GraphChIAr: genome-wide super-resolution reconstruction of protein-mediated remote chromatin interactions by augmenting hi-C interaction maps with multiple ChIP-seq profiles Authors: Hao Dong, Guo-Zheng Rao, Hao-Yu Wang, Xin-Ran Wu, Tong Xian, Bo-Qiang Wang, Pu-Feng Du Published: 2026 ABSTRACT AbstractProtein-mediated chromatin interactions are fundamental to gene regulation. However, experimental approaches such as Chromatin Interaction Analysis by Paired-End Tag sequencing are limited by data scarcity and high cost, while existing computational models are constrained by limited resolution and challenges in effectively integrating heterogeneous genomic data. To address these issues, we propose GraphChIAr, a regression-based deep learning framework that estimates chromatin interaction strength by augmenting Hi-C contact maps with various ChIP-seq profiles and genomic sequence information. A key advantage of GraphChIAr is its super-resolution capability, enabling accurate estimations of chromatin interactions from conventional resolutions down to ultra-high near-nucleosome resolution (e.g. 200 bp). By introducing genomic shift distance in GraphChIAr, we enabled it to predict remote interactions between distant genomic loci at a genome-wide scale. Cross-referencing results demonstrate high predictive accuracy for key mediating proteins such as CTCF, highlighting the benefits of integrating complementary genomic features. Together, GraphChIAr provides an effective computational tool to augment experimental data and advance the study of 3D genome organization. The source code of GraphChIAr is available at (https://github.com/don194/GraphChIAr).Keywords: chromatin interaction, ChIA-PET, ChIP-seq, super-resolution Introduction Chromatin is intricately organized within the nucleus in a 3D structure that governs essential cellular processes [1, 2]. The spatial organization of the genome facilitates interactions between distal regulatory elements and their target genes, which is crucial for proper gene expression regulation [3–6]. Understanding these complex 3D interactions is essential for deciphering gene regulation mechanisms and their implications in development and disease [7]. Over the past decade, several experimental techniques have been developed to investigate 3D genome organization. Hi-C [8, 9] and the higher-resolution Micro-C [10] have enabled genome-wide mapping of chromatin contacts, revealing hierarchical structures like compartments [11] and topologically associating domains (TADs) [12–14]. Chromatin Interaction Analysis by Paired-End Tag sequencing (ChIA-PET) [15], on the other hand, captures chromatin interactions mediated by specific proteins of interest, such as CTCF, RAD21, or RNA Polymerase II. In contrast to Hi-C, where a comprehensive view of all potential interactions is offered, ChIA-PET offers valuable insights into protein-specific chromatin loops that often have functional significance in gene regulation [16]. ChIP-Seq (Chromatin immunoprecipitation followed by high-throughput sequencing) is widely used for analyzing interactions between DNA and chromatin-associated proteins, providing crucial information about protein binding sites [17–22]. In general, ChIA-PET can be viewed as a hybrid of Hi-C and ChIP-seq [23]. However, ChIA-PET experiments are labor-intensive, time-consuming, and expensive [24]. The scarcity of publicly available ChIA-PET datasets is evident in the NCBI GEO [25] database. There are only 999 ChIA-PET experimental sets in the NCBI GEO database as of March, 2025. However, there are 10 128 Hi-C and 105 544 ChIP-seq sets at the same time. Therefore, artificially synthesizing ChIA-PET interactions from discrete Hi-C and ChIP-Seq data offers a promising strategy to expand the available ChIA-PET datasets [23]. Numerous computational methods have been developed to predict CTCF-mediated chromatin loops using different types of features, including DNA sequence-based approaches such as CLNN-loop [26], CTCF-MP [27], and DeepMILO [28], methods integrating genomic and epigenomic features such as Lollipop [29], and approaches combining CTCF binding and Hi-C data, such as LOOPbit [30]. However, these methods are not designed to learn and predict ChIA-PET interactions at a genome-wide level. DeepChIA-PET [23] partially addressed this limitation by integrating Hi-C and ChIP-seq data to predict ChIA-PET interactions. Nevertheless, it formulates the task as a pixel-wise binary classification problem on a contact matrix, which limits its ability to capture the continuous nature of chromatin interaction strengths. ChIPr [31], on the other hand, models interaction prediction as a regression task to estimate interaction strength. However, it operates on predefined pairs of genomic regions, making it difficult to generalize to genome-wide prediction. Recent advances in Hi-C prediction have demonstrated the effectiveness of integrating multi-modal data and modeling chromatin interactions across multiple genomic scales [32–41]. However, these approaches primarily focus on aggregate chromatin contacts and do not explicitly model protein-mediated interactions. These observations highlight the need for a computational framework that, similar to Hi-C prediction methods, integrates diverse sequence and epigenomic features to predict protein-specific ChIA-PET interaction strengths across the genome in a regression-based manner. To this end, we present GraphChIAr, a novel regression model that integrates genomic sequence features, ChIP-seq profiles, and Hi-C contact maps to synthesize ChIA-PET interaction maps digitally. Our approach constructs a genome interaction network that aggregates multi-modal features across the genome to estimate ChIA-PET interaction strength values. Our GraphChIAr significantly elevated the resolution of a contact map from conventional resolution to nucleosome-level resolution. Our method provides a valuable alternative approach to investigate high-resolution genome 3D conformation in the context that experimental ChIA-PET profiles are limited or unavailable. This also enables scientists to acquire high-resolution chromatin interaction maps with limited resources. Materials and methods Data preprocessing Hi-C and micro-C data for multiple cell types (GM12878, IMR90, and H1-hESC) were downloaded from the 4DN data portal [42, 43]. ChIA-PET datasets were downloaded from the ENCODE project [44, 45]. HiChIP datasets were downloaded from the NCBI Gene Expression Omnibus (GEO) database [25]. For the 200 bp resolution data, BAM files were downloaded and processed using the bam2pairs and juicer tools [46] to generate Hi-C contact matrix files in hic format. The complete list of all Hi-C, ChIA-PET, and micro-C datasets in this study is provided in Table S1 in the supplementary, with their sources and accession numbers. We extracted Hi-C contact matrices using the hic-straw [47] library. Four resolutions were tested in this study, including 10 kb, 5 kb, 1 kb, and 200 bp. Hi-C, micro-C, and ChIA-PET samples were extracted along the diagonal of a 2D contact matrix using a sliding window approach. For the 10 kb resolution, a window size of 200 × 200 bins with a step size of 50 bins was used. At 5 kb resolution, the window size was increased to 400 × 400 bins, with a step size of 100 bins. To accommodate GPU memory constraints while maintaining high resolution, a window size of 500 × 500 bins with a step size of 250 bins was used during training at the 1 kb resolution. During testing, the step size was reduced to 125 bins for more granular predictions. For the ultra-high 200 bp resolution, which is also computationally intensive, we used a 250 kb window (1250 × 1250 bins) with a 500 kb step. Unlike the conventional-resolution settings, the 200 bp prediction was formulated as a super-resolution task, in which 1 kb Hi-C data were used as input to predict chromatin interactions at 200 bp resolution. This design enables GraphChIAr to generate near-nucleosome-level interaction maps from lower-resolution Hi-C data. Epigenomic data, including ChIP-seq profiles for various histone modifications and transcription factors, were obtained from the ENCODE project [44, 45]. All datasets were downloaded as BAM files. The replicates for each epigenomic track were merged using pysam (https://github.com/pysam-developers/pysam), a Python wrapper for samtools [48]. The merged BAM files were then converted to BigWig format using bamCoverage from the deepTools [49] suite with a bin size of 10 bp and reads per genomic content (RPGC) normalization. A log(x + 1) transformation was applied on all normalized values in BigWig files before these values were used as inputs to the model. The complete list of ChIP-seq datasets used in this study is provided in Table S2 in the supplementary. Model architecture GraphChIAr is implemented in PyTorch and consists of four main components, as illustrated in Fig. 1, a feature extraction module, a DIFFormer aggregation module, a spatial pattern extraction module, and a prediction module. The architecture is designed to flexibly handle two distinct prediction tasks: standard intra-region prediction (within a single genomic window) and long-range inter-region prediction (between two distinct genomic regions, separated by a genomic distance). To enable genome-wide predictions, we introduce the offset mechanism, which allows the model to predict interactions between non-overlapping genomic regions separated by a specified genomic distance. Unlike traditional methods that focus only on intra-region interactions along the main diagonal, the offset mechanism shifts the model’s focus to other diagonals of the interaction matrix, enabling the prediction of long-range interactions. This approach enables the prediction of chromatin interactions across the entire genome at multiple resolutions. The architecture of the GraphChIAr model for ChIA-PET prediction. The model consists of four main components: (1) feature extraction module that processes DNA sequence and ChIP-seq data through convolutional layers with PE; (2) spatial pattern extraction module that captures chromatin interaction patterns from hi-C matrices; (3) graph node aggregation module using DIFFormer to aggregate node features across genomic bins while leveraging hi-C interaction data as the graph structure; and (4) prediction module that combines these features to generate ChIA-PET interaction matrices. This module employs sequential 2D residual blocks with dilation rates of (r1, r2, r3) = (1, 2, 4) to expand the receptive field. For the ultra-high resolution (200 bp) experiments, these dilated blocks were omitted to reduce computational complexity and memory usage. The DIFFormer component integrates graph convolutional networks with global attention mechanisms to capture both local and long-range genomic dependencies. Networks with global attention mechanisms to capture both local and long-range genomic dependencies. Feature extraction The model inputs consist of a matrix of c epigenomic signals, Xepi∈Rl × c, which is defined over a genomic window of length l. Depending on the specific configuration (e.g. for CTCF prediction), a one-hot encoded DNA sequence (Xseq∈Rl × 5) may also be included. These available features are concatenated along the feature axis and transposed to form the input tensor X(0), where the feature dimension reflects the number of included data tracks (i.e. c or c + 5). This tensor is processed by a series of 1D Residual Blocks to learn genomic bin representations. The r-th (r = 1, 2, …) ResBlock first down-samples its input feature map X(r - 1) using a 1D convolution with stride 2 (W0(r)). The result (H(r)) is then passed through a residual path containing two standard 1D convolutional layers (kernel size 3, stride 1, W1(r) and W2(r)). All three convolutional layers above are followed by a batch normalization (BN0(r), BN1(r) and BN2(r)). The first two batch normalizations are followed by a ReLU activation (σ). The block’s output is computed via a residual connection. The block can be formalized as follows: and where ‘*’ denotes the 1D convolution operation. Following each ResBlock, a 1D max-pooling layer with a kernel size of 5 and stride of 5 further reduces the sequence length. The number of convolutional blocks is dynamically determined based on the target number of genomic bins. Each full block reduces the sequence length by an approximate factor of 10, as a result of the combination of a stride-2 convolution and a max pooling operation. To produce a final n × n contact matrix, we introduce a reduction scale parameter α = l / n. The model automatically infers the number of required blocks to adapt α. If the exact reduction cannot be achieved using full blocks alone, the final block adaptively adjusts the down-sampling strategy by either setting the initial convolution stride to 1 (retaining only the max-pooling for a five-fold reduction) or omitting the max-pooling layer (retaining only the stride-2 convolution for a two-fold reduction) to ensure that the output length matches the contact matrix size. To preserve positional information, a fixed sinusoidal positional encoding (PE) is concatenated to the channel dimension of the final convolutional output. The PE for the i-th channel is defined as follows: and where pos is the genomic bin index, φ = 2i/hPE, and hPE is the number of PE channels. For intra-region prediction, pos is the relative index within the region (pos = 0, 1, …, n-1). For inter-region predictions, we use the genomic position in an expanded region. Suppose that we have two regions A and B, we first obtain two genomic bins’ relative index, respectively in their own region (posA = 0, 1, …, n-1 and posB = 0, 1, …, n-1). We convert posB to be relative to region A, by adding an offset δ to posB, where δ is the genomic distance between the two regions by bins. Concatenating PE with the convolutional output, denoted as X(final), yields the final representation where PE represents the PE matrix, and h = d + hPE denoting the combined channel dimension passed to downstream layers. Graph node aggregation We model n genomic bins as n nodes in a graph to capture long-range interactions. For inter-region prediction, the graph contains 2n nodes, with n nodes from region A and n nodes from region B. The interactions between these bins are modeled using the DIFFormer framework [50]. Let Zi(t) represents the state of the i-th bin at time t. The dynamic diffusion process in DIFFormer is governed by a partial differential equation. According to literature [50], the continuous time process can be discretized using the forward Euler scheme, with time interval τ. The discretized form of the equation can be represented as follows: where Zi(k) is the state of the i-th bin at step k, S(k) the learnable diffusive matrix, G = D-1/2 A D-1/2, A the adjacency matrix from the Hi-C contact map, and D the diagonal degree matrix of A. For intra-region predictions, A is a n × n matrix. For inter-region predictions, the adjacency matrix A is a 2n × 2n matrix constructed from the corresponding Hi-C contact data across the joint two-region system. The matrix consists of three components: intra-region contacts within region A, intra-region contacts within region B, and cross-region contacts between the two regions. Therefore, both intra-region and inter-region contacts were considered in inter-region predictions. We set τ = 0.5 to controls the rate of information diffusion. To save computational power and to improve computational efficiency, we applied the linear simple diffusive model of DIFFormer, as follows: where and ‖·‖2 the L2 norm. After passing through two consecutive DIFFormer layers, the final node embeddings, which we denote as Zfinal∈Rn × h, are used for node pair construction. For standard intra-region prediction, this step creates a pairwise representation tensor P∈Rn × n × 2h by concatenating the feature vectors for each pair of nodes: pij = [zfinal,i; zfinal,j]. For inter-region prediction, this tensor is constructed only between nodes from different regions, using the corresponding final embeddings zA and zB: pij = [zA,i; zB,j]. Spatial pattern extraction To extract local spatial patterns, the Hi-C contact map (F(0)) is processed by two consecutive convolutional modules (m = 1, 2). For intra-region predictions, the Hi-C contact map is an n × n matrix, which is restricted within the corresponding genomic window. For inter-region predictions, the contact map is still an n × n matrix, representing only contacts between two regions. The m-th convolutional module is defined as follows: where W(m)spatial represents a 3 × 3 convolutional kernel, and ‘*’ the 2D convolution operation. The output of the later module is the spatial feature tensor Fspatial∈Rn × n × h. The number of output channels is set to h to match the dimension of the graph node embeddings, facilitating subsequent feature fusion. Prediction The prediction head is responsible for integrating the graph-derived and spatial features to produce the final interaction map. First, the pairwise node representation P∈Rn × n × 2h and the spatial features Fspatial∈Rn × n × h are concatenated along the channel dimension, resulting in a combined feature tensor Min∈Rn × n × 3h. Subsequently, a linear transformation, implemented as a 1 × 1 convolution, is applied to this tensor. This step serves to both fuse the features from the distinct pathways and adjust the channel dimension from 3 h back to the model’s core dimension h. The resulting tensor, which we denote as M(0)∈Rn × n × h, is then passed to the subsequent residual blocks. Next, this tensor is processed by a stack of three sequential 2- Residual Blocks, indexed by j∈{1, 2, 3}. The operations within the j-th block are defined as: and where M(0) is the input to the first block, and * denotes the 2- convolution operation. Finally, a lightweight three-layer Multilayer Perceptron (MLP) acts as the final prediction head. It takes the output of the last 2D Residual Block (M(3)) as input to map the feature vector at each bin-pair position (i, j) to a single scalar value. This process yields the final prediction of the interaction map (Y′∈Rn × n). For computational efficiency, this MLP is implemented as a stack of three 1 × 1 convolutional layers. A dropout rate of 0.1 is applied throughout the network for regularization. The core hidden dimension for the entire model is h = 32. Model training and prediction We optimized the model using the mean squared error (MSE) loss. For intra-region predictions, the loss focuses on the upper triangular portion of the interaction matrix, excluding the main diagonal. However, for inter-region predictions involving the offset mechanism, the loss is computed over the entire n × n matrix. Specifically, for a predicted matrix Y′ and ground truth matrix Y, the loss L is calculated as: where yij and y’ij denote the experimental and predicted interaction values, n the window size (number of genomic bins), and ntri the number of elements in the upper triangular matrix excluding the main diagonal. All training procedures for GraphChIAr and DeepChIA-PET were performed on a workstation equipped with a single NVIDIA GeForce RTX 4060 Ti GPU with 16 GB memory. C.Origami and HiCDiffusion were trained on a single NVIDIA GeForce RTX 5090 GPU due to their higher computational power requirements. We trained GraphChIAr using a chromosome-based data partitioning strategy, as in previous studies [35]. Specifically, the training set comprised all chromosomes except chromosomes 5, 10, and Y, with chromosomes 5 and 10 held out for validation and testing, respectively [35]. To accommodate different data resolutions, we used resolution-specific batch sizes: 32 for 10 kb, 16 for 5 kb and 1 kb, and 4 for 200 bp resolution data. Model training was performed using the Adam optimizer with an initial learning rate of 10−4, coupled with a learning rate scheduler that applied linear warmup followed by cosine annealing. We employed early stopping with a patience of 10 epochs, and training was capped at a maximum of 80 epochs. The model checkpoint with the lowest validation loss was retained for inference. During prediction, the model processes the test chromosomes using the same sliding window approach as in training. This generates multiple overlapping prediction matrices for the targeted regions. To construct a single, coherent map for evaluation, the final value for any bin pair located in an overlapping area is calculated as the mean of all predictions covering that pair. All performance metrics were then computed by comparing this final, averaged prediction map against the corresponding regions of the ground truth experimental matrix. For cross-cell-type evaluation, models trained on one cell line were applied to data from other cell lines to assess the model’s capacity to capture both conserved and cell-type-specific chromatin interaction patterns. Performance evaluation We primarily evaluated model performance using two categories of correlation-based metrics that assess the concordance and structural similarity between the predicted Y′ matrix and the experimental Y matrix. First, to provide a global assessment of concordance between the predicted matrix Y′ and the experimental matrix Y, we calculated the Pearson correlation coefficient (PCC). The metric is computed over the upper triangular portion of the interaction matrix (excluding the main diagonal) as follows: where yij and y’ij denote the experimental and predicted interaction values, and ’ present their corresponding mean values within the upper triangular region. We calculated this metric in two ways: an overall PCC using all bin pairs, and a non-zero PCC, where the calculation was restricted to bin pairs with a non-zero experimental value yij > 0. Second, we employed the Stratum-adjusted Correlation Coefficient (SCC) [24] to evaluate structural similarity between predicted and experimental contact maps. SCC calculates Pearson correlation coefficients at each genomic distance d: where yij and y’ij denote the experimental and predicted interaction values, Dd represents interactions at distance d, and d and ’d are the mean values of contacts at this distance. Baseline comparisons with existing methods To comprehensively evaluate the performance of our model, we compared GraphChIAr against several representative baseline methods: DeepChIA-PET [23], CAESAR [40], HiCDiffusion [41], and C.Origami [35], all of which have been used to predict chromatin interactions. For DeepChIA-PET, a state-of-the-art method for ChIA-PET interaction prediction, we adapted its architecture to perform regression rather than binary classification by replacing the final classification layer with a regression output and changing the loss function from binary cross-entropy to MSE. Following its original design, DeepChIA-PET utilized Hi-C contact maps and ChIP-seq signals as input features. To ensure a fair comparison, we maintained the original residual network structure while adjusting the number of hidden units according to the resolution: 32 units for the 10 kb resolution and 16 units for the more memory-intensive 5 kb resolution. We also matched training conditions to those used in GraphChIAr, with batch sizes of 16 and 8 for 10 kb and 5 kb data, respectively. Furthermore, we evaluated two additional state-of-the-art models originally developed for predicting Hi-C contact maps: C.Origami and HiCDiffusion. C.Origami is an integrative framework that utilizes multiple genomic signals as inputs; in our study, it was configured with the same set of genomic features as GraphChIAr. In contrast, HiCDiffusion is a diffusion-based generative model designed as a sequence-to-map framework and was thus trained using pure DNA sequence data as its primary input according to its original design. To ensure a fair and computationally efficient comparison, both models were implemented using their official source code with a consistent hidden layer dimension of 256. For CAESAR, since the original implementation only supports 200 bp resolution, we reimplemented the contact profile model in PyTorch to accommodate our 1 kb resolution data. Our reimplementation follows the architectural principles described in the original publication, incorporating one-dimensional convolutional layers to extract local epigenomic features along the chromatin sequence and graph convolutional layers to model spatial interactions over the chromatin contact graph. For both CAESAR and our model, we used the same set of epigenomic features corresponding to each specific task as input signals. Specifically, our CAESAR implementation included two graph convolutional layers and two 1D convolutional layers, each with 32 hidden units. This adapted model serves as a baseline to assess the high-resolution predictive capability of GraphChIAr. Shapley-value-based input attribution analysis To assess the importance of different biological input sources, we performed an input-level attribution analysis based on exact Shapley values. For each task, the trained full model was fixed and evaluated under all possible combinations of input sources: CTCF ChIP-seq and DNA sequence for CTCF, RAD21 and SMC3 ChIP-seq for RAD21, and POLR2A, H3K27ac, and H3K4me3 ChIP-seq signals for POLR2A. Inputs excluded from a given combination were masked at inference time by setting the corresponding channels to zero. For each input subset S, the utility function v(S) was defined as the negative mean-squared error between predicted and observed contact maps on the held-out evaluation set. The Shapley value of each input source i was computed as: where M denotes the full set of modalities. Shapley values were normalized by the sum of absolute contributions to obtain relative importance scores. Aggregate peak analysis Aggregate Peak Analysis (APA) was performed using juicer tools [46] to evaluate the focal enrichment of chromatin interactions. To obtain high-resolution anchor centers, the pre-called loop lists were filtered to retain only those interactions whose anchors overlapped with ChIP-seq summits, ensuring spatial alignment at a fine scale. The resulting aggregated matrices were used to calculate APA scores, including the Peak-to-Lower-Left (P2LL) ratio, to quantify the signal enrichment at the center of the loops relative to the background regions. Results GraphChIAr accurately predicts chromatin interactions across conventional resolutions We benchmarked GraphChIAr on the GM12878 data against DeepChIA-PET, HiCDiffusion, and C.Origami. All models were trained on the GM12878 dataset with input features configured according to the architecture of each method (detailed in Section 2.5). For GraphChIAr, CTCF predictions utilized sequence information and CTCF ChIP-seq signals, Cohesin-associated interactions used RAD21 and SMC3 data, and transcription-associated interactions combined POLR2A ChIP-seq with H3K27ac and H3K4me3 signals. We assessed performance using Pearson correlation, Non-zero Pearson correlation, and stratum-adjusted Pearson correlation coefficients (SCC). As in Fig. 2, GraphChIAr consistently demonstrated superior predictive accuracy. In direct comparisons of correlation coefficients for CTCF, RAD21, and POLR2A at both 10 and 5 kb resolutions, GraphChIAr outperformed DeepChIA-PET, C.Origami, and HiCDiffusion (Fig. 2A). Moreover, for all tested proteins, GraphChIAr consistently achieved higher stratum-adjusted Pearson correlation coefficients across genomic distances compared to the baselines (Fig. 2B). The high predictive performance indicated by these metrics is visually corroborated by the strong agreement between GraphChIAr’s predictions and the experimental data (Fig. 2C and Fig. S1). To further evaluate generalizability, we performed cross-protocol validation on HiChIP datasets. GraphChIAr yielded consistent and robust predictive results, demonstrating its reliability across different 3D genomic experimental platforms (Table S3 and Fig. S2). Taken together, these results indicate that GraphChIAr provides accurate predictions of chromatin interactions across different target factors and resolutions. The consistent performance improvements observed for CTCF, RAD21, and POLR2A highlight the robustness and flexibility of the framework. Overall performance of GraphChIAr in predicting ChIA-PET interaction matrices compared to DeepChIA-PET. (A) Bar charts comparing the overall Pearson correlation (top) and nonzero Pearson correlation (bottom) between GraphChIAr and three methods: DeepChIA-PET, C.Origami, and HiCDiffusion. Performance is evaluated across CTCF, RAD21, and POLR2A targets at 10 kb and 5 kb resolutions. (B) Relationship between stratum-adjusted Pearson correlation coefficients and genomic distance (up to 2 Mb). GraphChIAr consistently maintains higher correlation across various genomic distances compared to other models. (C) Visualization of predicted versus experimental ChIA-PET interaction maps for a representative region on chromosome 10 (21.0–23.0 Mb). The top panel shows the GraphChIAr prediction, the middle panel shows the experimental ground truth, and the bottom track displays the corresponding CTCF ChIP-seq binding profile, illustrating the model’s ability to recover fine-grained spatial interactions. Integration of complementary features improves predictive performance To assess the contribution of complementary data sources beyond the primary target protein’s binding signal, we conducted ablation experiments for CTCF, RAD21, and POLR2A ChIA-PET predictions at both 10 and 5 kb resolutions (Fig. 3). We compared a baseline ablated model, which relied solely on the target protein’s ChIP-seq signal, with the full model, which additionally incorporated complementary features: DNA sequence information for CTCF, SMC3 ChIP-seq data for RAD21, and H3K27ac/H3K4me3 histone modification signals for POLR2A. Performance comparison between the full GraphChIAr model and an ablated model. (A) Bar charts comparing model performance. The top panel compares the overall Pearson correlation between GraphChIAr and DeepChIA-PET, while the bottom panel compares the nonzero Pearson correlation between the full model and the ablated model for CTCF, RAD21, and POLR2A at 10 and 5 kb resolutions. (B) Distance-dependent Pearson correlation curves for the full and ablated models across all protein targets and resolutions. The performance gap confirms the necessity of synergistic feature representation in capturing complex chromatin loops. (C) Quantitative contribution of input modalities determined by Shapley additive explanations. The donut charts illustrate the relative importance of sequence features (one-hot encoding) and epigenetic features (ChIP-seq) for CTCF, RAD21, and POLR2A models. Across all settings, integration of complementary features resulted in measurable performance gains. As shown in Fig. 3A (bottom), the full model consistently achieved a higher Non-zero Pearson correlation than the ablated model. Prediction quality was also evaluated using stratum-adjusted Pearson correlation coefficients (SCC), which showed the full model’s performance was strong and often moderately improved across genomic distances (Fig. 3B). To better understand the improved performance of the full model, we performed a Shapley-value-based analysis to estimate the contribution of each input type (Fig. 3C). The results showed that the primary target signal remained the main contributor for CTCF and RAD21 prediction, while complementary inputs also made substantial contributions. In contrast, POLR2A prediction relied more heavily on complementary epigenetic features, with histone marks contributing nearly half of the total importance. These results indicate that the improved performance of the full model arises from integrating multiple genomic signals rather than relying on a single input source. In summary, the ablation and interpretability analyses show that integration of complementary features enhances GraphChIAr’s predictive performance. The improved concordance with experimental data and the substantial contributions from complementary inputs highlight the importance of incorporating diverse genomic signals in chromatin interaction prediction. GraphChIAr enables cross-cell-type chromatin interaction prediction To evaluate GraphChIAr’s ability to predict chromatin interactions across different cell types, we conducted cross-cell-type evaluation experiments between GM12878 and IMR90 cells. This experiment tests whether our model can capture both conserved features and cell-type-specific chromatin organization patterns. As shown in Figs 4A and 4C, metrics demonstrate the robust generalization of GraphChIAr. The Pearson correlation coefficients for cross-cell-line predictions closely track the intra-cell-line benchmarks across genomic distances. For both IMR90 and GM12878, the cross-cell-line performance remains highly consistent with the results obtained from models trained and tested on the same cell type. The Nonzero Pearson Correlation similarly shows minimal performance decay in cross-cell-line tasks, indicating that the model effectively identifies true interaction regions even in a different cellular context. GraphChIAr demonstrates robust generalization in cross-cell-type prediction. (A) Cross-cell-line performance evaluation. The bar charts compare the Pearson correlation (top) and nonzero Pearson correlation (bottom) for intra-cell-line and cross-cell-line predictions. (B) Chromatin interaction matrices for a 2 Mb region on chromosome 10 (15.5–17.5 Mb) are shown. The left column displays the comparison between predictions and experimental ChIA-PET data in GM12878 cells. The right column illustrates the prediction for IMR90 cells using the model trained on GM12878. Circles highlight cell-type-specific interaction hotspots successfully captured by GraphChIAr. (C) Stratum-adjusted Pearson correlation as a function of genomic distance for IMR90 (left) and GM12878 (right). The nearly overlapping curves for intra- and cross-cell-line tasks underscore the biological robustness of the learned feature representations. Visual analysis further confirms these findings. As shown in Fig. 4B, we analyzed a genomic region on chromosome 10 spanning from 15.5 to 17.5 Mb. In GM12878 cells, our model successfully captured key interaction hotspots that match the experimental ChIA-PET data. Similarly, in IMR90 cells, GraphChIAr accurately predicted cell-type-specific interaction patterns. Particularly, GraphChIAr captures the strong interaction region (circled in Fig. 4B) that appears uniquely in IMR90 cells, even when using the model trained on GM12878 data. These findings indicate that GraphChIAr demonstrates robust performance in cross-cell-type prediction tasks. The model can generalize from training data to capture both conserved and cell-type-specific chromatin interaction patterns in a different cell line. This capability establishes GraphChIAr as a robust tool for predicting interactions in cell lines where experimental data are lacking. GraphChIAr enhances high-resolution chromatin interaction prediction from hi-C data A key advantage of GraphChIAr is its ability to generate chromatin interaction maps at high resolution. To evaluate this capability, we tested whether our model could predict Micro-C interactions at 1 kb resolution in H1-hESC cells. As shown in Fig. 5A, the predicted contact map recapitulates the interaction patterns observed in experimental Micro-C data, while appearing less sparse. For comparison, we reproduced the CAESAR [40] contact profile model as described in its original publication as a baseline. Pearson correlation analysis demonstrates that GraphChIAr consistently outperforms the CAESAR baseline across genomic distances (Fig. 5B, left). GraphChIAr enhances high-resolution and ultra-high-resolution chromatin interaction prediction. (A) Visualization of predicted micro-C contact map (top) compared to experimental micro-C data (middle) at 1 kb resolution in H1-hESC cells. The corresponding CTCF ChIP-seq signal is shown below (bottom). (B) Top: Pearson correlation coefficient as a function of genomic distance for 1 kb micro-C prediction in H1-hESC cells, comparing GraphChIAr to the CAESAR contact profile model. Bottom: Stratum-adjusted Pearson correlation coefficient (SCC) for 1 kb CTCF ChIA-PET prediction in GM12878 cells. (C) Multi-resolution (10 kb, 1 kb, and 200 bp) comparison of CTCF ChIA-PET interactions in GM12878 cells. The top row shows GraphChIAr’s predictions, the middle row shows the experimental ChIA-PET data, and the bottom row displays the corresponding CTCF protein binding sites. Dashed boxes indicate zoomed-in regions from lower to higher resolutions. We also examined the model’s performance in predicting ChIA-PET interactions at 1 kb resolution in GM12878 cells. Visually, GraphChIAr successfully recovered the key interaction patterns for CTCF (Fig. 5C, middle panel). However, the stratum-adjusted Pearson correlation (Fig. 5B, right panel) shows moderate concordance, due to the increased noise and data sparsity inherent to experimental ChIA-PET data at this high resolution. We further explored GraphChIAr’s capability to model chromatin interactions at ultra-high resolution (200 bp), approaching a near-nucleosome precision. To accommodate the increased computational load and memory usage at this resolution, we optimized our GraphChIAr by removing the dilated residual blocks from the prediction head. To generate 200 bp resolution interaction maps, GraphChIAr interpolates the 1 kb resolution Hi-C data to construct a graph-based representation. As shown in the multi-resolution zoom-in process for CTCF (Fig. 5C), the experimental data at this ultra-high resolution appear sparse and visually noisy. In contrast, GraphChIAr provides a sharper contact map visually, with more defined interaction hotspots that align closely with the corresponding protein binding sites (Fig. 5C, bottom track). The corresponding multi-resolution predictions for RAD21 and POLR2A, which show similar patterns, are provided in Fig. S3 in the supplementary. The SCC analyses for the 200 bp resolution are also included in Fig. S4 in the supplementary. To further evaluate the biological relevance of the predictions, we performed APA across multiple proteins and resolutions (Fig. 6). At both 1 kb and 200 bp resolutions, GraphChIAr predictions exhibit clear focal enrichment centered at loop anchors, forming the characteristic cross-shaped interaction patterns observed in experimental ChIA-PET data. In contrast, Hi-C signals alone show weaker and more diffuse enrichment. The P2LL scores of GraphChIAr are substantially higher than Hi-C and approach those of experimental ChIA-PET data across CTCF, RAD21, and POLR2A, indicating that the model effectively reconstructs biologically meaningful protein-mediated interactions at fine scale. Evaluation and application of GraphChIAr predictions. (A) APA of focal interaction enrichment. APA matrices compare hi-C, experimental ChIA-PET, and GraphChIAr predictions at 1 kb and 200 bp resolutions for three regulatory factors (CTCF, RAD21, POLR2A). The number of unique loops (N) used for aggregation is indicated above each panel. Corner APA scores quantify signal enrichment at the center relative to background regions, with P2LL scores shown in the bottom-left corner of each matrix. (B) Application example at the DPPA2/DPPA4 locus (chr3:109.25–109.40 Mb) across three cell types. Predicted RAD21-mediated interactions are shown as arcs (top) and the corresponding RAD21 ChIP-seq signal track (bottom). Arc color intensity represents predicted interaction strength, normalized globally across all three cell types. Candidate interactions were selected from the top-scoring long-range predictions above the 99.5th percentile, with a maximum of 10 arcs displayed per cell type. GM12878 is shown as the training-cell reference with few predicted interactions, H1-hESC shows prominent arcs, and K562 exhibits reduced interactions. DPPA2 and DPPA4 gene positions are indicated below each panel. As a locus-level application example, we further examined the DPPA2–DPPA4 locus, a development- [51] and cancer-related [52] region previously reported to exhibit cell-type-specific RAD21-mediated (cohesin-associated) chromatin loops. Previous RAD21 ChIA-PET analyses showed that this locus is embedded in prominent loops in stem cells, whereas these interactions are absent in GM12878 and attenuated in K562 [53]. Using the GM12878-trained RAD21 model with cell-type-specific Hi-C and RAD21 ChIP-seq inputs, GraphChIAr predicted few candidate RAD21-mediated interactions in GM12878, prominent interactions in H1-hESC, and reduced interactions in K562 (Fig. 6B). This example illustrates how GraphChIAr can prioritize candidate protein-mediated interactions at loci lacking matched ChIA-PET data. Overall, these results demonstrate that GraphChIAr effectively enhances chromatin interaction prediction from conventional Hi-C data to kilobase and sub-kilobase resolution, while preserving biologically meaningful loop enrichment signals. Genome-wide estimation of remote chromatin interactions To extend the utility of GraphChIAr from intra-window predictions to a genome-wide scale, we introduced an ‘offset’ mechanism. This mechanism enables the model to predict interactions between two non-overlapping genomic regions separated by a specified genomic distance (the offset). This is a critical step toward genome-wide scale analysis. We evaluated this functionality at both 10 kb and 200 bp resolutions. It is well-established that due to the physical properties of chromatin, most significant interactions occur within a limited distance, making long-range interactions inherently sparse. Our model accurately captures this biological reality. As shown in Fig. 7B (right panel), at 10 kb resolution, the model’s predicted inter-region contact map (showing interactions between the 21–23 and 23–25 Mb regions) exhibits a largely sparse background, which is highly consistent with experimental observations. Crucially, the model does not merely predict a flat background; instead, it successfully and precisely identifies true, CTCF-mediated long-range chromatin loops, which are visible as concentrated signals in the bottom-left corner of the map. GraphChIAr predicts chromatin interactions at multiple resolutions and captures long-range interaction properties. (A) Experimental CTCF ChIA-PET contact map for the chr10: 21–25 Mb region in GM12878 cells. The 1D track below shows the corresponding CTCF ChIP-seq signal. (B) Model predictions at 10 kb resolution. Left: Standard intra-window (on-diagonal) prediction for the 21–25 Mb region. Right: Inter-region (offset) prediction, showing interactions between the 21–23 and 23–25 Mb windows. The concentration of predicted signals in the bottom-left corner accurately reflects the distance-decay properties of chromatin interactions, confirming that most contacts occur within ~2 Mb. (C) Predicted 1 kb intra-window contact map. (D) Model predictions at 200 bp resolution. Left: Intra-window prediction for a zoomed-in region (chr10: 21.0–21.25 Mb). Right: Inter-region (offset) prediction, demonstrating the model’s capability to capture sparse, fine-scale long-range interactions at ultra-high resolution. All 1D tracks represent CTCF ChIP-seq binding signals. Furthermore, the offset mechanism is indispensable for ultra-high-resolution (200 bp) analysis. Due to computational constraints, the manageable window size at 200 bp resolution is limited and cannot directly span long-range interactions. By introducing the offset mechanism, we can systematically predict fine-scale interactions across the genome at a feasible computational cost. As illustrated in the 200 bp offset prediction (Fig. 7D, right panel), the model utilizes this mechanism to accurately predict fine-scale chromatin structures that span the window boundaries at a 200 bp resolution. In summary, the offset mechanism is a core function that successfully extends GraphChIAr’s capabilities to genome-wide prediction, enabling the systematic analysis of long-range interactions at both conventional and ultra-high resolutions. Conclusion We present GraphChIAr, a deep learning framework designed to estimate protein-mediated chromatin interactions by integrating Hi-C maps, ChIP-seq signals, and DNA sequence data. By utilizing graph neural networks, GraphChIAr effectively captures not only local chromatin interactions but also remote chromatin dependencies, enabling accurate modeling of chromatin conformation at a genome-wide scale. The model’s ability to integrate complementary features, such as DNA sequences and epigenomic profiles, significantly enhances its predictive performance. One of GraphChIAr’s key advantages is its multi-resolution capability, allowing for reliable interaction predictions down to an ultra-high 200 bp resolution. At this scale, the model effectively reduces noise in experimental data, revealing detailed chromatin structures that align with specific protein binding sites. GraphChIAr serves as a useful tool for studying chromatin interactions, especially in cases where experimental data are limited. This framework contributes to advancing our understanding of 3D genome organization and provides a means for exploring gene regulatory mechanisms in diverse biological contexts. Despite its strong performance, GraphChIAr has several limitations. The model relies on the availability of protein-mediated chromatin interaction data for training, which restricts its applicability to proteins and conditions with sufficient public datasets. In addition, the diversity of datasets explored in this study remains limited, and further validation across broader experimental platforms would be beneficial. A possible future direction is to extend GraphChIAr toward a unified multi-modal screening framework by integrating Hi-C with multiple ChIP-seq profiles to identify protein-specific 3D chromatin structures. In this way, the model could serve as both a predictor and a computational filter, and incorporating more diverse epigenomic features may further improve its robustness and applicability. Key Points GraphChIAr predicts protein-mediated chromatin interactions by integrating Hi-C, ChIP-seq, and genomic sequence data. GraphChIAr predicts protein-mediated chromatin interactions by integrating Hi-C, ChIP-seq, and genomic sequence data. The model shows strong performance across cell types, improving accuracy by integrating complementary genomic features. The model shows strong performance across cell types, improving accuracy by integrating complementary genomic features. The framework improves resolution to 200 bp, enabling near-nucleosome level chromatin mapping. The framework improves resolution to 200 bp, enabling near-nucleosome level chromatin mapping. Supplementary Material Contributor Information Hao Dong, School of Computer Science and Technology, College of Intelligence and Computing, Tianjin University, Tianjin 300350, China. Guo-Zheng Rao, School of Computer Science and Technology, College of Intelligence and Computing, Tianjin University, Tianjin 300350, China. Hao-Yu Wang, School of Computer Science and Technology, College of Intelligence and Computing, Tianjin University, Tianjin 300350, China. Xin-Ran Wu, School of Computer Science and Technology, College of Intelligence and Computing, Tianjin University, Tianjin 300350, China. Tong Xian, School of Computer Science and Technology, College of Intelligence and Computing, Tianjin University, Tianjin 300350, China. Bo-Qiang Wang, School of Computer Science and Technology, College of Intelligence and Computing, Tianjin University, Tianjin 300350, China. Pu-Feng Du, School of Computer Science and Technology, College of Intelligence and Computing, Tianjin University, Tianjin 300350, China. Author contributions HD collected the data, designed the model, implemented the algorithm, performed experiments, analyzed results, and partially wrote the manuscript. HYW participated in algorithm design and implementations. XRW, TX, and BQW participated in performing experiments, GZR and PFD supervised the whole study, conceptualized the algorithm, analyzed the results, and partially wrote the manuscript. Funding This work is partially supported by the National Natural Science Foundation of China [NSFC 62372320, and NSFC 61872268]. Data availability The code and data for reproducing the results of this paper are available in a GitHub repository (https://github.com/don194/GraphChIAr). References Associated Data Supplementary Materials Data Availability Statement The code and data for reproducing the results of this paper are available in a GitHub repository (https://github.com/don194/GraphChIAr).