bar indicate the percentage of loop anchors overlapping the corresponding TF binding peaks.

Extended Data Fig. 5

a. Box plots show the distribution of normalized TSA-seq and DamID scores on distinct SPIN states in H1 and HFFc6 cell lines. The number of 25 kb bins in H1 (from left to right) is 11,073, 15,446, 10,002, 18,327, 8,771, 5,129, 6,705, 81,96, and 29,226, respectively. The number of 25 kb bins in HFFc6 (from left to right) is 9,693, 14,111, 9,032, 15,749, 9,018, 18,411, 8,249, 7,900, and 20,712, respectively. Box plot shows median (centre line), interquartile range (IQR; box), and whiskers extending 1.5×IQR. b. A confusion matrix shows the comparison of SPIN states in K562 between the previous version and the new result based on the joint modelling across four cell lines. Note that the new result is based on a new nucleolus TSA-seq data. The numbers in the heatmap indicate the number of 25 kb bins. c. The differences of the distributions of normalized TSA-seq and DamID between any two pairs of SPIN states are tested by the two-sided Wilcoxon rank sum test. Colours of heatmap indicate the p-value under the null hypothesis that two distributions are derived from the same population. P-value is classified as p ≤ 1e-10, 1e-10 <p ≤ 1e-5, 1e-5 <p ≤ 0.05, or p > 0.05.

Extended Data Fig. 6. 3D Structure-function relationship and its cell-to-cell variability.

a. SPIN states define spatial genome compartments. Heatmaps show enrichment of caRNAs and histone marks (columns) across SPIN states (rows) in HFFc6. Fold-change is calculated as the ratio of observed signals over genome-wide expectation. b. Single-cell genome structure of H1 with regions coloured by SPIN state. c. Slice of the structure in (a) showing a subset of chromosomes and predicted speckle locations (red spheres). d. Log2-enrichment of structural features across SPIN states from model populations over genome-wide expectation. Structure features as defined in Fig. 4f. e. Box plots of average radial positions of chromatin by SPIN state in HFFc6. Box plots show median (centre line), interquartile range (IQR; box), and whiskers extending 1.5×IQR (n = 7 H1, 5 HFFc6). f. Box plots of speckle association frequencies (SAF) by SPIN state in HFFc6. Box plots as in (e). g. Violin plots of speckle distance z-score differences (HFFc6 - H1) for genes with significant changes in expression (| log₂ fold-change | > 9, FDR < 0.05). The percentages and numbers in brackets indicate the fraction and number of genes with significantly increased (top) or decreased (bottom) speckle-distance z-scores in HFFc6 relative to H1 (FDR 0.05). Box plots within the violin show the median (centre line), interquartile range (box), and whiskers extending to 1.5X the interquartile range (n = 121, 37 for down-regulated and up-regulated genes). h. Log₂-fold enrichment of 14 structural features for POU3F1, THBS1 and ACTB genes in H1 and HFFc6. i. (left) ACTB gene expression in H1 and HFFc6. (Right) Joint nearest speckle and lamina distance distribution for ACTB in 1,000 single-cell models. Box plots show the median (centre line), interquartile range (box), and whiskers extending to 1.5X the interquartile range (n = 7 H1, 5 HFF). j. Speckle distance (SpD) distributions of genes across gene expression categories in HFFc6. Within each violin, the white dot represents the median, and the vertical black line represents interquartile range. Pairwise group differences were assessed using two-sided independent-samples t-tests (Student’s t-test; no multiple-testing correction applied). Significance is indicated as ns (p > 0.05), * (p ≤ 0.05), ** (p ≤ 0.01), *** (p ≤ 0.001), and **** (p ≤ 0.0001). k. Gene expression distributions for housekeeping vs. non-housekeeping genes stratified by SAF. Box plots show the median (centre line), interquartile range (box), and whiskers extending to the minimum and maximum data points within 1.5X the interquartile range; individual points represent outliers. l. Representative HFFc6 genome structure with Class I (red) and Class II (blue) highly expressed genes. A representative single 3D genome structure of HFFc6 showing the locations of all highly expressed class I genes (red) and class II genes (blue). m. t-SNE of TSS-containing regions based on 3D structure features for 25% highest and lowest expressed genes and gene-less regions in HFFc6. Counts of total (and housekeeping) genes are indicated. n. Local genome architecture near TSS of Class I gene LMNA and Class II gene FBN2. Tracks (top to bottom): Hi-C (10 kb bins, KR norm. O/E), SON-TSA-seq, Lamina B1 DamID, H3K4me1, H3K4me3, H3K27ac, plus/minus strand RNA-seq. o. Schematic illustration explaining reduced long-range contact contributions for Class I vs. increased long-range interactions for Class II genes.

Extended Data Fig. 7. Housekeeping genes are engaged in extensive enhancer-promoter loops.

All boxplots in this figure follow the same format: the centre line indicates the median, the box limits represent the upper and lower quartiles, and the whiskers extend to 1.5 times the interquartile range. P-values were calculated using a two-sided Mann-Whitney U test. a. Size distributions of enhancer-promoter (EP) loops (n = 79,708) and CTCF-mediated loops (n = 99,447). Data from H1 and HFFc6 were combined. b. Gene transcription levels versus the number of interacting enhancers in HFFc6. The number of genes for each group (from left to right) is 6,846, 1,626, 1,505, 1,354, 1,276, 1,164, 1,833, 1,894, and 2,150, respectively. c. Fold change in transcription levels for genes with more enhancers in H1 (n = 7,109), the same number of enhancers in both cell types (n = 3,147), or more enhancers in HFFc6 (n = 4,535). d. Expression breadth of genes with different numbers of interacting enhancers in HFFc6. e. Distance-normalized contact signals between housekeeping gene promoters and ±30 kb regions surrounding their interacting enhancers. f. Comparison of enhancer counts between two housekeeping gene classes defined in the 3D modelling section. The number of enhancers in class I and class II genes is 1,262 and 86 in H1, and 1,259 and 79 in HFFc6, respectively. g. Number of loop anchors linked to housekeeping genes, categorized by whether they contain enhancers, promoters, both, or neither. h,i. Aggregate peak analysis of cell-type-specific loops connecting distal enhancers to housekeeping gene promoters across platforms at 5-kb resolution. Panels h and i show the results for loops uniquely detected in H1 and HFFc6, respectively. j. Expression breadth across 32 samples for genes grouped by enhancer count percentiles in each sample. k. Enrichment of housekeeping genes across gene sets defined by both the number of interacting enhancers and the number of supporting samples. Each bin represents a specific combination of these two factors. For example, the top-right bin corresponds to genes with more than the 90^th percentile of interacting enhancers in over 10 samples.

Extended Data Fig. 8. Enhancer-promoter loops within the nuclear lamina and their relationship to gene regulation.

All boxplots in this figure follow the same format: the centre line indicates the median, the box limits represent the upper and lower quartiles, and the whiskers extend to 1.5 times the interquartile range. P values were calculated using a two-sided Mann-Whitney U test. a,b. Distributions of the number of interacting enhancers for genes in different SPIN states in H1 (a) and HFFc6 (b). The number of genes in each group (from left to right) is 5,092, 2,626, 2,966, 1,769, 1,743, 2,450, 975, 928, and 965 for H1, and 5,250, 1,770, 4,521, 1,792, 1,540, 3,145, 303, 714, and 479 for HFFc6, respectively. c. Comparison of transcription levels between genes with or without interacting enhancers in the Lamina SPIN state. The number of genes with and without interacting enhancers is 421 and 544 in H1, and 108 and 371 in HFFc6, respectively. d. Comparison of the number of linked enhancers between active and inactive genes in the Lamina SPIN state. The number of active and inactive genes is 337 and 628 in H1, and 74 and 405 in HFFc6, respectively. e. Example genome browser views showing that expressed genes and their interacting enhancers are often synergistically looped out of the nuclear lamina to facilitate gene regulation. Blue arcs represent chromatin loops linking the gene at the centre of each region to distal enhancers. f. Lamin-B1 DamID-seq signals surrounding lamina-associated genes and their interacting enhancers in HFFc6. TSS, transcription start site; TES, transcription end site; TPM, transcripts per kilobase million.

Extended Data Fig. 9

Compartment and SPIN integration with replication timing, RNA-seq, and nascent transcripts from iMargi. a, b. Averaged Hi-C, replication timing (16 fraction Repli-seq), nascent transcription (iMargi), and mRNA levels (RNA-seq) for h1ESCs at all A/B compartments (column 1) and SPIN states either co-registered or co-localized within A/B compartments (columns 2–10). All genomics data is plotted as the average signal across all genomic intervals representing SPINs in a particular column. SPIN genomic intervals of (SPIN genomic interval +/− flanks of 60% of the size of the genomic interval) are stretched laterally to scale by size before average signal is computed. a. All A compartments or select SPINS co-registered or within compartment A and b. All B compartments or select SPINS co- registered or within compartment B. Tracks show pileups in H1 for Hi-C Aggregate-Peak- Analysis (APA), A/B compartment, 16 fraction Repli-seq, median RNA-seq signal, condensed RNA-seq reads, median averaged iMARGI (+) and iMARGI (−) signal, condensed iMARGI (+) and iMARGI (−) reads, median iMARGI (+) signal, condensed iMARGI (+) reads, median iMARGI (−) signal, and condensed iMARGI (−) reads.

Extended Data Fig. 10. Predicting the effect of genomic variants on 3D genome folding with deep learning.

a. Example of a 345 bp deletion (chr1: 47262830-47263175) at the TAL1 locus. We trained a model on HFFc6 Micro-C using the Akita architecture and predicted contact maps for the reference human genome sequence (WT; ~1 Mb region) and the sequence carrying the deletion (Mut). Red: higher than expected interaction frequencies given genomic distance (log(observed/expected)); Blue: lower than expected. The effect of the deletion (Mut - WT) is plotted below. Purple: increased chromatin interactions; Green: decreased. Genes in the locus are plotted below the contact maps with TAL1 highlighted in red. The deleted region has a CTCF binding site and is located in a TAD boundary. Mirroring the experimental deletion in HEK293T cells (Hnisz et al.), the HFFc6 model predicts increased contact frequency between TAL1 and adjacent regions (black rectangle), as did an H1 model (data not shown). b. In silico mutation of transcription factor motifs (replacing motifs with random sequences) affects deep learning predictions of nearby chromatin interactions. An example POU2F1::SOX2 motif (left, chr13: 81872756-81872772) and FOSL1::JUND motif (right, chr16: 12340569-12340578) were generated using models with the Akita architecture trained on H1 or HFFc6 Micro-C data, respectively. Motif logos generated via model importance scores using DeepExplainer are shown below the maps. These resemble the canonical motifs, though not precisely. Colour scales are the same as in (a), and motif sites are centred on the contact maps. Star symbols indicate regions with altered chromatin interaction predictions.