UniChrom: a universal deep learning architecture for cross-scale chromatin interaction prediction

TL;DR

UniChrom is a deep learning model that accurately predicts chromatin interactions across different cell types and genomic scales, offering insights into genome organization and gene regulation.

Contribution

UniChrom introduces a novel attention-based deep learning framework for cross-scale and cross-lineage chromatin interaction prediction.

Findings

UniChrom outperforms existing methods in predicting chromatin interactions across multiple cell lines.

It achieves high accuracy for long-range interactions exceeding 1.77 megabases (AUC: 0.976).

The model generalizes well to endothelial cells, showing strong cross-lineage performance (AUC: 0.962).

Abstract

Chromatin interactions regulate gene expression and genome organization, but computational prediction across cell types remains challenging. We developed UniChrom, a deep learning framework integrating DNA sequences and epigenomic features through attention-based neural networks to predict chromatin interactions. Evaluation across human lymphoblastoid, leukemia, and fibroblast cell lines demonstrates superior performance compared to existing methods, with fivefold cross-validation and Wilcoxon tests confirming statistical significance (p < 0.05). Distance-stratified analysis reveals robust performance across all genomic scales, including long-range interactions exceeding 1.77 megabases (AUC: 0.976). Independent validation on endothelial cells confirms cross-lineage generalization (AUC: 0.962). Bootstrapping analysis with 1,000 iterations validates performance stability with tight 95% confidence intervals. DeepSHAP interpretability identifies CTCF and cohesin components as dominant features alongside cell-type-specific histone modifications, while DeepLIFT reveals functional regulatory motifs at interaction anchors. UniChrom provides a statistically validated framework for investigating genome architecture across cellular contexts with potential applications in understanding gene regulation in development and disease.

The online version contains supplementary material available at 10.1186/s12864-026-12625-x.

Figures

UniChrom architecture for chromatin interaction prediction. The framework integrates two parallel processing modules: a sequence module processing DNA sequences from interaction anchors through one-hot encoding, convolutional layers, and BiGRU networks; and an epigenomic module processing cell-type-specific ChIP-seq features through BiGRU and dense layers. Upper left: Hi-C data processing workflow showing chromatin loops identified by HICCUPS as positive labels and distance-matched non-interacting loci as negative labels. Upper right: chromosome-based data splitting strategy with chromosomes 7

Performance comparison of UniChrom-sequence with comparative methods. A AUC values for UniChrom-sequence and comparative methods (TargetFinder, SPEID, EPIshilbert, FusNet) across GM12878, IMR90, and K562 cell lines. B AUPR values across the same methods and cell lines

Feature integration improves cross-cell-type prediction stability. A AUC comparison of UniChrom (integrated model), UniChrom-sequence (sequence-only), and UniChrom-Histones (genomic features-only) across GM12878, IMR90, and K562 cell lines. B AUPR comparison showing performance advantage of feature integration, particularly evident in K562 where individual feature modules show larger performance decline

Multi-modal method comparison across cell types. A-C ROC curves comparing UniChrom against ChINN, CLNN_loop, Enhancer_MDLF, and EPI_Trans in GM12878, IMR90, and K562 cell lines, with corresponding AUC values. D-F Precision-recall curves for the same methods and cell lines, with AUPR values demonstrating performance differences across cellular contexts

Cross-cell-type transfer learning performance. Boxplot comparison of UniChrom (dark blue) and ChINN (cyan) across five metrics (ACC, AUC, AUPR, F1, Recall) for cross-cell-type prediction (trained on GM12878, tested on IMR90 and K562). Wilcoxon signed-rank test *p*-values indicate statistically significant performance differences between methods

B). Performance evaluation revealed robust prediction across all distance categories (Supplementary Fig. 6A). Notably, UniChrom achieved the highest AUC (0.976) and accuracy (0.932) for large-distance interactions, even surpassing shorter-range performance. While AUPR, F1, and recall showed modest reductions for large-distance interactions (AUPR = 0.894, decreased by 0.086 relative to small interactions; F1 = 0.840, decreased by 0.043 and 0.090 relative to medium and small interactions; Recall = 0.850, decreased by 0.037 and 0.094), all metrics remained substantially high. These results demons

Cell-type-specific feature importance patterns. Horizontal bar charts showing mean absolute SHAP values for genomic features in K562, GM12878, and IMR90 cell lines, ranked by contribution to interaction predictions. CTCF exhibits highest importance in K562 (SHAP = 0.142), while RAD21 shows highest importance in GM12878 (SHAP = 0.156) and IMR90 (SHAP = 0.109). Genomic distance and histone modifications display cell-type-specific importance patterns

Multi-scale sequence interpretability analysis of regulatory elements. Top: RUNX3 gene structure showing protein domains and COSMIC mutations. Middle: Three complementary interpretability analyses of an interaction anchor region (chr6:41556711–41556866): DeepLIFT nucleotide contribution scores, tiling deletion effects, and systematic mutagenesis sensitivity map. Orange box highlights a functionally important region (nucleotides 118–126) showing convergent evidence across all three analyses and containing a RUNX3 binding motif (sequence logo). Color scale indicates contribution magnitude (red:

Full text

Introduction

The intricate three-dimensional organization of chromatin within eukaryotic nuclei constitutes a fundamental regulatory mechanism that governs gene expression. This hierarchical spatial architecture is characterized by complex loop structures that facilitate precise spatiotemporal interactions between distant genomic elements [1]. At the molecular level, the CCCTC-binding factor (CTCF) serves as a master regulator of this organization, orchestrating chromatin loop formation through homodimer interactions and establishing critical topological domains [2]. A particularly significant aspect of this spatial organization is manifested in enhancer-promoter interactions (EPIs), wherein distal regulatory elements achieve precise spatial proximity with their target promoters to modulate transcriptional dynamics [3, 4]. The importance of these chromatin interactions extends beyond basic cellular processes, as perturbations in three-dimensional genome architecture have been implicated in diverse pathological conditions, suggesting potential therapeutic strategies through targeted manipulation of spatial genome organization [5].

Understanding the physical interactions between distal regulatory elements within chromatin architecture requires experimental approaches capable of capturing spatial proximity. Researchers have developed various chromosome conformation capture techniques, including 3 C and its derivatives (4C, 5C), as well as chromatin interaction analysis by paired-end tag sequencing (ChIA-PET) [6–8]. Recent advances in genome-wide technologies such as Hi-C, Capture Hi-C, and HiChIP have enabled generation of high-resolution maps of chromatin interactions [9]. However, these experimental approaches require substantial sequencing depth and can be resource-intensive when applied across multiple cell types or conditions. Computational prediction methods offer complementary approaches to experimental techniques, enabling exploration of chromatin interactions in contexts where experimental data may be limited or unavailable [10].

The complexity of chromatin interactions has catalyzed the development of diverse computational approaches, with machine learning architectures demonstrating particular promise in this field. SPEID pioneered the integration of Convolutional Neural Networks (CNN) for sequence feature extraction with Long Short-Term Memory (LSTM) networks, establishing a framework for capturing complex interaction patterns [11]. Building on this foundation, EPIshilbert introduced an innovative approach by transforming linear genomic sequences into two-dimensional representations through Hilbert curves [12], enabling sophisticated spatial feature analysis via residual networks [13]. EPI-Trans further advanced the field by harnessing transformer architectures to model intricate chromatin interaction patterns [14], while EnContact demonstrated superior performance in predicting enhancer-enhancer interactions across multiple cell lines [15]. Despite these advances, systematic evaluation has revealed fundamental challenges in model assessment, particularly regarding class imbalance and the critical need for validation strategies that rigorously account for sequence similarity between training and test sets [16].

The integration of epigenomic features into machine learning frameworks has emerged as a critical advancement in chromatin interaction prediction. TargetFinder established an important precedent by leveraging gradient boosting algorithms to predict EPIs through epigenomic signatures [17]. This approach was further refined by CTCF-MP, which integrated CTCF binding motifs with chromatin accessibility data in gradient boosting frameworks [18], while ChINN introduced a sophisticated deep learning architecture capable of simultaneously processing genomic sequences and spatial information [19]. FusNet extended these capabilities through ensemble learning strategies, enabling the prediction of protein-mediated chromatin interactions via systematic model integration [20]. Despite these advances, existing computational methods face significant context-dependent limitations, with prediction accuracy varying markedly across interaction types and cellular environments. These methodological constraints highlight the critical need for more versatile prediction approaches capable of effectively modeling diverse chromatin interactions across multiple cellular contexts.

To address the challenge of predicting chromatin interactions across diverse cellular contexts, we developed UniChrom, a deep learning framework that integrates DNA sequence information with cell-type-specific epigenomic features. UniChrom processes three categories of input: DNA sequences from interaction anchor regions, epigenomic features including histone modifications and architectural protein binding, and genomic distance between anchors. The framework employs convolutional neural networks to extract sequence features, bidirectional recurrent networks to capture long-range dependencies, and attention mechanisms to weight feature contributions. Through systematic evaluation across multiple cell lines and interaction types, we demonstrate that integration of sequence and epigenomic features provides more stable cross-cell-type prediction compared to individual feature modalities. Interpretability analysis using DeepSHAP and DeepLIFT reveals that the model learns biologically relevant features, including established transcription factor binding sites and cell-type-specific epigenetic signatures, providing insights into sequence and epigenetic determinants of chromatin organization.

Material and methods

Dataset preparation and processing

Our model development and evaluation utilized high-resolution chromatin interaction data from four human cell lines: GM12878 (lymphoblastoid), K562 (chronic myelogenous leukemia), IMR90 (fetal lung fibroblasts), and HUVEC (human umbilical vein endothelial cells). GM12878, K562, and IMR90 were used for model training and primary evaluation, while HUVEC was reserved exclusively for independent cross-cell-line validation. We obtained in situ Hi-C datasets from the Gene Expression Omnibus database (accession: GSE63525) [21]. Raw interaction data underwent comprehensive processing through the HiC-Pro pipeline [22], implementing systematic corrections for GC content, sequence mappability, and restriction fragment length biases. We identified significant chromatin loops using the HICCUPS algorithm [21] at 5-kilobase (kb) resolution, applying stringent filtering criteria (false discovery rate < 0.1) to ensure high-confidence interaction calls. The 5 kb resolution used by HICCUPS directly determined the sequence length L = 5,000 bp for model inputs at each interaction anchor. To accommodate variable-length genomic regions during model application, we implemented standardized preprocessing: for sequences longer than 5,000 bp, we extracted a central fragment of 5,000 bp; for sequences shorter than 5,000 bp, we performed symmetric zero-padding extending equally upstream and downstream from the sequence center to achieve the required 5,000 bp length. This standardization ensures consistent input dimensions across all samples while preserving the biological information at interaction anchor regions.

Dataset construction followed a rigorous methodology to ensure robust model training and evaluation. We designated experimentally validated chromatin loops identified by HICCUPS as positive samples. For negative sample generation, we implemented a distance-matched sampling strategy whereby non-interacting genomic regions were first stratified into bins based on their linear separation (5 kb to 2 megabases), then systematically sampled to match the genomic distance distribution of positive interactions. To enable unbiased model evaluation and prevent data leakage, we employed chromosome-based partitioning rather than random sample splitting. This decision addresses documented overfitting issues in chromatin interaction prediction: post-publication analyses have demonstrated that random splitting of window-based datasets can substantially inflate performance metrics due to substantial overlap between positive samples in training and test sets [23–25]. Such overlap enables models to achieve inflated cross-validation performance through memorization of overlapping patterns rather than genuine biological feature learning. Our chromosome-based strategy—chromosomes 7–8 for validation, chromosomes 9–10 for testing, remaining chromosomes for training—ensures complete genomic separation between datasets, eliminating overlapping regions and providing realistic assessment of generalization to truly unseen genomic contexts (Supplementary Table 2). All comparative methods were evaluated under this stringent framework. The GM12878 dataset comprised 9,448 positive and 9,448 negative samples, which were partitioned into training (7,657), validation (1,046), and test (745) sets based on chromosomal assignment. To characterize the composition of interaction types in our datasets, we performed overlap analysis between HICCUPS-identified loops and ChIA-PET CTCF data. Using GM12878 as a representative example, we compared all positive training samples against ChIA-PET CTCF-mediated interactions with an overlap threshold of 0.8. The analysis revealed that 2,651 of 7,657 samples (34.6%) overlapped with CTCF data, indicating the presence of diverse interaction types including CTCF-mediated loops, enhancer-promoter interactions, and enhancer-enhancer interactions in the training datasets.

Neural network architecture and implementation

UniChrom implements a hierarchical deep learning framework engineered for comprehensive feature integration (Fig. 1, Supplementary Table 1). The architecture comprises three specialized functional modules: a sequence processing module, a genomic feature integration module, and an attention-based fusion module, enabling systematic integration of both sequence-level and genomic features for chromatin interaction prediction. The sequence processing module leverages CNN and RNN architectures to extract complex patterns from DNA sequences at anchor regions, while the genomic feature module processes cell-line-specific signals including histone modifications, transcription factor binding profiles, and spatial distances between interacting regions. A sophisticated fusion attention module then integrates these complementary feature representations, enabling dynamic weighting of diverse information sources for precise interaction prediction.Fig. 1. UniChrom architecture for chromatin interaction prediction. The framework integrates two parallel processing modules: a sequence module processing DNA sequences from interaction anchors through one-hot encoding, convolutional layers, and BiGRU networks; and an epigenomic module processing cell-type-specific ChIP-seq features through BiGRU and dense layers. Upper left: Hi-C data processing workflow showing chromatin loops identified by HICCUPS as positive labels and distance-matched non-interacting loci as negative labels. Upper right: chromosome-based data splitting strategy with chromosomes 7–8 for validation, chromosomes 9–10 for testing, and remaining chromosomes for training. Features from both modules are concatenated and processed through BiLSTM and attention layers for final interaction prediction

Input processing and feature extraction

The sequence processing module implements sophisticated feature extraction from DNA sequences at interaction anchor regions. Input sequences undergo one-hot encoding, generating a binary matrix:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R\in {\mathrm{0,1}}^{L\times 4}$$\end{document}

where L represents sequence length, with each nucleotide position encoded as a four-dimensional vector corresponding to the four possible bases (A, C, G, T). The encoded sequences are processed through dual-stream neural networks: convolutional layers implement sliding window operations optimized for detecting sequence motifs and regulatory elements, while bidirectional gated recurrent units (BiGRU) capture long-range sequence dependencies in both forward and reverse directions. This bidirectional approach enables comprehensive feature extraction that accounts for both local sequence patterns and their broader genomic context.

The genomic module processes cell-type-specific epigenomic features, including ten distinct histone modifications (H3K4me1/2/3, H3K9ac/me3, H3K27ac/me3, H3K36me3, H3K79me2, H4K20me1), three architectural proteins (CTCF, RAD21, SMC3), and spatial distance information between potential interaction regions. These features are carefully normalized and transformed to preserve their biological significance while facilitating effective integration with sequence-derived features. The fusion attention module then dynamically weights the contribution of different feature types, allowing the model to adapt its predictions based on the relative importance of sequence and epigenomic signals in different genomic contexts.

Attention mechanism implementation

The attention mechanism implements a multi-stage feature integration process designed to capture complex interactions between sequence and genomic features. This mechanism comprises four key computational stages:

1. Feature Projection and Query-Key Generation: The integrated features undergo transformation into query ( \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q$$\end{document} ) and key ( \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K$$\end{document} ) representations through linear projections:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Query-Key Pairs=\left[Q;K\right]$$\end{document}

where [ \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q$$\end{document} ; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K$$\end{document} ] denotes the concatenated query-key representation that captures feature relationships across different genomic contexts.

• 2.Similarity Computation: A feedforward neural network calculates similarity scores according to:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Similarity={W}_{2}\times tanh({W}_{1}\times [Q;K]+b)\times v$$\end{document}

where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W}_{1}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W}_{2}$$\end{document} represent trainable weight matrices that learn feature relationships, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b$$\end{document} denotes the bias term enabling flexible feature mapping, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v$$\end{document} represents a learnable vector that projects the transformed features to similarity scores.

• 3.Attention Weight Normalization: The computed similarity scores undergo normalization through the softmax function to generate interpretable attention weights:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Attention\_weights}_{i}=\frac{exp({Similarity}_{i})}{\sum_{j}exp({Similarity}_{j})}$$\end{document}

where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Similarity}_{i}$$\end{document} represents the similarity score at position i, ensuring the attention weights sum to unity across all positions.

• 4.Feature Integration: The final output is computed through weighted aggregation of value vectors:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Output=\sum_{i}({Attention\_weights}_{i}\times {Values}_{i})$$\end{document}

where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Attention\_weights}_{i}$$\end{document} represents the normalized attention weight at position i, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Values}_{i}$$\end{document} represents the corresponding value vector, enabling the model to dynamically focus on relevant feature combinations for interaction prediction.

Prediction layer

The model implements a hierarchical prediction structure culminating in a dense layer with a single neuron and sigmoid activation function for final classification:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P\left(interaction\right)=\sigma \left(W\times F+b\right)$$\end{document}

where W represents the learnable weight matrix, F denotes the integrated feature vector from the attention module, b represents the bias term, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma$$\end{document} is the sigmoid activation function. The prediction mechanism generates probability scores bounded between 0 and 1, with samples classified as positive interactions when the prediction value exceeds the empirically determined threshold of 0.5, and negative otherwise. This binary classification framework enables quantitative assessment of chromatin interaction likelihood based on the comprehensive feature integration performed by the upstream network components. The probability scores provide a continuous measure of interaction strength, facilitating both binary classification of chromatin interactions and analysis of interaction likelihood in different genomic contexts.

Model evaluation metrics

To systematically assess the model’s prediction performance, we implemented a comprehensive evaluation framework incorporating multiple complementary metrics: Area Under the Receiver Operating Characteristic Curve (AUC), Accuracy (ACC), F1-score, Recall, and Precision. These metrics were calculated using standardized evaluation protocols, with all competing models subjected to identical hyperparameter configurations and data preprocessing procedures to ensure fair comparison.

The fundamental binary classification metrics are defined as:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Accuracy=\frac{TP+TN}{TP+TN+FP+FN}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Recall=\frac{TP}{TP+FN}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Precision=\frac{TP}{TP+FP}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F1=2\cdot \frac{Precision\times Recall}{Precision+Recall}$$\end{document}

where TP (True Positives) represents correctly predicted chromatin interactions, TN (True Negatives) indicates correctly identified non-interacting regions, FP (False Positives) denotes incorrectly predicted interactions, and FN (False Negatives) represents missed interaction predictions. These metrics collectively provide a multi-faceted assessment of model performance across different aspects of prediction accuracy.

The AUC serves as a threshold-independent performance measure. The ROC curve is constructed by plotting the True Positive Rate (TPR) against the False Positive Rate (FPR) across varying classification thresholds:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$TPR=\frac{TP}{TP+FN}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$FPR=\frac{FP}{FP+TN}$$\end{document}

To address potential class imbalance inherent in chromatin interaction datasets, we additionally employed the Area Under the Precision-Recall curve (AUPR) metric, which provides particularly insightful evaluation for datasets with uneven class distributions:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$AUPR=\int Precision\left(Recall\right)dRecall$$\end{document}

Training strategy

UniChrom was trained on NVIDIA A800 GPUs using the Adam optimizer for adaptive learning rate adjustment, with an initial learning rate of 0.001. The training ran for 100 epochs with a batch size of 128 to balance efficiency and memory usage. To prevent overfitting, we applied early stopping (patience = 5), which automatically halts training if validation performance shows no significant improvement for 5 consecutive epochs, thereby ensuring stable generalization performance. The model uses binary cross-entropy as the loss function for chromatin interaction prediction.

Comparative method implementation and evaluation framework

To ensure fair comparison, all comparative methods (SPEID, EPIshilbert, FusNet, TargetFinder, ChINN, CLNN_loop, EPI_Trans, Enhancer_MDLF) were evaluated under identical experimental conditions. Methods were categorized by input modality: some primarily process DNA sequences (e.g., SPEID, EPIshilbert, FusNet), while others integrate sequence with epigenomic features (e.g., ChINN, EPI_Trans, CLNN_loop, Enhancer_MDLF). All methods were trained using NVIDIA A800 GPUs with Adam optimizer (learning rate: 0.001), batch size of 128, 100 epochs, and early stopping (patience = 5). Data partitioning was consistent across methods: chromosomes 7–8 for validation, chromosomes 9–10 for testing, and remaining chromosomes for training. Each method received inputs according to its design requirements, with epigenomic features (histone modifications, architectural proteins, genomic distance) provided where applicable. Model architectures followed original publications.

Feature importance and model interpretability

To systematically assess the contribution of each genomic feature to chromatin interaction prediction, we employed the DeepSHAP (Deep SHAP) method [26], an implementation of SHapley Additive exPlanations (SHAP) optimized for deep neural networks [27]. DeepSHAP is based on the Shapley values from cooperative game theory, assigning each input feature a value that represents its marginal contribution to the model output. In our study, the DeepSHAP analysis pipeline for the genomic-feature branch of the UniChrom model includes the following steps:

• First, we train UniChrom’s genomic-feature module until convergence, outputting chromatin interaction prediction probabilities. Then, we randomly sample 2,000 instances from the training set as a background reference set to establish a baseline for feature contributions and ensure robustness of the computed results.

Next, we initialize the explainer using shap.DeepExplainer from the Python SHAP library, providing the trained model and the background dataset, and compute SHAP values for every sample in the full test set. DeepSHAP traces the change of each feature from its background value to its actual input value and quantifies the impact of this change on the model’s prediction, thereby decomposing the model output into the sum of contributions from all features.

To quantify feature importance, we take the absolute value of the SHAP values for each sample, aggregate the absolute SHAP values for the same type of genomic feature (e.g., H3K4me3, CTCF, etc.) across all positions, and then average across test samples to obtain a global importance score for each feature category:

\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\overline{\phi }}_{i}=\frac{1}{N}\sum_{j=1}^{N}|{\phi }_{ij}|$$\end{document}

Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\phi }$$\end{document} denotes the mean absolute contribution of feature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i$$\end{document} , \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N$$\end{document} is the number of test-set samples, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\phi }_{ij}$$\end{document} is the SHAP value of feature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i$$\end{document} for sample \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$j$$\end{document} . Finally, the feature importance results are aggregated and visualized with a bar chart to analyze the relative contributions of different genomic features and their differential patterns across cell types.

Sequence interpretability analysis using DeepLIFT

To dissect the sequence-specific determinants of chromatin interactions at nucleotide resolution, we implemented Deep Learning Important FeaTures (DeepLIFT) [28], a state-of-the-art deep learning interpretation framework that quantifies the contribution of individual sequence elements to interaction predictions. The DeepLIFT methodology systematically evaluates feature importance by comparing prediction outcomes between input sequences and carefully constructed reference sequences, enabling precise identification of functionally relevant DNA motifs and regulatory elements. For each analyzed sequence, reference sequences are generated through dinucleotide shuffling, a sophisticated randomization procedure that maintains local sequence composition characteristics such as GC content and dinucleotide frequencies while disrupting potential regulatory motifs. This controlled shuffling approach ensures that observed differences in predictions can be attributed specifically to the disruption of functional sequence patterns rather than changes in basic sequence properties. The contribution scores computed through comparison between original and reference sequence predictions provide quantitative measures of each nucleotide position’s importance in determining chromatin interaction propensity, facilitating the identification of key regulatory sequences that mediate specific chromatin interactions.

In silico sequence analysis

To systematically characterize sequence-dependent regulatory mechanisms of chromatin interactions, we implemented a comprehensive computational framework that simulates targeted genetic perturbations. The analysis pipeline incorporates two complementary approaches: tiling-deletion scanning and systematic nucleotide mutation analysis [29]. The tiling-deletion protocol implements a high-resolution scanning strategy that systematically evaluates the functional importance of local sequence contexts. Using a sliding window of 10 base pairs with single-nucleotide resolution, we computationally simulate precise deletions by replacing nucleotides within each window with 'N' characters. Each modified sequence undergoes prediction through our trained model, generating a detailed map of sequence regions critical for maintaining chromatin interactions. This approach effectively identifies functionally essential sequence elements by quantifying the impact of their deletion on interaction probability. Complementing the deletion analysis, we performed comprehensive in silico mutagenesis to evaluate the sequence specificity of identified regulatory regions. This analysis systematically generates all possible single-nucleotide variants at each position in the input sequence, computing interaction strength predictions for each sequence variant. The resulting mutation effect profiles are visualized through position-specific scoring matrices, where each position is evaluated against all possible nucleotide substitutions. This systematic perturbation analysis reveals the sequence specificity of regulatory elements and identifies nucleotide positions crucial for chromatin interaction stability.

Bootstrapping analysis

To more robustly evaluate model performance and quantify the uncertainty of each metric, we incorporated the bootstrap (bootstrapping) method into our results analysis [30]. Specifically, we performed repeated sampling with replacement on the test set of each cell line to generate multiple bootstrap resampled datasets. For each resample, we recalculated the model’s performance metrics, including AUC, AUPR, ACC, F1, and Recall. We conducted N = 1,000 resampling iterations to adequately approximate the empirical distributions of these metrics. After aggregating all resampling results, we used the percentile method to compute the mean and the 95% confidence interval for each metric. These confidence intervals reflect the statistical variability of model performance without relying on parametric assumptions.

Results

Sequence module performance exhibits cell-type specificity

To evaluate the contribution of DNA sequence features to chromatin interaction prediction, we systematically compared UniChrom's sequence module against established methods including SPEID, EPIshilbert, FusNet, and TargetFinder under chromosome-based evaluation framework. In the GM12878 lymphoblastoid cell line, UniChrom-sequence demonstrated robust predictive performance with AUC of 0.890 and AUPR of 0.911, substantially outperforming comparative methods: EPIshilbert (AUC: 0.717, AUPR: 0.675), SPEID (AUC: 0.720, AUPR: 0.693), FusNet (AUC: 0.720, AUPR: 0.704), and TargetFinder (AUC: 0.739, AUPR: 0.717) (Fig. 2).Fig. 2. Performance comparison of UniChrom-sequence with comparative methods. A AUC values for UniChrom-sequence and comparative methods (TargetFinder, SPEID, EPIshilbert, FusNet) across GM12878, IMR90, and K562 cell lines. B AUPR values across the same methods and cell lines

To verify the statistical significance of these performance differences, we conducted fivefold cross-validation across all comparative methods and performed pairwise Wilcoxon signed-rank tests. The analysis confirmed statistically significant superior performance of UniChrom-sequence across all metrics (AUC, AUPR, ACC, F1, Recall) with p-values < 0.05 when compared against each baseline method (Supplementary Fig. 1). Additionally, bootstrapping analysis on resampled GM12878 data demonstrated that UniChrom-sequence achieves more stable performance relative to comparative methods (Supplementary Fig. 4A).

Notably, prediction performance using sequence information exhibited pronounced cell-type specificity across all evaluated methods. UniChrom-sequence performance showed gradual decline from GM12878 (AUC: 0.890, AUPR: 0.911) to IMR90 (AUC: 0.824, AUPR: 0.834) and K562 cells (AUC: 0.740, AUPR: 0.648), a pattern consistently observed across comparative methods. This cell-type-dependent variation suggests that while DNA sequence encodes fundamental interaction determinants, sequence features alone cannot fully capture the complex, context-specific regulatory mechanisms governing chromatin organization. These findings underscore the necessity of integrating additional cell-type-specific features for robust prediction across diverse cellular contexts.

Feature integration establishes robust cross-cell-type prediction

To assess the contribution of feature integration to prediction performance, we compared three model configurations: UniChrom (integrating sequence and genomic feature modules), UniChrom-sequence (sequence-only), and UniChrom-Histones (genomic features-only). This ablation analysis evaluates whether multi-modal feature integration provides advantages beyond individual feature sets.

Performance evaluation across three cell lines revealed distinct patterns for different feature configurations (Fig. 3). UniChrom consistently achieved the highest performance across all cell types, with AUC values of 0.956 (GM12878), 0.955 (IMR90), and 0.941 (K562), and corresponding AUPR values of 0.960, 0.952, and 0.911. In contrast, individual feature modules exhibited larger performance variations across cell types. UniChrom-Histones showed strong performance in GM12878 (AUC: 0.943, AUPR: 0.954) and IMR90 (AUC: 0.924, AUPR: 0.927), but substantial decline in K562 (AUC: 0.794, AUPR: 0.808). UniChrom-sequence displayed progressive performance decrease from GM12878 (AUC: 0.890, AUPR: 0.911) through IMR90 (AUC: 0.824, AUPR: 0.834) to K562 (AUC: 0.740, AUPR: 0.648). Comprehensive evaluation across five metrics (Precision, Recall, F1, ACC, AUC) confirmed this pattern, with UniChrom maintaining more consistent performance across cellular contexts (Supplementary Fig. 2).Fig. 3. Feature integration improves cross-cell-type prediction stability. A AUC comparison of UniChrom (integrated model), UniChrom-sequence (sequence-only), and UniChrom-Histones (genomic features-only) across GM12878, IMR90, and K562 cell lines. B AUPR comparison showing performance advantage of feature integration, particularly evident in K562 where individual feature modules show larger performance decline

The observed stability of integrated prediction reflects complementary information captured by different feature modalities. Genomic features represent cell-type-specific regulatory states that vary across cellular contexts, while sequence features encode genomic architecture that remains constant. Integration of both modalities enables the model to leverage cell-invariant and cell-specific determinants simultaneously, improving robustness across diverse cellular environments. The integrative strategy proves particularly valuable in challenging prediction contexts, where individual feature types show performance limitations. While the integrated model exhibits some variation across cell types, this variation is substantially smaller than that observed for single-modality approaches. These findings demonstrate that multimodal integration provides more stable cross-cell-type prediction, though challenges remain in achieving uniform performance across all cellular contexts.

Systematic comparison reveals advantages of integrative modeling

To comprehensively evaluate UniChrom's multi-modal learning capability, we compared it against five published methods that integrate chromatin interaction prediction with various genomic features: ChINN (employing convolutional networks for CTCF and RNA Pol II-mediated interactions), EPI_Trans (transformer-based architecture for enhancer-promoter interactions), Enhancer_MDLF (multi-task deep learning for enhancer recognition), and CLNN_loop (convolutional LSTM for chromatin loops). All methods were evaluated under identical chromosome-based data partitioning to ensure fair comparison.

Performance analysis across three cell lines revealed consistent patterns. In GM12878, UniChrom achieved AUC of 0.956 and AUPR of 0.960, compared to ChINN (AUC: 0.86, AUPR: 0.899), CLNN_loop (AUC: 0.945, AUPR: 0.954), EPI_Trans (AUC: 0.757, AUPR: 0.739), and Enhancer_MDLF (AUC: 0.742, AUPR: 0.712). In IMR90, UniChrom maintained AUC of 0.955 and AUPR of 0.952, while comparative methods showed varying performance levels (ChINN: AUC 0.944, AUPR 0.934; CLNN_loop: AUC 0.948, AUPR 0.946; EPI_Trans: AUC 0.732, AUPR 0.698; Enhancer_MDLF: AUC 0.713, AUPR 0.662). In K562, UniChrom achieved AUC of 0.941 and AUPR of 0.911, with similar relative performance patterns observed across baseline methods (Fig. 4). The consistent cross-cell-type performance suggests that UniChrom's hierarchical feature integration effectively captures cell-type-invariant chromatin organization principles while accommodating cell-specific regulatory patterns.Fig. 4. Multi-modal method comparison across cell types. A-C ROC curves comparing UniChrom against ChINN, CLNN_loop, Enhancer_MDLF, and EPI_Trans in GM12878, IMR90, and K562 cell lines, with corresponding AUC values. D-F Precision-recall curves for the same methods and cell lines, with AUPR values demonstrating performance differences across cellular contexts

To verify statistical significance of observed performance differences, we conducted fivefold cross-validation across all three cell lines and performed pairwise Wilcoxon signed-rank tests comparing UniChrom against each baseline method. The analysis confirmed statistically significant superior performance of UniChrom across all metrics (AUC, AUPR, ACC, F1) with p-values < 0.05 for all pairwise comparisons (Supplementary Fig. 3). Bootstrapping analysis with 1,000 iterations on GM12878 data further validated these findings, demonstrating robust performance with tight 95% confidence intervals across all metrics (Supplementary Fig. 4B). These statistical validations confirm that UniChrom's performance advantages are reliable and not attributable to random variation.

Transfer learning enables cross-cell-type chromatin interaction prediction

To assess cross-cell-type generalization capability, we evaluated UniChrom's transfer learning performance through a training-transfer paradigm: models were trained on GM12878 data and subsequently applied to predict chromatin interactions in IMR90 and K562 cell lines without additional fine-tuning. We compared UniChrom against ChINN, which also supports transfer learning through feature extraction from pre-trained models. This evaluation addresses a practical challenge in chromatin interaction prediction, as generating comprehensive training data for each cell type is resource-intensive.

Transfer learning results demonstrated substantial performance differences between methods. When trained on GM12878 and transferred to IMR90, UniChrom achieved AUC of 0.933 and AUPR of 0.92, compared to ChINN's AUC of 0.821 and AUPR of 0.86. Transfer to K562 yielded similar patterns, with UniChrom achieving AUC of 0.914 and AUPR of 0.88, while ChINN achieved AUC of 0.626 and AUPR of 0.61 (Supplementary Table 3). Across both transfer scenarios, UniChrom maintained performance levels substantially closer to cell-type-specific training performance, suggesting effective learning of generalizable chromatin organization features.

Statistical validation through Wilcoxon signed-rank tests confirmed that UniChrom's transfer learning advantages are statistically significant across all evaluation metrics (ACC: p = 7.81e-03, AUC: p = 1.17e-02, AUPR: p = 7.42e-03, F1: p = 7.81e-03, Recall: p = 7.81e-03) (Fig. 5). These results indicate that UniChrom's multi-scale feature integration and attention mechanisms enable robust transfer of chromatin organization principles across diverse cellular contexts.Fig. 5. Cross-cell-type transfer learning performance. Boxplot comparison of UniChrom (dark blue) and ChINN (cyan) across five metrics (ACC, AUC, AUPR, F1, Recall) for cross-cell-type prediction (trained on GM12878, tested on IMR90 and K562). Wilcoxon signed-rank test p-values indicate statistically significant performance differences between methods

Performance comparison of UniChrom with methods targeting specific interaction types

To evaluate UniChrom's multi-task learning capability, we compared it against three methods developed for specific interaction types using established benchmark datasets: ChINN for CTCF-mediated interactions, EnContact for enhancer-enhancer interactions (EEI), and SPEID for enhancer-promoter interactions (EPI). The evaluation datasets comprised 55,280 CTCF-mediated, 28,762 enhancer-enhancer, and 3,988 enhancer-promoter interaction samples, with strictly balanced positive-to-negative ratios (1:1) to avoid performance inflation from class imbalance [15, 17, 19]. Samples were partitioned into training, validation, and test sets at an 8:1:1 ratio.

Performance evaluation revealed differential prediction difficulty across interaction types (Fig. 6). For CTCF-mediated interactions, UniChrom achieved AUC of 0.964, outperforming ChINN (AUC: 0.897), a method specifically designed for CTCF interaction prediction. For enhancer-promoter interactions, UniChrom achieved AUC of 0.737, compared to SPEID (AUC: 0.577) and EnContact (AUC: 0.462). Notably, ChINN showed limited performance on EP interactions (AUC: 0.474), despite strong performance on its target CTCF interactions. This discrepancy reflects ChINN's design optimization for CTCF-mediated interactions and the removal of dataset artifacts present in earlier EP benchmark datasets. Previous EP datasets often contained severe class imbalance (1:20 positive-to-negative ratios) and sample redundancy that artificially inflated performance metrics. Our strictly balanced dataset construction provides more realistic evaluation conditions. Enhancer-enhancer interaction prediction proved challenging for all methods, with UniChrom achieving AUC of 0.588 compared to EnContact's AUC of 0.577.Fig. 6. Performance comparison across chromatin interaction types. ROC curves showing prediction performance for CTCF-mediated interactions (left), enhancer-enhancer interactions (middle), and enhancer-promoter interactions (right). UniChrom is compared against methods designed for specific interaction types: ChINN (CTCF-mediated), EnContact (enhancer-enhancer), and SPEID (enhancer-promoter). UniChrom achieves AUC of 0.964 for CTCF-mediated interactions, 0.737 for enhancer-promoter interactions, and 0.588 for enhancer-enhancer interactions

These results demonstrate that while specialized methods achieve strong performance on their target interaction types, they often show limited generalization to other interaction mechanisms. UniChrom's multi-task architecture enables consistent performance across diverse interaction types, suggesting effective learning of shared chromatin organization principles.

Feature importance analysis reveals regulatory hierarchy in chromatin organization

To understand the regulatory mechanisms underlying chromatin interaction prediction, we applied DeepSHAP analysis to UniChrom's genomic feature module across GM12878, IMR90, and K562 cell lines. The analysis revealed both conserved and cell-type-specific patterns in feature importance (Fig. 7). In GM12878 and IMR90 cells, the cohesin component RAD21 exhibited the highest importance scores (SHAP values: 0.15561 and 0.10891, respectively), while K562 cells showed highest importance for CTCF (SHAP: 0.14179), indicating differential reliance on architectural proteins across cellular contexts.Fig. 7. Cell-type-specific feature importance patterns. Horizontal bar charts showing mean absolute SHAP values for genomic features in K562, GM12878, and IMR90 cell lines, ranked by contribution to interaction predictions. CTCF exhibits highest importance in K562 (SHAP = 0.142), while RAD21 shows highest importance in GM12878 (SHAP = 0.156) and IMR90 (SHAP = 0.109). Genomic distance and histone modifications display cell-type-specific importance patterns

The high importance of CTCF and cohesin complex components aligns with established biological understanding of chromatin organization. CTCF and cohesin mediate chromatin loop formation by promoting long-range interactions between distal regulatory elements, with most chromatin loops containing convergent CTCF motifs that strongly influence interaction patterns [31, 32]. The substantial predictive contribution of cohesin components (SMC3 and RAD21) in our model is consistent with previous reports of their high predictive power [17], which can be attributed to their enrichment at open chromatin regions [33] where their binding variability plays important roles in determining chromatin interaction patterns.

Beyond architectural proteins, the analysis revealed additional regulatory features with consistent or cell-type-specific contributions. Genomic distance showed stable importance across all cell types (SHAP values: 0.02374 to 0.03647), reflecting fundamental physical constraints on chromatin folding. Histone modifications displayed cell-type-specific importance patterns: H3K4me1 showed higher importance in K562 cells (SHAP: 0.03288) compared to other cell types. These cell-type-specific epigenetic signatures suggest distinct regulatory mechanisms for modulating chromatin accessibility and interaction specificity across different cellular contexts.

Sequence interpretability analysis identifies regulatory motifs

To dissect sequence-level determinants of chromatin interactions, we implemented complementary interpretability analyses through DeepLIFT, in silico tiling deletion, and systematic mutagenesis approaches. Analysis of a representative anchor sequence from GM12878 (chr15:63,386,917–63,387,072) within the RUNX3 gene region revealed convergent evidence for functional regulatory elements (Fig. 8).Fig. 8. Multi-scale sequence interpretability analysis of regulatory elements. Top: RUNX3 gene structure showing protein domains and COSMIC mutations. Middle: Three complementary interpretability analyses of an interaction anchor region (chr6:41556711–41556866): DeepLIFT nucleotide contribution scores, tiling deletion effects, and systematic mutagenesis sensitivity map. Orange box highlights a functionally important region (nucleotides 118–126) showing convergent evidence across all three analyses and containing a RUNX3 binding motif (sequence logo). Color scale indicates contribution magnitude (red: positive effect, blue: negative effect on interaction prediction)

DeepLIFT analysis identified a critical region spanning nucleotides 118–126 with substantially elevated contribution scores relative to the sequence background. Parallel tiling deletion experiments demonstrated marked interaction disruption upon removal of this sequence segment, with this deletion producing the largest effect among all tested windows. Systematic single-nucleotide mutagenesis provided independent validation of this region's functional significance, with the mutagenesis sensitivity map revealing pronounced effects of base alterations within the identified segment, particularly at positions corresponding to a RUNX3 binding motif.

The convergence of three independent analytical approaches—computational scoring, deletion analysis, and mutagenesis—identifying the same genomic region provides evidence that the model has learned biologically relevant sequence features. RUNX3 encodes a transcription factor with documented roles in chromatin organization and gene regulation, and its binding sites have been associated with chromatin loop anchors in multiple cell types. The identification of an established transcription factor binding site through model interpretation, without prior knowledge of binding site locations, indicates that deep learning frameworks can capture sequence determinants with demonstrated biological function. These findings illustrate how interpretability methods can identify functional regulatory elements from chromatin interaction patterns, offering a framework for discovering sequence features underlying three-dimensional genome architecture. It should be noted that the above analyses are primarily based on computational predictions and in silico perturbation experiments, and therefore do not directly demonstrate a causal role of the identified sequence regions in chromatin interactions under in vivo conditions.

Model robustness assessment through baseline comparison and distance-scale evaluation

To assess whether the model’s performance might stem from incidental learning of a particular type of feature, we used the GM12878 cell line as an example to tally the ratio of positive to negative samples in the test set and introduced random prediction and distance-only baselines for comparison. The results show that the test set contains 754 positive samples and 806 negative samples, indicating that the dataset is generally balanced. Under the same data conditions, we evaluated random guessing and the distance-based baseline and compared them with our method. The results indicate that for both random guessing and distance-only prediction, AUC, AUPR, Accuracy, F1, and Recall are all around 0.5 (Supplementary Fig. 5A–B). This further demonstrates that the performance gains achieved by our method are not accidental but arise from the model’s effective learning of biological features.

We further evaluated UniChrom's capability across different chromatin loop distance scales to address potential distance-dependent biases. Samples were stratified into three equal-sized groups based on genomic distance distribution: Small (< 439,718 bp), Medium (439,718–1,770,146 bp), and Large (> 1,770,146 bp) (Supplementary Fig. 6B). Performance evaluation revealed robust prediction across all distance categories (Supplementary Fig. 6A). Notably, UniChrom achieved the highest AUC (0.976) and accuracy (0.932) for large-distance interactions, even surpassing shorter-range performance. While AUPR, F1, and recall showed modest reductions for large-distance interactions (AUPR = 0.894, decreased by 0.086 relative to small interactions; F1 = 0.840, decreased by 0.043 and 0.090 relative to medium and small interactions; Recall = 0.850, decreased by 0.037 and 0.094), all metrics remained substantially high. These results demonstrate that UniChrom maintains reliable performance across the full spectrum of chromatin interaction distances without substantial degradation at long genomic ranges.

Cross-cell-line validation on independent HUVEC dataset

To assess generalization capability beyond the training cell lines, we evaluated UniChrom on HUVEC, an endothelial cell line not used during model development. This evaluation addresses whether models trained on lymphoblastoid, cancer, and fibroblast cell lines can accurately predict chromatin interactions in functionally distinct cell types.

We conducted comprehensive evaluation comparing both UniChrom-sequence and the full UniChrom model against all baseline methods on HUVEC data. For sequence-based prediction, UniChrom-sequence achieved AUC of 0.739 and AUPR of 0.753, compared to EPIshilbert (AUC: 0.728, AUPR: 0.715), FusNet (AUC: 0.703, AUPR: 0.687), SPEID (AUC: 0.712, AUPR: 0.678), and TargetFinder (AUC: 0.613, AUPR: 0.587) (Supplementary Fig. 7A). For multi-modal prediction, UniChrom achieved AUC of 0.962 and AUPR of 0.965, compared to ChINN (AUC: 0.920, AUPR: 0.909), CLNN_loop (AUC: 0.799, AUPR: 0.780), Enhancer_MDLF (AUC: 0.702, AUPR: 0.671), and EPI_Trans (AUC: 0.673, AUPR: 0.642) (Supplementary Fig. 7B).

The HUVEC results demonstrate that UniChrom maintains high performance on an unseen cell type, with AUC exceeding 0.96 for multi-modal prediction. The relative performance advantages over baseline methods remain consistent with observations on training cell lines, and the performance on HUVEC approaches levels achieved on training cell lines, indicating effective learning of generalizable chromatin organization principles.

Discussion

Chromatin interaction prediction has traditionally focused on specific interaction types or protein-mediated contacts, with limited integration across multiple regulatory layers. We developed UniChrom as a framework that combines DNA sequence, cell-type-specific epigenomic features, and genomic distance to predict chromatin interactions. Systematic evaluation across multiple cell lines demonstrated that this integrative approach provides more stable performance compared to models relying on individual feature modalities, particularly when predicting interactions in cell types with distinct regulatory landscapes.

Performance evaluation on an independent cell line (HUVEC) demonstrated that UniChrom's architectural design generalizes effectively to unseen cellular contexts. When trained and tested on HUVEC data, UniChrom achieved AUC of 0.962 for multi-modal prediction and 0.739 for sequence-based prediction, maintaining performance levels comparable to those observed on the primary cell lines (GM12878, IMR90, K562). In contrast, baseline methods showed varying degrees of performance on HUVEC, with some experiencing substantial decreases compared to their performance on other cell types. This consistency across functionally distinct cell types—from lymphoblastoid and cancer cells to fibroblasts and endothelial cells—indicates that UniChrom's integration strategy captures regulatory principles that generalize across cellular lineages. Cross-cell-type transfer experiments further demonstrated that models trained on one cell type can predict interactions in other cell types, though with reduced performance compared to cell-type-specific training, highlighting the value of cell-type-specific epigenomic information.

Feature importance analysis using DeepSHAP identified architectural proteins and histone modifications contributing to interaction predictions. CTCF and cohesin complex components (RAD21, SMC3) consistently ranked among the most important features across cell types, consistent with their established roles in chromatin loop formation through loop extrusion mechanisms [21, 34, 35]. The relative importance of these factors varied across cell lines—CTCF showed highest contribution in K562 (SHAP: 0.142), while RAD21 dominated in GM12878 (SHAP: 0.156) and IMR90 (SHAP: 0.109)—suggesting cell-type-specific utilization of architectural proteins. Histone modifications displayed cell-type-specific importance patterns, with H3K4me1 and H3K27ac showing elevated importance in contexts where enhancer activity is prominent. These modifications have documented roles in regulating enhancer activity and enhancer-promoter communication [36], while H3K36me3 is associated with transcriptional activity and chromatin states [37]. Sequence interpretability analysis identified regulatory motifs, including RUNX3 binding sites, at interaction anchor regions, demonstrating that the model learns sequence features associated with transcription factor binding and chromatin organization.

Although UniChrom achieves strong performance in predicting multiple types of chromatin interactions, several aspects merit further investigation. First, UniChrom relies to some extent on key epigenomic features, which may limit its transferability and generalization to cell types where such epigenomic information is missing or incomplete. Second, most current deep learning approaches—including UniChrom—model genomic sequences using fixed window sizes (5 kb in this study). However, recently proposed models such as Enformer and AlphaGenome have demonstrated that integrating substantially larger sequence contexts (approximately 100 kb to 1 Mb) is critical for capturing distal regulatory signals [38, 39]. Therefore, incorporating longer-range sequence context may further enhance the prediction of chromatin interactions spanning megabase-scale genomic distances. Additionally, chromatin interactions exhibit dynamic changes during cellular differentiation and development, with enhancer-promoter contacts forming progressively to activate developmental gene expression programs [40, 41]. Whether similar developmental dynamics characterize other interaction types, and how computational models might predict context-dependent interaction formation, represents an important direction for future investigation. Experimental validation of computationally predicted interactions, particularly for novel regulatory contacts not present in training data, will be essential for assessing functional relevance and refining prediction algorithms.

The integration of sequence-based and epigenomic approaches for chromatin interaction prediction provides a framework for investigating genome organization principles across cellular contexts. As experimental technologies continue to generate chromatin interaction data at increasing resolution and across diverse biological conditions, computational methods that effectively integrate multi-modal information will become increasingly valuable for understanding the regulatory architecture underlying gene expression programs in development and disease.

Supplementary Information

Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Supplementary Material 4. Supplementary Material 5. Supplementary Material 6. Supplementary Material 7. Supplementary Material 8. Supplementary Material 9. Supplementary Material 10.