CRM-TI: an enhanced pipeline for computationally assigning the target genes of cis-regulatory modules by considering comprehensive long-range regulation mechanisms

Abstract

Background

Metazoan transcriptional regulation is mediated by modular DNA segments termed cis-regulatory modules (CRMs). The precise identification of CRM target genes is essential for understanding cellular functions and disease mechanisms. While reporter assays can validate CRM targets, selecting candidates remains challenging because CRMs often act over long distances, bypassing nearby genes. Simple nearest-gene assignments are therefore unreliable, and methods based on epigenetic correlations are still limited. Enhancer-promoter interaction (EPI) prediction has improved CRM target identification. However, lots of existing EPI tools suffer from overfitting and overlook alternative mechanisms such as non-coding RNAs (ncRNAs) transcribed from CRMs.

Results

To address these gaps, we developed CRM-Target Identifier (CRM-TI), a deep learning pipeline that integrates mechanisms of chromatin interactions and CRM ncRNA-transcription-related gene regulation. Using a chromosome-wise partitioning scheme, CRM-TI achieved an unbiased test auROC of 90.0%, outperforming existing methods by at least 10%. Additionally, considering both CRM-gene interactions and ncRNA-transcription-related regulation further improved auROC performance by 21.9% over the model that considers only CRM-gene interactions.

Conclusions

These results demonstrate that combining comprehensive transcriptional regulatory mechanisms on CRMs provides a more complete framework for CRM-target gene assignment. CRM-TI is available at https://github.com/cobisLab/CRM-TI.

Supplementary information

The online version contains supplementary material available at 10.1186/s13040-026-00555-z.

Introduction

Transcriptional regulation in metazoans is cooperatively mediated by modular cis-regulatory elements, collectively referred to as cis-regulatory modules (CRMs) [1]. CRMs can be further classified into enhancers, promoters, and insulators [2]. Precise identification of the target genes of these elements is crucial for elucidating specific cellular functions and advancing the understanding of numerous diseases [3]. Consequently, characterizing CRM target genes and mapping transcriptional regulatory networks within cells are essential tasks.

Although the regulatory targets of CRMs can be verified in vivo through experiments such as reporter transgenic assays [2], selecting candidate genes for testing remains challenging, since regulatory sequences can act on genes located millions of base pairs away [4]. Reporter assays are widely used to demonstrate CRM activity, but they typically involve placing a candidate CRM upstream of a minimal promoter driving a reporter gene. Such designs confirm whether the CRM has the intrinsic ability to activate transcription but do not account for the native genomic context or long-range chromatin looping. Thus, reporter assays are most informative for testing CRM activity in isolation, whereas experiments such as CRISPR-based perturbations or chromatin conformation capture methods are generally required to unravel the regulatory mechanisms of CRM-target genes. A naive approach to assigning CRM targets is to link each CRM to its nearest gene. However, enhancers can bypass neighboring genes and distally regulate more distant targets via long-range interactions [5]. As a result, nearest-gene assignments or distance-based thresholds often fail to provide accurate predictions. Some algorithms attempt to improve accuracy by incorporating epigenetic signal correlations between enhancers and promoters [6]. Nevertheless, the performance of such correlation-based methods remains limited, as enhancers and promoters already share many similar epigenetic features [7].

An improved approach to assigning CRM target genes is through enhancer-promoter interaction (EPI) identification. Biologists have shown that distal CRM regulation can be mediated by chromatin interaction loops [8, 9]. By forming such loops, CRMs are brought into close spatial proximity with distal target genes, enabling transcriptional activation. Chromosome conformation capture experiments coupled with high-throughput sequencing (Hi-C) [10] or chromatin interaction analysis by paired-end tag sequencing (ChIA-PET) [11] can detect these chromatin interactions and thereby assign genes showing EPIs as targets of enhancer CRMs. Based on this concept, several computational tools have been developed to pre-screen potential EPIs and infer candidate target genes of enhancers [12–16]. However, it has been noted that many EPI prediction tools suffer from inflated performance due to the overlapping training and test sets in the TargetFinder dataset [17]. As a result, the reported usage of many EPI-based CRM-target gene assignment tools requires careful re-evaluation.

The limitations of existing methods for CRM-target gene assignment stem not only from poorly designed model training processes but also from incomplete representation of CRM regulatory mechanisms. Recent studies have shown that CRMs employ multiple mechanisms to control the transcriptional programs of their target genes. In addition to chromatin loop-mediated interactions, CRMs can also produce non-coding RNAs (ncRNAs), or the functional RNA molecules that do not encode proteins, to help stabilize chromatin loops and promote the transition of RNA polymerase II into productive elongation at promoters, thereby facilitating long-range gene regulation [18]. Enhancer RNAs (eRNAs) represent a subclass of ncRNAs transcribed bidirectionally or unidirectionally from active enhancer CRMs [18]. The discovery of eRNAs highlights mechanisms beyond EPIs that should be considered when computationally assigning CRM target genes. Thus, the absence of a comprehensive mechanistic framework in existing CRM target assignment tools leaves significant room for improvement. More sophisticated algorithms incorporating diverse regulatory mechanisms should be developed, accompanied by rigorously monitored model training processes.

In this study, we developed a deep-learning pipeline called CRM-Target Identifier (CRM-TI) to enhance the accuracy of CRM-target gene assignment. CRM-TI incorporates dual predictive blocks for comprehensive modeling of CRM regulatory mechanisms: one capturing direct CRM-gene interactions mediated by chromatin loops and another modeling the regulatory influence of ncRNAs transcribed from CRMs. We used Drosophila as a proof-of-concept system, with the framework designed to be transferable to mammalian systems. Using a chromosome-wise data partitioning scheme, CRM-TI achieved an unbiased test auROC of 90.0% for CRM-target gene assignment, significantly outperforming existing tools and baseline methods by at least 10%. Moreover, incorporating both CRM-gene interactions and CRM ncRNA-transcription-related gene regulation improved performance by 21.9% compared with the model that considers only CRM-gene interactions. These results demonstrate that integrating diverse transcriptional regulatory information provides a more comprehensive and mechanistically informed strategy for CRM-target gene prediction, offering valuable insights into the complex organization of transcriptional regulatory networks. CRM-TI is available at https://github.com/cobisLab/CRM-TI.

Data and methods

The ground-truth data used in this research

Three types of ground-truth data were prepared in this research for training and evaluating CRM-TI: the CRM-gene interaction ground-truth data, the non-overlapping CRM-target gene and non-target gene ground-truth data, and the CRM-CRM interaction data. A summary of all datasets used in this research is provided in Fig. 1. The Drosophila CRMs were downloaded from the REDfly database (version 9.61, dm6 with genome annotation version dm 6.48; downloaded on 2022/11/21).

Fig. 1.

The prepared ground-truth datasets used in this study. (a) Type 1 positive data: confident interacting CRM-gene pairs extracted form Hi-C experiments. (b) Type 1 negative data: confident non-interacting and distanced-matched CRM-gene pairs based on Hi-C experiments. (c) Type 2 positive data: non-overlapping CRM-target gene pairs in which the CRMs do not overlap with their target genes. (d) Confident non-interacting CRM-target gene pairs further extracted from the non-overlapping CRM-target gene pairs in Penal c). (e) Type 2 negative data: distance-matched CRM-non-target gene pairs based on literature mining

Type 1 ground-truth data (the interacting and non-interacting CRM-gene pairs)

To help obtain models capable of recognizing the interactions between CRMs and genes, we prepared the CRM-gene interaction ground-truth dataset consisting of confident interacting CRM-gene pairs (Fig. 1-a) and confident non-interacting CRM-gene pairs (Fig. 1-b).

Confident interacting CRM-gene pairs

CRMs can interact with the proximal promoter regions of genes through chromatin loop formation, thereby potentially influencing these genes [19]. We defined the proximal promoter region of a gene as 1,024 bp upstream and 128 bp downstream of its transcription start site. To identify CRM-gene interactions supported by confident experimental evidence, we downloaded and merged Hi-C data from male embryo cells generated by Pal et al. [20]. The accession numbers of these Hi-C experiments are provided in Table S1 of the Supplementary File. Hi-C chimeric read results were first mapped to the Drosophila genome (version v6.48) using BWA [21], and interacting segment pairs were computed with Pairtools [22]. Only pairs with “UU,” “UR,” or “RU” tags marked by Pairtools were retained. Default parameters and commands were applied in the above steps. Since Hi-C experiments are inherently noisy and require high read depths to achieve reliable resolution, we applied the following stringent criteria on the Hi-C results to define confident CRM-gene interactions (Fig. 1-a): (1) the chimeric read depth for a confident Hi-C pair had to be at least 7; and (2) at least 5 confident pairs were required to support interaction between a CRM and a gene’s proximal promoter. Using these filters, we identified 6,372 confident interacting CRM-gene pairs across all genes and experimentally verified CRMs in Drosophila. It should be noted, however, that a confident interacting CRM-gene pair does not necessarily indicate regulatory activity; additional experimental validation is required to confirm the CRM targeting regulation.

Confident non-interacting CRM-gene pairs

To confidently identify CRM-gene non-interacting pairs, we first defined loosely interacting interaction signals. Using the same processed Hi-C data described above, a CRM-gene pair was considered to have loosely defined interaction signals if at least one chimeric read pair (read depth 1) overlapped both the CRM and the proximal promoter region of the gene. Pairs with no loosely defined interaction signal were designated as confident non-interacting CRM-gene pairs. Since most non-interacting CRM-gene pairs have large genomic distances between the given CRMs and genes, we sampled a subset of these confident non-interacting pairs while maintaining a similar CRM-gene distance distribution, in 10,000 bp bins, to that of the confident interacting CRM-gene pairs. This procedure avoided introducing artificial distance bias into the dataset. After sampling, 5,988 confident non-interacting CRM-gene pairs were obtained (Fig. 1-b).

Type 2 ground-truth data (the non-overlapping CRM-target gene and CRM-non-target gene pairs)

To address the CRM-target gene assignment task, we gathered literature-curated CRM-target gene pairs. Specifically, manually curated Drosophila CRM-target gene pairs were downloaded from REDFly [23]. In total, 10,538 experimentally verified CRM-target gene pairs were retrieved from REDFly (version 9.61, genome version 6.48; downloaded on 2022/11/21). Chromosomal locations of genes (version v6.48) were retrieved from FlyBase [24].

For CRMs that directly overlapped their target genes, such as enhancers located within exons [25] or promoters within coding regions [26], we treated these as direct overlapping CRM-target gene pairs. To systematically capture the overlaps between CRMs and genes, gene body regions were extended to include the core promoter regions (defined as 128 bp upstream and downstream of the transcription start site). Among the 10,538 curated CRM-target gene pairs, 7,644 pairs involved direct overlaps between a given CRM and its target gene. In this study, we focused instead on non-overlapping cases, which are more challenging to identify their regulatory relationships. Hence, the remaining 2,894 non-overlapping CRM-target gene pairs were adopted as CRM-target gene ground-truth data in this research (Fig. 1-c).

To prepare CRM-non-target gene pairs, which served as negative samples, we employed natural language processing techniques on CRM-related literature. In our previous work, we developed a literature comprehension pipeline named DMLS [27] to analyze CRM-related articles. Using DMLS, we extracted all gene names from 438 REDFly-curated CRM articles. Because REDFly experts had already manually curated the true CRM-target genes in these articles, any gene names mentioned in the texts but not annotated as the targets of the CRMs under study were regarded as non-target genes of these CRMs. To further increase confidence in the CRM-non-target gene pairs, we applied one additional filter: the CRMs under study did not overlap with the mentioned genes in these CRM-related articles. Applying this criterion yielded 27,861 confident non-overlapping CRM-non-target gene pairs. However, most of these pairs have large genomic distances between the given CRMs and genes (median distance around 6677.4 kbp, in comparison with a median distance of only 4.5 kbp for the non-overlapping CRM-taget-gene pairs), which may introduce an artificial distance bias. To mitigate this, we sampled a subset of pairs from these 27,861 pairs while preserving a similar CRM-gene distance distribution, in 5,000 bp bins, to that of the non-overlapping CRM-target gene pairs. This resulted in a final set of 352 non-overlapping CRM-non-target gene pairs (Fig. 1-e).

Further categorization of the non-overlapping CRM-target gene pairs

To design regulation-mechanism-aware deep-learning blocks in CRM-TI, the CRM-target gene ground-truth data were further categorized. Specifically, we investigated the regulatory effects of chromatin loop formation between CRMs and their target genes. From the “Type 1 ground-truth data (the interacting and non-interacting CRM-gene pairs)” subsection, we obtained confident interacting and non-interacting CRM-gene pairs. By intersecting the non-overlapping CRM-target gene pairs with these confident interacting CRM-gene pairs and the confident non-interacting CRM-gene pairs separately, we further partitioned the non-overlapping CRM-target gene ground-truth pairs into three groups (Fig. 1-c): (1) Confident interacting CRM-target gene pairs: 386 of the 2,894 non-overlapping CRM-target gene pairs showed confident evidence of chromatin loop interactions. (2) Confident non-interacting CRM-target gene pairs (Fig. 1-d): among the remaining 2,508 pairs, 1,019 were supported as confident non-interacting pairs. (3) CRM-target gene pairs without clear evidence: the remaining 1,489 pairs lacked sufficient experimental evidence to be confidently classified as interacting or non-interacting. These unresolved pairs were not utilized in training the networks of CRM-TI. However, pairs located on Chr X within this 1,489 pairs were still retained in the non-overlapping CRM-target gene pair test set for evaluating CRM-target gene assignment.

Type 3 ground-truth data (the CRM-CRM interacting pairs and non-interacting pairs)

Since only a limited number of experimentally verified CRM-target gene pairs were available, the models in the CRM-TI pipeline were first pre-trained using independently prepared CRM-CRM interaction ground-truth data. By pre-training the deep-learning networks with the CRM-CRM interaction dataset and subsequently fine-tuning them on pairs with specific molecular mechanisms, we maximized the utility of both the literature-curated CRM-target gene pairs and the confidently extracted CRM-non-target gene pairs. The CRM-CRM interaction dataset was independently prepared from male and female embryonic Hi-C experiments from Pal et al. [20]. Briefly, this dataset comprises confident CRM-CRM interacting pairs and confident CRM-CRM non-interacting pairs. Confident interacting pairs were extracted by applying strict filtering criteria to the experiment results, requiring: (1) confident Hi-C pairing signals from chimeric reads with depths 10, (2) Hi-C pair types classified as UU/UR/RU by Pairtools [22], and (3) confident CRM-CRM interacting pairs require supports from at least six confident Hi-C pairs. Confident non-interacting CRM-CRM pairs were sampled from CRM pairs with no Hi-C chimeric read signal across all experimental runs. These non-interacting pairs were sampled by keeping the same CRM-CRM distance distribution as that of CRM-CRM interacting pairs. In total, 22,821 confident CRM-CRM interacting pairs and 10,782 confident CRM-CRM non-interacting pairs were collected, all restricted to CRMs located on chromosomes 2 R, 2 L, 3 L, and 3 R to avoid potential data leakage.

Characterization of basic properties of CRMs and CRM-target gene pairs in the ground-truth dataset

To better characterize the CRM-gene pairs considered in this study, we added additional analyses describing the basic properties of the ground-truth datasets. The distributions of CRM lengths and CRM-gene interaction distances are presented in Fig. S1 of the Supplementary File. In Fig. S1, S1-a shows the CRM length distribution for all CRMs considered in this study, while S1-b presents the CRM length distribution for CRMs involved in the non-overlapping CRM-target gene pairs. Fig. S1-c illustrates the CRM-gene distance distribution of the non-overlapping CRM-target gene pairs. Transcription factor binding site (TFBS) content within CRMs can provide additional characterization of the datasets. To address this point, we incorporated TFBS annotations from the REDfly database (version 9.7.0, downloaded on 2026/03/09, dm 6.48 genome annotation). These TFBS annotations were overlapped with CRMs included in the ground-truth datasets to examine the TF binding content within the regulatory elements used in this study. The resulting distributions of TFBS counts within CRMs from all CRM-target gene pairs and from the non-overlapping CRM-target gene pairs are presented in Fig. S1-d and Fig. S1-e, respectively, in the Supplementary File. These analyses provide a basic characterization of the contents of the CRMs considered in the ground-truth dataset.

Input epigenetic scores used by CRM-TI

Since previous studies have shown that epigenetic profiling is closely linked to chromatin contact formation and transcriptional regulation [28, 29], CRM-TI incorporates epigenetic data as input to its deep neural networks. In this study, we collected and processed relevant Drosophila datasets, including histone modification chromatin immunoprecipitation sequencing (ChIP-seq) experiments, chromatin-binding protein ChIP-seq experiments, and DNase-seq results. Histone modification and chromatin-binding protein ChIP-seq data were obtained from the modENCODE consortium [30], while DNase-seq datasets were downloaded from studies by Thomas et al. [31] and Mavrich et al. [32]. A complete list of the experiments used is provided in Table S2 of the Supplementary File. In total, CRM-TI incorporated 77 histone modification ChIP-seq experiments, 57 chromatin-binding protein ChIP-seq experiments, and 7 DNase-seq experiments.

For each experiment, base-resolution sequencing scores were computed using our previously established method [33, 34]. We briefly sketched the calculation of these scores. For ChIP-seq data, the score at each base was calculated as the logarithm of the ratio between total read coverage (in reads per million, RPM) at that base in the ChIP experiments and the corresponding coverage in the input controls. For DNase-seq, scores were derived by aggregating read counts (in RPM) at each base followed by logarithmic transformation. All sequencing scores were then normalized to a range between 0 and 1 using chromosome-wise min-max normalization, as described in our earlier work [33, 34].

The CRM-TI pipeline

The overview of CRM-TI

CRM-TI is a CRM-target gene assignment pipeline designed to account for the full range of known CRM regulatory mechanisms, encompassing regulation mediated by chromatin loops and regulation facilitated by ncRNAs transcribed from CRMs. In this study, CRM-TI specifically focuses on non-overlapping CRM-gene pairs. To achieve this goal, CRM-TI incorporates two predictive blocks to assist in CRM-target gene assignment (Fig. 2-a): (1) The CRM-gene interaction recognition block evaluates the presence of interactions between a CRM and the proximal promoter of a gene, providing evidence that the CRM may regulate the gene through chromatin loop formation. (2) The CRM ncRNA-transcription-related target gene regulation identification block identifies transcription events occurring on CRMs and assesses the regulatory potential of the resulting ncRNAs, which may stabilize loops or remodel promoter environments. This block thus helps reveal regulatory effects mediated by CRM-transcribed ncRNAs.

Fig. 2.

(a) Overview of the CRM-TI pipeline for CRM-target gene assignment. CRM-TI integrates two complementary blocks to predict CRM-target gene regulatory relationships. Outputs from the two blocks are combined using max-aggregation to produce final assignments. (b) The deep network architecture for estimating CRM-gene interaction probability. Input ChIP-seq-based feature matrices are processed through convolutional filters and multi-head attention layers, followed by pooling and fully connected layers, to yield the interaction probability between CRMs and gene proximal promoters. (c) The deep network architecture for estimating the probability that the ncRNA transcribed from the given CRM regulates the given gene. Selected transcription-related ChIP signals for CRMs and core promoters are encoded with convolutional filters and passed through dense layers, resulting in the predicted probability of the occurrence of CRM ncRNA-transcription-related gene regulation. In this figure, the “1D conv” and “fully connected” refer to sequences of 1D convolution operations and fully connected operations, respectively, with the corresponding numbers of kernels and hidden nodes listed alongside

The two blocks were both constructed using the convolution operations that help extract important epigenetic contents. Previous studies have shown that CRMs can exhibit characteristic chromatin signatures when performing their regulatory functions [34]. For example, promoters are typically enriched with marks such as H3K4me3, whereas enhancers are often associated with modifications such as H3K4me1 and other regulatory histone marks. In the CRM-TI framework, this distinction is implicitly addressed through the separate processing of CRM regions and gene promoter regions within the model architecture. Specifically, CRM and promoter inputs are passed through several convolutional layers that learn region-specific epigenetic patterns directly from the data. Because the model processes CRM and promoter region signals independently before integrating their representations, the convolutional filters can help capture the distinct epigenetic compositions characteristic of enhancers and promoters. Before the convolution operations in each block, the corresponding signal values for each given genomic regions are queried from the precomputed genome-wide epigenetic score tracks. The retrieved scores from all epigenetic datasets are then assembled to form the input feature matrices used by the convolution operations. This step, referred to as the ChIP score query in Fig 2, extracts epigenetic signal profiles for the specified genomic region from available experiments. The detailed architectures and training procedures of the two deep-learning blocks are described in the following subsections.

CRM-TI Block 1: CRM-gene interaction recognition

It is widely accepted that eukaryotic gene transcription can be facilitated by chromatin loops formed between distal CRMs and the proximal promoter regions of their target genes [35]. To capture this regulatory mechanism in this block, we focused on the proximal promoter region of each gene. The proximal promoter of a gene was defined as 1,024 bp upstream and 128 bp downstream of its transcription start site. Potential interactions between a CRM and the proximal promoter of a gene were identified through three steps (Fig. 2-b): regulatory code extraction, mutual feature merging, and interaction classification.

Regulatory code extraction

CRM-TI first constructs the input tensors from the collected epigenetic data scores (see the “Input epigenetic scores used by CRM-TI” section for details) for a given CRM () and the proximal promoter region of a gene (). Because most verified CRMs are approximately 512 bp in length, both CRM and promoter tensors are first resized to a length of 512, following previous recommendations [33]. Regions longer than 512 bp are bilinearly downsampled to 512 bps, while shorter regions are zero-padded to 512 bps. The resulting and tensors each have dimensions of 512 × 141, where 512 represents the fixed input length, and 141 corresponds to the number of epigenetic signal channels. In total, CRM-TI incorporates 77 histone modification ChIP-seq experiments, 57 chromatin-binding protein ChIP-seq experiments, and 7 DNase-seq experiments, resulting in 141 epigenetic tracks. These input tensors are then separately passed through a parameter-shared sequence of convolution operations to extract the underlying epigenetic code patterns. The sequence of convolution operations is computed as follows:

1

2

where is the 1-D convolution operator on tensor with kernels of size using stride 1 and same-padding, is the ReLU (rectified linear unit) activation function, and and are the extracted feature tensors for the given CRM and the promoter region, respectively. Note that the sequence of convolution operations applied to CRMs and the proximal promoter regions of given genes shares the same parameters in , meaning that the CRM and proximal promoter region inputs are processed using the same sequence of convolution operations. In other words, convolution kernels with the same parameters are applied to both inputs so that regulatory features are extracted in a consistent manner. The feature tensors, and , are provided for the next part.

Mutual feature merging

CRM-TI next evaluates the position-wise mutual correlations between the epigenetic codes of the CRM and the proximal promoter region of the gene. These positional mutual correlations are computed using a multi-head attention mechanism [36, 37]:

3

4

where is the 2-headed attention operation using embeddings of size 64, is the correlated tensor between and , and is the merged feature tensor. In the above formula, the epigenetic codes extracted from the given CRM are first attended to the epigenetic codes extracted from the proximal promoter region of the gene. A residual 1-D convolution layer then further distills the merged attended tensors and lessens the issue of gradient vanishing [38]. The two-headed attention operation, using embeddings of size 64, is as follows [39]:

5

The final attended and merged feature tensor integrates the epigenetic codes of the CRM and the proximal promoter region of the given gene and is then passed to the classification part for CRM-gene interaction recognition.

Interaction classification

The final attended and merged feature tensor is subsequently non-linearly transformed into a linearly separable space for CRM-gene interaction recognition:

6

where // are the trainable weight matrices of the fully connected layers all with hidden node size 128, is the trainable weight matrix of the last dense layer with hidden node size 2 for the softmax layer, represents the PReLU (parametric rectified linear unit), stands for the global average pooling operation, and is the recognition probability that the given CRM interacts with the proximal promoter region of the given gene.

Network training for CRM-TI block 1

To optimize the network, we first pre-trained CRM-TI Block 1 on the CRM-CRM interaction training-validation set and subsequently fine-tuned it on the CRM-gene interaction training-validation set. Importantly, both datasets excluded samples from Chr X during model pre-training and fine-tuning. Pre-training and fine-tuning of the CRM-gene interaction recognition block were performed using the Keras-based deep learning framework RDDL [40]. The following hyperparameters were applied during fine-tuning of this block: (1) training batch size = 64; (2) dropout rate = 0.4 after each fully connected layer; (3) training epochs = 60; (4) learning rate schedule = linearly warmed up to within the first 12 epochs, followed by exponential decay to before the 55th epoch; (5) loss function: categorical cross-entropy; and (6) parameters in the regulatory code extraction layers and multi-head attention module were frozen, while all others were updated during fine-tuning. Dropout rates were tuned to prevent overfitting, learning rates were selected to ensure stable optimization, and attention head sizes were chosen to capture diverse feature interactions. These values were determined through a combination of coarse search and small-range random search guided by cross-validation. All pre-training and fine-tuning steps were conducted on an NVIDIA RTX4080 GPU.

CRM-TI Block 2: CRM ncRNA-transcription-related gene regulation identification

Transcription occurring on CRMs can generate ncRNAs that help stabilize chromatin loop formation [41] and facilitate the transition of RNA polymerase II from pausing to elongation at the core promoter region of the gene [42–44]. In this block, the core promoter region was defined as 128 bp upstream and downstream of the transcription start site [45]. CRM-TI Block 2 is designed to identify ncRNA-transcription-related CRM-target gene regulation. The identification process involves two steps: regulatory code extraction and CRM ncRNA-transcription-related gene regulation identification (Fig. 2-c).

Regulatory code extraction

Since transcription within the CRM region and its gene regulation mechanism are the primary focus of this block, the input tensors for the CRM and the core promoter region of the gene are constructed using only transcription initiation-related signals. These include ChIP-seq results for RNA Polymerase II CTD, RNA Polymerase II, H3K4me1, and H3K23ac across different developmental stages. While H3K4me3 is widely recognized as a canonical promoter-associated mark linked to active transcription initiation, other histone modifications have also been reported to correlate with transcription-related regulatory activity in the genome. In particular, H3K4me1 has been associated with active and poised regulatory regions, including enhancers, and has been implicated in transcription-coupled chromatin regulation [46]. H3K23ac has also been reported to participate in transcription-related chromatin modifications and is functionally coupled with other histone acylation processes involved in transcription regulation [47]. While H3K4me3 is indeed a well-established transcription initiation marker, the inclusion of other transcription-related histone modifications, such as H3K4me1 and H3K23ac, allows the model to capture complementary regulatory signals associated with CRM activity. Details of the included experiments are listed in Table S3 of the Supplementary File. The input tensors for a given CRM and the core promoter region of the gene are resized to a length of 256, following the same procedures used in the CRM-gene interaction recognition block, yielding the and tensors, respectively. The choice of 256 reflects the defined length constraints of core promoter regions. Both and tensors have dimensions of 256 × 15. The sequences of convolution operations for regulatory code extraction are defined as follows:

7

8

where and have the same definitions as in Block 1 of CRM-TI. Because the CRM and the core promoter region of the given gene may exhibit distinct transcriptional patterns, the convolutional sequences for these two regions were designed with separate parameter sets. These independent convolutional operations produce the regulatory code tensors and for the given CRM and the core promoter region of the given gene, respectively.

CRM ncRNA-transcription-related gene regulation prediction

The extracted regulatory code tensors and are summed to generate a single transcription feature tensor. This merged tensor is then passed through multiple fully connected layers, projecting it into a high-dimensional space to enable identification of ncRNA-transcription-related regulation from the CRM on the given gene:

9

where and are the trainable weight matrices of the dense layers with hidden node size 16 and 8, is the trainable weight matrix of the last fully connected layer with hidden node size 2 for the softmax layer, is the PReLU activation function, and is the probability of the occurrence of CRM ncRNA-transcription-related gene regulation.

Network training for CRM-TI Block 2

The CRM ncRNA-transcription-related gene regulation identification block (CRM-TI Block 2) focuses on identifying the regulatory effects of ncRNAs transcribed from CRMs on their target genes. Because only a limited number of experimentally verified non-interacting, non-overlapping CRM-target gene pairs were available (Fig. 1-d and -e), we addressed this challenge by restricting the input tensors to transcription initiation-related epigenetic features and pre-training the deep network on CRM-CRM interaction pairs. The final model was then fine-tuned through cross-validation on the curated non-interacting CRM-target gene pair training-validation set. Details of the pair set are summarized in the “Summary of the prepared sets used in different tasks” section. Both the CRM-CRM interaction pre-training set and the non-interacting CRM-target gene training-validation set excluded samples from Chr X. Pre-training and fine-tuning were also carried out using the RDDL framework [40]. The following hyperparameters were taken during fine-tuning: (1) training batch size = 128; (2) dropout rate = 0.6 after each fully connected layer; (3) training epochs = 120; (4) learning rate schedule = linearly warmed up to during the first 24 epochs, followed by exponential decay to before epoch 109; (5) loss function: categorical cross-entropy; and (6) parameters in the regulatory code extraction layers were frozen, while weights in the CRM ncRNA-transcription-related gene regulation prediction layers were fine-tuned. All pre-training and fine-tuning processes of this block were performed on an NVIDIA RTX4080 GPU.

The summary block

After considering the two aspects of CRM regulatory mechanisms, the final targeting probability of the CRM on the given gene is aggregated using the following formula:

10

where is the recognition probability of CRM-gene interaction, is the identification probability of the occurrence of CRM ncRNA-transcription-related gene regulation, and is the final probability that the CRM targets and regulates the gene. Conceptually, taking the maximum of and corresponds to a logical OR rule: a CRM-gene pair is predicted as positive if either CRM ncRNA-transcription-related gene regulation or chromatin interaction provides strong evidence. In the final CRM-TI pipeline, we adopted a threshold of for CRM-target gene assignment. When the aggregated for a given CRM-gene pair exceeds this threshold, the pair is assigned as a CRM-target gene pair.

Performance evaluation metrics for CRM-target gene assignment

Assigning the target genes of given CRMs can be framed as a standard classification problem. Therefore, to evaluate the performance of CRM-TI and to compare it with related works, we computed the confusion matrix that produces the following evaluation metrics [48–50]:

11

12

13

14

where TP (true positive) counts the number of correctly identified CRM-target gene pairs, FP (false positive) sums the number of random CRM-gene pairs mistakenly assigned as targeting pairings, TN (true negative) calculates the number of correctly checked random CRM-gene pairings, and FN (false negative) is the number of missed CRM-target gene pairs. A default probability threshold of 0.5 was used for final CRM-target gene assignments. For comparisons between CRM-TI and other tools and baseline methods, we employed receiver operating characteristic (ROC) curves to minimize threshold bias across different methods. ROC curves were generated by plotting the recalls against the corresponding false positive rates (FPR) across varying thresholds. To quantify the results from ROC curves, the areas under the ROC curves (auROC) were integrated. A higher auROC value indicates stronger intrinsic predictive performance. Note that precision-recall (PR) curves were not used for comparison because positive samples (CRM-target gene pairs) substantially outnumbered negative samples (CRM-non-target pairs) in the ground-truth dataset, creating a high prevalence rate. In such cases, elevated auPRC values (area under PR curves) can be trivially achieved and thus are not informative.

We adopted the following rationale for metric selection throughout the rest of the article. The overall performance of the CRM-TI pipeline and its ablation variants was reported in full using a comprehensive set of metrics (auROC, auPRC, F1, precision, recall, and specificity) to provide a complete view of the behavior of CRM-TI. In learning curves, the F1-score values from the training/validation folds across epochs were used to monitor model convergence and over-fitting during fine-tuning. ROC curves for both validation and independent test sets will then be used to assess the threshold-free generalizability and final performance of the model. Finally, performance comparisons across different methods and aggregation strategies were based on auROC, as auROC is threshold-independent and therefore enables fair comparison across models with different optimal decision cut-points.

Results and discussions

Summary of the prepared sets used in different tasks

In this study, we collected three genres of ground-truth data to construct and evaluate the CRM-target gene assignment pipeline. As detailed in the “Data and Methods” section, these datasets include: interacting and non-interacting CRM-gene pairs, non-overlapping CRM-target gene and CRM-non-target gene pairs, and CRM-CRM interaction data. Based on these three datasets, we prepared the following data groups for model training and evaluation: (1) CRM-gene interaction pair set (Type 1 ground-truth dataset): used to fine-tune the deep network for CRM-gene interaction recognition (CRM-TI Block 1, CRM-gene interaction recognition). It includes 6,372 confident interacting CRM-gene pairs as positives (Fig. 1-a) and 5,988 confident non-interacting CRM-gene pairs as negatives (Fig. 1-b). Notably, interacting CRM-gene pairs may or may not correspond to CRM-target gene pairs. (2) Non-overlapping CRM-target gene pair set (Type 2 ground-truth dataset): used to design and evaluate the final CRM-TI pipeline for CRM-target gene assignment. This set consists of 2,894 non-overlapping CRM-target gene pairs as positive samples (Fig. 1-c) and 352 sampled CRM-non-target gene pairs as negative samples (Fig. 1-e). (3) Non-interacting CRM-target gene pair set (positives from the Type 2 subset lacking any Hi-C signals, with the same negatives from the Type 2 dataset): used to fine-tune the network for detecting CRM ncRNA-transcription-related target gene regulation (CRM-TI Block 2). This set encompasses 1,019 confident non-interacting CRM-target gene pairs as positives (Fig. 1-d) and 352 sampled CRM-non-target gene pairs as negatives (Fig. 1-e). (4) CRM-CRM interaction pair set (Type 3 ground-truth dataset): used for pre-training both deep networks in CRM-TI to mitigate data sparsity. Excluding CRMs on chromosome X, this set contains 22,821 CRM-CRM interacting pairs as positives and randomly selected non-interacting 10,782 CRM-CRM pairs as negatives.

In all prepared data groups, pairs from chromosomes 2 R, 2 L, 3 R, and 3 L were used for model training, whereas all pairs from chromosome X were reserved exclusively for testing to prevent data snooping. The detailed statistics of the training-validation and test sets for each data group are summarized in Table S4 of the Supplementary File.

Because the CRM-target gene ground-truth set was relatively small, we adopted a transfer learning strategy in CRM-TI by pre-training the models followed by fine-tuning. Specifically, the networks were first pre-trained on a general CRM-CRM interaction dataset defined by Hi-C signals (see section “Type 3 ground-truth data (the CRM-CRM interacting pairs and non-interacting pairs)”). The CRM-CRM interaction dataset provided abundant transcriptional regulatory signals for effective pre-training. To avoid overfitting, we applied the above chromosome-wise partitioning scheme for cross-validation throughout the pre-training and fine-tuning processes. We pre-trained two deep networks (Block1: the CRM-gene interaction recognition block and Block 2: the CRM ncRNA-transcription-related gene regulation identification block), each with different input information. These pre-trained networks were then fine-tuned on their respective CRM-gene datasets. This strategy enabled CRM-TI to maintain robust training performance even with the limited number of CRM-target gene samples available.

CRM-TI can accurately identify the target genes of CRMs

CRM-TI is a pipeline that integrates current knowledge of CRM regulatory mechanisms to assign CRM-target genes. To achieve this, two deep network blocks were designed. The first block is the CRM-gene interaction recognition model (CRM-TI Block 1), which evaluates whether a CRM forms an interaction loop with the proximal promoter region of a given gene. The second block is the CRM ncRNA-transcription-related gene regulation identification model (CRM-TI Block 2), which assesses whether CRM-transcribed ncRNAs contribute to stabilizing CRM-gene chromatin loops or to remodeling epigenetic marks around the gene’s core promoter. Together, these blocks provide a comprehensive framework for modeling CRM-target gene regulation.

The two prediction blocks were first independently pre-trained on the CRM-CRM interaction pair set using the epigenetic scores selected for each model. They were then fine-tuned separately for their respective tasks. To ensure model convergence and proper parameter fitting for both blocks, we applied learning curve techniques that tracked the average performance across training-validation folds, using chromosome-wise cross-validation on all chromosomes except Chr X during both pre-training and fine-tuning. For performance evaluation, we first assessed the task-specific accuracy of each block independently, followed by an evaluation of the overall performance of CRM-TI in CRM-target gene assignment using comprehensive performance metrics. The detailed fine-tuning results for the two blocks and the overall CRM-target gene assignment performance were presented in the following subsections.

The CRM interaction recognition block (CRM-TI Block 1) accurately identifies interactions between CRMs and the promoter regions of genes

The first block in CRM-TI is the CRM-gene interaction recognition model. This block identifies CRM-gene pairs connected by direct chromatin contacts between CRMs and the proximal promoter regions of genes. To achieve this, CRM-TI Block 1 was fine-tuned on the CRM-gene interaction pair set (Fig. 1-a and b), which comprises confident interacting and non-interacting pairs from chromosomes 2 L, 2 R, 3 L, and 3 R. As shown in the learning curves (Fig. 3-a), the model converged and was well-fitted during training. The average validation performance on CRM-gene interaction recognition reached an auROC of 87.2% (Fig. 3-b), revealing the capacity of CRM-TI Block 1 to detect direct chromatin contacts between CRMs and gene proximal promoters. In addition, the test auROC for CRM-gene interaction identification was 90.7%, closely matching the validation results and indicating good model generalization. It should be noted, however, that CRM-gene interaction identification does not necessarily equate to CRM-target gene assignment, a distinction further examined in the “Integrating mechanisms of CRM-gene interaction and ncRNA-transcription-related gene regulation can boost CRM-target gene identification” section.

Fig. 3.

The training process and performance evaluation of CRM-TI Block 1 for CRM-gene interaction recognition using the CRM-gene interaction pair data. (a) The fine-tuning learning curves. The model converges rapidly, with F1-scores stabilizing around epoch 20 and maintaining consistent performance thereafter, indicating that the model generalizes well without signs of overfitting. (b) The validation (cyan curves) and test (red line) ROC curve results for CRM-gene interaction recognition. The model exhibits high sensitivity across a broad range of false positive control thresholds, with validation and test performance closely aligned

The ncRNA-transcription-regulated CRM-target gene identification block (CRM-TI Block 2) identifies CRM-target-gene pairs that lack CRM-gene interactions

The second block in CRM-TI identifies CRM ncRNA-transcription-related gene regulation. This block was fine-tuned on the non-interacting CRM-target gene pair set (Fig. 1-d and e) to extract CRM-target pairs potentially regulated by ncRNAs transcribed from CRMs. The learning curves of CRM-TI Block 2 (Fig. 4-a) on the training-validation folds of this dataset showed convergence without signs of overfitting. Because data imbalance was addressed using a random over-sampling technique, the validation results used for over-fitting confirmation in the learning curves were based on the over-sampled folds [40]. Notice that in Fig. 4-a, the F1 value increments between the end of Epoch 1 and the end of the last Epoch reflected the fast convergence of the fine-tuning process, rather than direct performance gain by the fine-tuning process. On average, CRM-TI Block 2 achieved an auROC of 83.2% on the (original) validation data (Fig. 4-b), manifesting its ability to identify CRM-target gene pairs that might be regulated by CRM-transcribed ncRNAs. The test set performance reached an auROC of 88.4% on the reserved Chr X pairs, further confirming the generalization power of CRM-TI Block 2 for predicting CRM ncRNA-transcription-related gene regulation.

Fig. 4.

The training process and performance evaluation of CRM-TI Block 2 for CRM ncRNA-transcription-related gene regulation identification using the non-interacting CRM-target gene pair set. (a) The fine-tuning learning curves. Model performance improves rapidly within the first 20 epochs and stabilizes thereafter, indicating convergence without overfitting. (b) The validation (light blue) and test (green) ROC curve results for CRM ncRNA-transcription-related gene regulation identification. The test results achieve a good auROC value, demonstrating that CRM-TI Block 2 effectively captures CRM ncRNA-transcription-related gene regulation

We further explicitly compared the performance of the pre-trained Block 2 model and the fine-tuned CRM-TI Block 2 model on the non-interacting CRM-target gene pair test set (from Fig. 1-d and e) to calculate the performance gain from fine-tuning. The pre-trained Block 2 model achieved a test auROC of only 63.0% (See Supplementary Fig. S2a) in comparison with the 88.4% auROC of the fine-tuned version for identifying CRM-target-gene pairs that lack CRM-gene interactions, confirming that fine-tuned adaptation to the specific biological context of CRM ncRNA-transcription-related gene regulation is essential and novel in CRM-TI. Notice that although the epigenetic signal distributions used for pre-training CRM-TI Block 2 are not identical to those encountered during fine-tuning, prior studies have shown that ncRNA transcription at CRMs can also facilitate the formation of CRM-CRM interactions [18], making this selection of pre-training task to share characteristic epigenetic patterns (e.g., Pol II occupancy and active histone marks) with the CRM ncRNA-transcription-related target gene regulatory processes. While more task-matched pre-training data could potentially further facilitate the convergence of CRM-TI Block 2 during fine-tuning, the purpose of pre-training is not to learn a fully task-complete model, but rather to provide a stable and biologically informed initialization for subsequent fine-tuning on scarce pairs. The shared epigenetic patterns allow pre-training to capture coarse regulatory features that can be easily refined during fine-tuning. Consistent with this observation, the pre-trained Block 2 model alone achieved limited predictive power, whereas fine-tuning is essential for adapting the model to CRM ncRNA-transcription-related regulatory mechanism identification. Hence, pre-training serves primarily as a starting point for subsequent parameter optimization, while fine-tuning is the critical step for learning CRM-specific regulatory signals.

Note that since the non-interacting CRM-target gene pair set used for fine-tuning CRM-TI Block 2 consists of only 1,019 positive CRM-target gene pairs (confident non-interacting CRM-target gene pairs) and 352 negative CRM-non-target gene pairs, this scale is insufficient to reliably train a generalizable deep network from random initialization. When using random initialization, the model trained from random-initialization achieved a test auROC of only approximately 57.5% (See Supplementary Fig. S2-b), indicating near-random predictive performance. By contrast, pre-training on the Type 3 CRM-CRM interaction dataset, which contains 22,821 confident CRM-CRM interacting pairs and 10,782 confident CRM-CRM non-interacting pairs, provided a biologically informed and statistically robust initialization that stabilized optimization and enabled subsequent fine-tuning. This pre-training accordingly enabled the fast convergence of the fine-tuning process, as indicated in Fig. 4-a.

Overall CRM-target gene assignment performance by integrating the two blocks in CRM-TI

After validating the task-specific performance of the two prediction blocks, we next evaluated the overall CRM-target gene assignment performance of the complete CRM-TI pipeline. By incorporating multiple CRM regulatory mechanisms, CRM-TI was expected to achieve strong assignment accuracy. On the non-overlapping CRM-target gene pair test set (from the Type 2 ground-truth dataset), which encompasses diverse CRM regulatory mechanisms, CRM-TI achieved a final test auROC of 90.0% (Fig. 5), indicating robust CRM-target gene assignment regardless of the CRM regulatory modes. Additional performance metrics for the full CRM-TI pipeline are listed in Table 1. These test results confirmed our expectations. In summary, CRM-TI represents a well-trained deep-learning pipeline capable of accurately identifying CRM-target gene pairs across diverse regulatory mechanisms.

Fig. 5.

The ROC performance comparison among CRM-TI, the model with only the CRM-gene interaction recognition block, and the model with only the CRM ncRNA-transcription-related gene regulation identification block on the non-overlapping CRM-target gene pair test set. The full CRM-TI pipeline (blue line), which integrates both the CRM-gene interaction recognition block and the CRM ncRNA-transcription-related gene regulation identification block, achieves the highest auROC. When the model considers only the CRM ncRNA-transcription-related regulation mechanism (green line), predictive performance remains strong but slightly reduced. By contrast, when only the CRM-gene interaction recognition block is used (red line), performance drops markedly, underscoring that CRM proximity alone might be insufficient for accurate CRM-target gene assignment

Table 1.

The CRM-target gene assignment performance of the full CRM-TI pipeline compared with its individual components (CRM-TI Block 1 and CRM-TI Block 2) on the non-overlapping CRM-target gene pair test set. CRM-TI Block 1: CRM-gene interaction recognition block; CRM-TI Block 2: CRM ncRNA-transcription-related gene regulation identification block

Pipelines	auROC	auPRC	F1	Precision	Recall	Specificity
The whole CRM-TI pipeline	90.0%	97.9%	0.909	0.963	0.861	0.781
Model with only CRM-TI Block 1	68.1%	93.1%	0.526	0.950	0.364	0.875
Model with only CRM-TI Block 2	88.7%	97.8%	0.826	0.968	0.720	0.844

Note that in Table 1, the specificity of the whole CRM-TI pipeline (0.781) is lower than that of the individual ablation models (Model with only CRM-TI Block 1: 0.875; Model with only CRM-TI Block 2: 0.844). The reduced specificity of the full CRM-TI pipeline reflects the intentional design choice of integrating two complementary regulatory mechanisms via a union (max-aggregation) strategy. Specifically, CRM-TI assigns a CRM-gene pair as a positive regulatory pair if either the CRM-gene interaction recognition block or the ncRNA-transcription-regulated gene regulation identification block provides strong evidence. This aggregation rescues true CRM-target gene pairs that are supported by one regulatory mechanism but not the other, thereby substantially increasing recall relative to either individual block. As a consequence of this union-based integration, specificity is expected to decrease modestly, reflecting a standard recall-specificity trade-off. Importantly, despite this trade-off, the overall discriminative performance of CRM-TI remained superior, as evidenced by its higher auROC (90.0%) compared with either Block 1 or Block 2 alone. This indicated that CRM-TI provides a more accurate global ranking of CRM-target gene pairs with the least recall-specificity trade-off. Note that specificity can be increased to levels comparable to the individual blocks by adjusting the decision threshold, at the cost of a moderate reduction in recall. Because CRM-TI outputs continuous probabilities, users can readily select an operating point that best suits their application (e.g., high-confidence predictions versus broader hypothesis generation).

Chromosome-wise data partitioning prevents data leakage and eliminates systematic bias in the performance evaluation

In this study, all training steps deliberately excluded pairs from Chr X, which was held out entirely as the independent test set. Specifically, CRM-gene pairs from Chr X were reserved for testing, while pairs from Chr 2 L, 2 R, 3 L, and 3 R were assigned to the first, second, third, and fourth training-validation folds, respectively. This chromosome-wise partitioning scheme ensured that test set reservation and cross-validation fold preparation avoided data leakage, while also enabling independent evaluation of each autosome via the cross-validation folds. In contrast, random genome-wide assignment could introduce data leakage through overlapping CRMs or shared genes among the test set and training-validation folds and artificially inflate performance estimates. This issue has been previously comprehensively discussed by Cao and Fullwood [17], who demonstrated that random splitting can lead to overly optimistic performance estimates in enhancer-promoter interaction prediction because highly similar genomic regions from the same loci may appear in both training and test sets. By using a completely held-out chromosome through chromosome-wise data partitioning, we ensured strict separation between the training/validation data and the test data, thereby eliminating potential overlap in regulatory regions and preventing leakage of local epigenetic patterns. Although each chromosome has distinct genomic characteristics (e.g., size and gene density), using Chr. X as the independent test set provides a more conservative and realistic assessment of model generalization, as the model must make predictions in a chromosomal background that was not seen during training.

As shown in Fig. 3-a (CRM-gene interaction recognition) and Fig. 4-a (CRM ncRNA-transcription-related gene regulation identification), the averages of the training and validation F1 values tracked closely and converged, which are standard indicators that the two models were not overfitted. Moreover, both CRM-TI Block 1 (CRM-gene interaction recognition) and CRM-TI Block 2 (CRM ncRNA-transcription-related gene regulation identification) yielded comparable auROC values between validation averages and test results, further confirming the absence of overfitting and demonstrating that the models generalize well across autosomes and the sex chromosome. Together, these safeguards ensured that the two blocks of CRM-TI avoided overfitting, while the consistent block performance across Chr X and autosomal cross-validation folds indicated that no systematic bias was introduced during data partitioning in the blocks of CRM-TI. The test results on Chr X and the average of cross-validation results on autosomal chromosomes provided a stringent and unbiased evaluation of the performance and generalizability of CRM-TI.

Architecture differences between Block 1 and Block 2 in CRM-TI

CRM-TI Block 1 (CRM-gene interaction recognition) and Block 2 (ncRNA-transcription-regulated CRM-target gene identification) are designed to model distinct regulatory mechanisms and therefore employ different input feature sets and input tensor lengths, which makes their architectural components, particularly convolutional kernel sizes, not directly comparable. Block 1 processes a broad spectrum of epigenetic signals over CRMs and gene proximal promoter regions to capture chromatin loop-mediated gene regulation, whereas Block 2 focuses specifically on ncRNA-transcription-related signals (RNA Pol II signals and transcription-related histone marks) over CRMs and gene core promoter regions to model ncRNA-associated regulatory effects. As a result of the different feature sets, genomic region focuses, and input tensor lengths, the effective receptive fields and signal characteristics differ substantially between the two blocks. Consequently, the convolutional kernel sizes in the two blocks did not need to be matched across blocks, but were instead determined independently via cross-validation in each task to best suit the corresponding genomic region resolution and biological objective. This independent optimization ensures that each block is appropriately tuned for its specific regulatory mechanism, rather than constrained by architectural symmetry that would not be biologically or computationally justified. The differences between CRM-TI Block 1 and CRM-TI Block 2 reflect their distinct functional roles rather than inconsistencies in model design.

CRM-TI is robust against the distance-match sampling of the negative samples

The steps in constructing the negative dataset were to produce a distance-bias-free CRM-non-target gene pairing set. In preparing the negative ground-truth dataset, CRM-non-target gene pairs were distance-matched to the positive non-overlapping CRM-target gene pairs, an essential step for eliminating strong distance-driven confounding effects. The majority of NLP-mined CRM-non-target gene pairs were separated by large genomic distances (median distance around 6677.4 kbp, in comparison with a median distance of only 4.5 kbp for the non-overlapping CRM-taget gene pairs). Training models on such a distance-mismatched set would introduce an artificial distance bias to the model, in which simple proximity-based heuristics dominate performance and obscure whether regulatory mechanisms are learned. Distance matching thus enables a bias-controlled training and evaluation setting that more accurately reflects the model’s ability to capture biological mechanisms. To further evaluate the robustness of CRM-TI in a distance-unmatched screening scenario, we performed an additional evaluation using the non-overlapping CRM-target gene pair test set that replaced the negative samples with the unsampled (without distance-matching) set of CRM-non-target gene pairs located on Chr X (test set from the full 27,861 pairs). In this setting, CRM-TI still achieved a test auROC of 87.3% (see Table 2), demonstrating that CRM-TI remains effective when confronted with a large, highly unmatched negative set. Other evaluation metrics for this setting reported similar conclusions (also see Table 2 for all evaluation metrics). Notice that the low auPRC, F1, and precision stem from the low prevalence in the unsampled (distance-unmatched) test set. Overall, distance-matching during the construction of negative CRM-gene pairs enabled unbiased learning in CRM-TI while maintaining robustness across genome-wide screening.

Table 2.

The CRM-target gene assignment performance of the full CRM-TI pipeline on the non-overlapping CRM-target gene pair test set and the test set with distance-unmatched negative samples

The whole CRM-TI pipeline	auROC	auPRC	F1	Precision	Recall	Specificity
On the distance-matched test set	90.0%	97.9%	0.909	0.963	0.861	0.781
On the un-sampled test set	87.3%	44.5%	0.385	0.248	0.861	0.718

Integrating mechanisms of CRM-gene interaction and ncRNA-transcription-related gene regulation can boost CRM-target gene identification

In this section, we further investigated the biological and mechanistic aspects of the designed pipeline. In CRM-TI, two complementary blocks were designed to capture distinct CRM regulatory mechanisms and thereby enable comprehensive identification of CRM-target gene pairs. These two blocks are the CRM-gene interaction recognition block and the CRM ncRNA-transcription-related gene regulation identification block. The CRM-gene interaction recognition block recognizes regulatory effects mediated by chromatin loops between a CRM and its target gene. On the other hand, the CRM ncRNA-transcription-related gene regulation identification block identifies regulatory effects arising from transcription on the CRM itself, which may generate ncRNAs that contribute to stabilizing chromatin loops or remodeling epigenetic states of their target genes. In the previous section, we demonstrated that both blocks achieved strong performance on their respective objective tasks. We now examine their ability to identify CRM-target gene pairs with diverse regulatory mechanisms.

For this purpose, we compared the performance of the full CRM-TI pipeline with models that used only the CRM-gene interaction recognition block or only the CRM ncRNA-transcription-related gene regulation identification block on the non-overlapping CRM-target gene pair test set (Type 2 ground-truth data, as shown in Fig. 1-c and -e), which contains pairs regulated through diverse mechanisms. The test results are summarized in Fig. 5 and Table 1. Using only the CRM-gene interaction recognition block yielded an auROC of 68.1% on this test set, while using only the CRM ncRNA-transcription-related gene regulation identification block achieved an auROC of 88.7%. By comparison, the complete CRM-TI pipeline reached an auROC of 90.0%, representing a 21.9% improvement over the interaction-only model and a 1.3% boost over the ncRNA-only model. These findings reveal that chromatin interactions alone may be insufficient to predict CRM-target gene relationships (recall = 36% for the interaction-only model), as some CRMs can exert regulatory effects on the genes without clear physical contacts with their proximal promoters. By integrating CRM-gene interactions and CRM ncRNA-transcription-related gene regulation, CRM-TI provides more accurate target gene assignments, suggesting the notion that ncRNA-transcription-related gene regulation might play an essential role in long-range CRM regulation. The large proportion of confident non-interacting CRM-target gene pairs in the ground-truth dataset (Fig. 1-d) further reinforces this observation. In summary, CRM-TI integrates complementary CRM regulatory mechanisms and achieves state-of-the-art performance in assigning the target genes of CRMs.

The formation of chromatin loops and the regulatory effects of ncRnas transcribed from CRMs might be complementary in CRM regulatory programs

In the CRM-TI deep-learning pipeline, results from the CRM-gene interaction recognition block and the CRM ncRNA-transcription-related gene regulation identification block were originally combined by taking the maximum value of the prediction probabilities from these two blocks. In this subsection, we investigated alternative strategies for aggregating the probabilities from the two blocks. Specifically, we also evaluated the minimum aggregation approach, in which the lower of the two prediction probabilities was taken. These aggregation methods highlight different biological perspectives. The maximum aggregation method for the two prediction blocks reflects the complementary effects of chromatin loop formation and the regulatory functions of CRM-transcribed ncRNAs. In contrast, the minimum aggregation method requires that both chromatin loop formation and ncRNA-mediated regulation be present to assign a target gene of a CRM. We evaluated the performance of the minimum aggregation strategy on the non-overlapping CRM-target gene pair test set. As shown in Fig. 6, the maximum aggregation approach (CRM-TI, auROC = 90.0%) outperformed the minimum aggregation approach (auROC = 76.2%) by 13.8%. The result suggested that chromatin loop formation and the regulatory activity of CRM-transcribed ncRNAs represent independent and complementary mechanisms in metazoan species.

Fig. 6.

The ROC performance of different aggregation strategies for the two prediction blocks in CRM-TI on the non-overlapping CRM-target gene pair test set. CRM-TI (blue line) applies max-aggregation across its two predictive blocks, ensuring the existence of either CRM-gene interactions or ncRNA-transcription-related CRM-target gene regulation for each CRM-gene pair. In contrast, the alternative approach (brown line) uses min-aggregation of the two blocks, requiring both mechanisms to be observed for the given pair. Max-aggregation in CRM-TI achieves a higher auROC value, indicating CRM-gene interactions and CRM-derived ncRnas act as complementary mechanisms

Although the mechanisms considered in the CRM-TI pipeline have been validated previously, biological confirmation of the new CRM-gene targeting predictions remains essential for establishing their functional relevance. In CRM-TI, two probabilities are provided for each CRM-gene pair: the CRM-gene interaction probability and the CRM ncRNA-transcription-related gene regulation probability. These values highlight distinct potential regulatory mechanisms for a given CRM-gene pair. Direct functional testing of these targeting relationships could be performed using reporter assays (e.g., luciferase or GFP constructs). Building on the predictions from these two biological aspects, more targeted validation can be achieved with CRISPR-based perturbations. For example, silencing CRMs to disrupt enhancer-promoter looping or mutating transcription initiation sites within CRMs to specifically block ncRNA transcription, followed by RNA-seq or PRO-seq, would allow measurement of downstream expression changes. More detailed molecular observations on these two mechanisms can be considered based on these analyses.

In the above results, we focused on max and min aggregation because these two strategies have direct biological interpretations in the context of CRM regulation: max aggregation corresponds to the scenario in which either chromatin looping-related or ncRNA-transcription-associated regulation is sufficient for CRM activity, whereas min aggregation reflects a more stringent requirement that both mechanisms should be simultaneously present. Although the mean aggregation might not provide a biologically meaningful way, mean aggregation as a standard ensemble technique was also considered. Using average aggregation of the two block probabilities, CRM-TI achieved an auROC of 89.5% (Supplementary Fig. S3), which is slightly lower than max aggregation (90.0%) but substantially higher than min aggregation. The result suggested that while mean aggregation offered a balanced integration of the two mechanisms, max aggregation remained the more biologically meaningful and best-performing strategy in terms of overall discriminative power for CRM-target gene assignment. Hence, max aggregation was retained as the default choice in CRM-TI.

Comparison with related works and baseline methods

Several tools have been developed to address the enhancer-target gene assignment problem, including PETModule [51], PEGASUS [52], CENTRE [53], the ABC model [54], and PreSTIGE [55]. To assign the target genes of enhancers, PETModule integrates conservation, distance, epigenetic correlation, and functional features using a Random Forest model. PEGASUS assigns CRM targets based on conservation of linkage, while CENTRE employs XGBoost with histone modification and expression data for CRM target gene assignment. The ABC model combines DNase, Hi-C, and H3K27ac features within a defined scoring scheme for enhancer-gene relations, and PreSTIGE predicts enhancer-gene associations from normalized H3K4me1 signals and RNA-seq data. Although these methods have had significant impacts on human regulatory genomics, they were not included in our comparison for two main reasons. First, all were designed specifically for human or mouse data and did not provide retraining codes to extend their applications to the demonstration organism Drosophila. Second, unlike our manually curated CRM-target gene pairs, the ground truths used in these models were largely inference-based (e.g., predicted CRMs or inferred enhancer-gene regulatory relationships from interaction data) rather than experimentally validated, which reduced their reliability. For these reasons, we restricted our comparisons to tools and baseline methods that could be retrained directly using the collected literature-curated CRM-target gene pairs and excluded tools that were not suitable.

Other available approaches for CRM-target gene assignment in Drosophila include EnTDefs and Enhancer Atlas. Functionally, CRM-TI unifies and extends the central ideas of these tools by integrating multiple regulatory mechanisms within a single deep-learning framework. To clarify the comparison, we first reviewed these tools and highlighted the complementary advances introduced by CRM-TI. EnTDefs [6] links CRMs to their target genes using correlations between CRMs and genes computed from four datasets: ChIA-PET data, DNase signals, gene expression profiles, and loop boundaries defined by CTCF motifs. CRM-TI complements this approach by directly modeling the chromatin interactions through epigenetic marks. Enhancer Atlas [56] incorporates the earlier tool EAGLE [57], which predicts enhancer-target gene pairs based on gene expression, genomic distances, and enhancer-associated signals. CRM-TI builds on this idea to further incorporate the concept of CRM ncRNA-transcription-related gene regulation, providing functional evidence of CRM activity that static genomic features alone cannot capture. TargetFinder [13] is another popular tool that links enhancers to their target genes by predicting enhancer-promoter interactions using epigenetic marks. However, TargetFinder has been shown to suffer from model overfitting [17] and is not directly applicable to Drosophila. Since TargetFinder is re-trainable, we retrained the TargetFinder algorithm using our curated fly CRM-gene interaction dataset. CRM-TI complements TargetFinder by mitigating overfitting through a chromosome-wise training scheme while also integrating additional regulatory modalities beyond enhancer-promoter interactions. Although several other enhancer-promoter interaction prediction tools exist, they were originally trained on the TargetFinder dataset and thus inherited its overfitting issues. The performance values of these tools remain unclear, and they were therefore excluded from our comparison. For benchmarking, we further implemented six baseline methods: (1) Pearson correlation coefficients of DNase (nucleosome depletion) signals between CRMs and genes; (2) Pearson correlations of histone mark signals for given CRMs and genes; (3) Pearson correlations of chromatin-binding site signals for given CRMs and genes; (4) genomic distances between CRMs and genes; (5) the nearest-gene assignment method; and (6) CRM-gene linking by thresholding the number of confident Hi-C contacts. For the correlation-based baselines, ROC curves were generated by varying the assignment correlation thresholds. For the distance baseline method, curves were obtained by computing confusion matrices under different distance cutoffs. For the Hi-C baseline method, the ROC curve was generated by varying the thresholds of acceptable confident Hi-C pair counts for CRM-gene pairs.

We compared the performance of CRM-TI with the retrained TargetFinder model, EnTDefs, Enhancer Atlas, and the six baseline methods on the non-overlapping CRM-target gene pair test set. The results of these comparisons are summarized in Fig. 7. As shown in Fig. 7, CRM-TI (auROC = 90.0%) outperformed all existing tools and baseline methods by at least 10% in auROC, with the second-best tool, EnTDefs, achieving 80.0%. Previous studies have reported that genomic distances between CRMs and target genes may serve as a predictive feature for the regulatory relationship [58]. In this comparison, distance-based assignment achieved an auROC of 70.0%, indicating that distance alone provides moderate predictive power. However, this performance was substantially lower than that of CRM-TI (auROC = 90.0%) and simple histone signal correlations (auROC = 74.0%), indicating that distance may be less informative than CRM-related epigenetic marks. Note that the ground-truth dataset was designed to be bias-free with respect to genomic distances, ensuring proper model training and fair tool comparison. PETModule represents another important line of work that utilizes ChIP-seq features to identify CRM-target genes through enhancer-promoter interactions. In PETModule, epigenetic signal (DNase I hypersensitivity signals, DHS) correlations were implemented as a key feature [51]. Following its design, the concept of ChIP-seq signal baseline methods using histone modification ChIP-seq signals, chromatin-binding protein ChIP-seq signals, and DNase-seq signals, which together cover the primary epigenetic inputs leveraged by PETModule-like approaches. Our results showed that models relying solely on ChIP-seq-derived features underperformed, achieving auROC values at least 16% (histone-ChIP signal baseline: auROC = 74.0%) lower than CRM-TI. This observation aligns with our central conclusion that single-modality epigenetic features are still insufficient to resolve CRM-target gene relationships, particularly for distal and non-interacting pairs. The key advance of CRM-TI lies in its mechanism-aware pipeline design, which explicitly integrates complementary regulatory evidence from chromatin loop-mediated interactions and CRM ncRNA-transcription-associated gene regulation. Finally, as expected, all baselines and tools relying solely on CRM-gene interactions failed to produce reliable CRM-target gene assignments, underscoring the need to incorporate additional regulatory mechanisms. In summary, integrating CRM-gene interactions and CRM ncRNA-transcription-related gene regulation substantially improves CRM-target gene assignment, explaining the superior performance of CRM-TI over existing tools and baseline methods.

Fig. 7.

The ROC performance comparison of CRM-TI with related tools and baseline methods on the non-overlapping CRM-target gene pair test set. The test set ROC curves show that CRM-TI (blue line, auROC = 90.0%) outperforms all baseline methods and existing tools for CRM-target gene assignment. Correlation- and distance-based baselines, including considering DNase, chromatin ChIP, histone ChIP, distance thresholds, and Hi-C evidence counts, performed poorly, as did the nearest-gene assignment. Among established tools, EnTDefs reached an auROC of 80.0%, around 10% lower than CRM-TI. These results highlight CRM-TI’s superior accuracy, which is achieved through the integration of CRM-gene interactions and CRM ncRNA-transcription-related gene regulation

Functional analysis of genes regulated by different CRM regulatory mechanisms

To further investigate the biological implications of the data partitioning analysis, we compared the 261 non-redundant genes from the confident interacting CRM-target gene pairs with the 230 non-redundant genes from the confident non-interacting CRM-target gene pairs. Gene ontology (GO) term enrichment analysis [59–61] (see Fig. S4-a and -b in the Supplementary File), performed using g:Profiler [62] with a significance threshold of 0.00001 and FDR correction, revealed that both gene sets share enrichment in DNA-binding transcription factor activity (GO:0003700) and cellular anatomical structure (GO:0110165). Notably, however, the two groups still displayed divergent functional signatures. Genes targeted by ncRNA-facilitated CRMs were distinguished by enrichment in developmental processes and morphogenetic pathways (GO:0048856, KEGG:04391). In contrast, CRM targets linked through chromatin interactions were enriched in categories associated with neurobiological receptor activity and metabolic functions (GO:0038023, GO:0004993, GO:0042417, KEGG:04080). These findings suggest that ncRNA-linked CRMs emphasize developmental and morphogenetic control, whereas interaction-linked CRMs play a broader role in neuronal signaling and organismal physiology.

In addition, expression analysis using embryo RNA-seq data from the modENCODE repository [30] (see Table S5 for the list of experiment SRA IDs) indicated that genes regulated through CRM interactions exhibit expression levels comparable to those of the genes regulated by ncRNAs transcribed from CRMs (see Fig. S4-c in the Supplementary File). The RNA-seq data were processed using the default parameters of Bowtie [63], Cufflinks, and Cuffdiff [64]. This result suggested that expression levels alone do not clearly distinguish between the two regulatory mechanisms considered in CRM-TI. Gene expression levels can be influenced by multiple regulatory inputs and therefore do not necessarily reflect the specific contribution of individual CRMs. Therefore, assigning CRM-gene relationships based on gene coexpression could still introduce substantial uncertainty. Notably, we also observed that no individual CRM regulates some genes via direct CRM-gene interaction while simultaneously regulating others through the ncRNA-transcription-related regulation mechanism. Together, these analyses reinforce the view that the two predictive blocks in CRM-TI capture complementary but distinct regulatory modalities, each contributing unique insights into CRM-target gene assignments.

Case study

We further conducted a case study in Drosophila melanogaster to demonstrate both the biological relevance of CRM-TI predictions and the interpretability of its outputs. The hairy gene in Drosophila is a classic pair-rule gene essential for embryonic segmentation and peripheral neurogenesis. Its regulation by CRMs has been extensively investigated. Specifically, stripe-specific enhancers were experimentally characterized by Hays et al. [65]. These well-validated enhancers provide ideal test cases for evaluating CRM-TI on a biologically meaningful locus. In particular, we investigated the enhancer located at chr3L:8711128-8720910 to examine its predicted regulatory program.

When CRM-TI was applied to this selected enhancer-gene (hairy) pair, the two mechanism identification blocks yielded distinct predictions (See Fig. S5 in the Supplementary File for the screen-shots of these CRM-TI prediction results). The CRM-gene interaction recognition block (CRM-TI Block 1) assigned a relatively modest probability of approximately 0.251, indicating weak evidence for regulation through chromatin looping. In contrast, the CRM ncRNA-transcription-related gene regulation identification block (CRM-TI Block 2) produced a much higher probability of approximately 0.822, suggesting strong CRM ncRNA-transcription-associated regulatory potential. Using CRM-TI’s maximum aggregation scheme, this enhancer-gene pair was confidently predicted as a true regulatory link, with Block 2 providing the dominant mechanistic evidence. The CRM targeting prediction is consistent with the experimental validation of the hairy-stripe enhancer reported by Hays et al. [65]. Independent transcriptional evidence further supports the inferred mechanism: publicly available Drosophila embryo cap analysis gene expression and deep sequencing (CAGE-seq) datasets from modENCODE [30], processed with Bowtie [63], show clear transcription initiation signals overlapping the hairy-stripe enhancer. These CAGE signals are indicative of eRNA transcription, reinforcing the inference of CRM-TI that transcriptional activity at the enhancer, rather than stable chromatin looping alone, is the major driver of hairy regulation. Together, this case study illustrated the biological applicability of CRM-TI in assigning CRM-target gene pairs while also providing insight into the underlying regulatory mechanisms.

Ultimately, CRM-TI can serve as a discovery tool to generate testable hypotheses of CRM-target gene pairs with potential regulatory mechanism inferences. Previously, we developed several CRM-related computational tools: regCNN [33] for predicting CRM locations, CFA [34] for inferring CRM functions, and DMLS [27] for systematically screening CRM-related literature to curate transcription factor-gene pairs verified to be linked via CRMs. In this work, CRM-TI extends the framework by predicting the target genes of CRMs. Together, these computational tools can be integrated with evidence from chromatin interactions (supported by Hi-C data) and transcription on CRMs (supported by CAGE data) to build a comprehensive CRM repository in which genome-wide CRMs are predicted, their functions inferred, their targets assigned, and their regulatory evidence systematically documented. This database, currently under development, will enable researchers to explore CRM predictions across multiple levels of experimental support and inference and prioritize candidates for validation experiments.

Biological limitations with CRM-TI

CRM-TI is a CRM-target gene assignment pipeline that incorporates comprehensive CRM-gene regulatory mechanisms. By integrating chromatin interactions and the regulatory effects of CRM-transcribed ncRNAs, it achieves the state-of-the-art performance of auROC = 90.0% in CRM-target gene assignment. The potential for further improvement indicates that additional factors likely contribute to distal CRM regulatory mechanisms. Because CRM-TI is a flexible framework, it can be extended to incorporate new biological aspects of CRM regulation. As novel insights into modular transcriptional regulation emerge, these advances can be integrated into CRM-TI to further enhance its predictive power. However, certain limitations remain in CRM-TI.

CRM-TI focuses on only non-overlapping CRM-gene pairs

At present, CRM-TI focuses only on non-overlapping CRM-gene pairs and does not provide specific predictions for overlapping pairs. Previous studies have shown that genes can also be regulated by exonic CRMs [66] and promoter CRMs [2], suggesting that genes with their promoter regions sharing sequence overlap with CRMs are likely to be their targets. We further evaluated this scenario in our dataset. Among the 10,551 curated CRMs with known targets, 10,988 overlapping CRM-gene pairs can be formed. Of these, approximately 69.6% have been experimentally verified as true CRM-target gene pairs, supporting the view that CRMs overlapping gene bodies or their core promoter regions are highly likely to regulate these genes. The remaining 30.4% of overlapping pairs may also represent true regulatory relationships, but further experimental validation will be necessary to confirm their functional relevance.

Potential cross-species and cross-cell-type supports of CRM-TI

Our decision to focus on Drosophila in this proof-of-concept study was guided by the availability of comprehensive, literature-curated CRM annotations in this model system. This rich dataset enabled us to rigorously develop and benchmark CRM-TI. A natural follow-up question is whether CRM-TI can be applied to mammals. The central concepts of CRM-TI rest on fundamental mechanisms of CRM-target gene regulation, such as enhancer-promoter looping and transcriptional activation through enhancer RNAs. These mechanisms are broadly conserved across metazoans [67]. In particular, the histone codes that distinguish active and repressive states, including marks such as H3K27ac, H3K4me1, and H3K27me3 [68, 69], are remarkably consistent between flies and mammals, providing a shared biochemical language that CRM-TI can capture. These conservation mechanisms suggest that the features learned from Drosophila data could be biologically transferable to other systems. Building on this biological foundation, the deep-learning blocks in CRM-TI can potentially be adapted to other species through transfer learning. By collecting the necessary datasets in mammals, such as curated CRM annotations, high-resolution Hi-C data, and ChIP-seq data for histone marks and chromatin-binding proteins, the current CRM-TI blocks could serve as the pretrained weights for fine-tuning in mammalian systems. Extending CRM-TI to mammalian systems will therefore be an important direction for future work once these datasets become available.

Currently, several practical limitations still prevent a fair and reliable cross-species comparison for CRM-TI. First, currently available Hi-C datasets in human and mouse lack sufficient read depth to confidently identify interacting CRM-gene relationships at the resolution required by CRM-TI. Second, the epigenetic datasets required by CRM-TI, including the full ChIP-seq datasets of histone modification, chromatin-binding proteins, and chromatin accessibility experiments, are not yet consistently available across species, particularly in humans. Third, the availability of high-confidence CRM-target gene ground-truth data in human or mouse is currently limited. Many existing enhancer-target gene datasets in these organisms were derived from computational inference based on epigenetic correlations or proximity rules, which may introduce substantial false positives and therefore are not suitable as reliable ground-truth benchmarks. Because the CRM-TI framework was designed to be evaluated using literature-curated and experimentally validated CRM-target gene relationships, we focused this study on Drosophila melanogaster, where such curated datasets are currently more complete.

Cell-type or condition-specific gene expression is another critical dimension of transcriptional regulation. In its current version, CRM-TI was trained on a ground-truth set of CRM-target gene pairs curated from mixed experimental conditions, enabling it to capture general patterns of CRM-gene regulatory programs rather than signals confined to specific cellular contexts. Owing to its modular design, CRM-TI could be fine-tuned to support cell-type-specific prediction if high-resolution chromatin interaction data and a moderate number of validated cell-type-specific CRM-target gene pairs become available. Such datasets would provide the fine-grained transcriptional information needed to model regulatory signals unique to particular cell types or conditions. Alternatively, CRM-TI’s general predictions could be refined post hoc using functional perturbation data to introduce condition-specificity. For example, cell-type-specific knockout or CRISPR perturbations of CRMs, followed by RNA-seq, could be used to map context-dependent targets. These strategies also highlight how future versions of CRM-TI may address cell-type-specific transcriptional regulation once more comprehensive datasets are available.

The effects of the resolution of chromatin interaction datasets used by CRM-TI

In this work, we selected the Hi-C dataset generated by Pal et al. [20], which profiled chromatin interactions in male and female Drosophila embryos, because it provides the most consistent and high-resolution chromatin interaction maps currently available. Nonetheless, relying exclusively on embryonic Hi-C data can introduce noise, as chromatin contacts can differ across developmental stages and tissues. Although Hi-C datasets from larval CNS cells and whole adult flies are available [70], naively merging them with embryonic data into a single dataset would be problematic. Stage-specific chromatin contacts could be still diluted, thereby introducing substantial label noise. Moreover, the number of confident non-interacting CRM-target gene pairs, along with the corresponding distance-matched non-target subsets, would become too small to support robust deep network training since more noise is also incorporated by dataset merging. In light of these challenges, the most appropriate strategy is to develop stage-specific CRM-TI models. Such models could leverage stage-matched Hi-C and epigenomic data to capture developmental and tissue-specific regulatory mechanisms without incurring excessive noise associated with cross-stage integration.

Epigenetic signal noise needs to be controlled when further adapting CRM-TI to other biological systems

When extending CRM-TI toward cell type-specific applications and mammalian systems, the impact of noise in ChIP-seq datasets must be carefully considered. In CRM-TI, ChIP-seq signals provide the genomic information used by the two deep-learning blocks to guide CRM-target gene assignment. However, these signals are susceptible to batch effects, sequencing depth, and antibody variability, all of which collectively introduce noise. Such noise can typically be mitigated through the use of appropriate control libraries. However, the convolutional layers in the deep networks still rely on ChIP-seq peaks with sufficient signal-to-noise ratios. Datasets with very weak signals may therefore provide information too limited to be effectively utilized. Although CRM-TI’s architectures, which include several layers of convolution filters, are designed to be robust against background fluctuations, its predictive accuracy ultimately depends on the quality and reliability of the input epigenetic data. Thus, epigenetic noise remains an important factor to account for when adapting CRM-TI to new biological systems.

Note that ATAC-seq (Assay for Transposase-Accessible Chromatin with sequencing) is another commonly used method to probe chromatin accessibility by identifying nucleosome-depleted regions that are likely to correspond to regulatory elements [71]. In the current CRM-TI framework, this type of information is already incorporated through DNase-seq datasets, which also measure chromatin accessibility by detecting regions of DNase I hypersensitivity. Both ATAC-seq and DNase-seq rely on the same biological principle of identifying open chromatin regions and therefore provide highly related signals regarding regulatory activity. In CRM-TI, DNase-seq signals from seven experiments were specifically selected to be included as part of the epigenetic input channels used by the model. Because these datasets already capture chromatin accessibility information, incorporating both ATAC-seq and DNase-seq data together would likely provide redundant signals with similar biological interpretation. However, ATAC-seq results can serve as an alternative to the DNas-seq data for capturing chromatin accessibility information.

CRM-TI can help analyze the potential alternative TSSs in cells

The transcription start site (TSS) information used in this study was obtained directly from the FlyBase genome annotation [24] (release dm6.48), which provides curated coordinates for gene models and their associated transcription start sites. During dataset construction, promoter regions were defined relative to the annotated TSS positions provided in the FlyBase annotation. Specifically, the core promoter and proximal promoter regions were determined based on the coordinates of the corresponding TSSs, allowing CRM-gene relationships to be evaluated with respect to the correct transcription initiation sites. Recently, it has been shown that alternative TSSs can generate multiple transcript isoforms and distinct promoter contexts [72]. The CRM-TI framework can naturally accommodate these cases by using the annotated TSS coordinates as the anchor points for promoter region extraction. In practice, when a gene contains multiple TSSs, each TSS can be treated separately to define distinct proximal and core promoter regions, and CRM-gene relationships can be evaluated with respect to these alternative promoter regions. This approach ensures that regulatory relationships associated with different transcription initiation sites are not overlooked.

CRM-TI is suitable for identifying target genes of CRMs within the same TADs

Topologically associating domains (TADs) are widely recognized as structural units of chromatin organization that tend to confine regulatory communication within their boundaries. In the datasets constructed for this study, most CRM-target gene pairs are separated by genomic distances within 25 kb (Supplementary Fig. S1-c). Previous studies have shown that Drosophila TAD sizes are typically on the order of 120–140 kb [73]. Given that the curated CRM-target gene distances are substantially smaller than typical TAD spans in Drosophila, the regulatory mechanisms analyzed here are expected to occur predominantly within the same TAD rather than across TAD boundaries. This observation is consistent with previous studies showing that regulatory contacts are generally constrained by TAD architecture. While explicit TAD annotations were not incorporated as input features in the current CRM-TI model, the pairs used for training and evaluation were restricted to genomic distances that would typically fall within individual TADs. Therefore, CRM-TI is primarily suitable for identifying CRM regulatory programs within individual TADs. CRM-target gene pairs that may span across TAD boundaries remain to be further investigated in future work.

CRMs within a TAD can form higher-order modules that cooperatively regulate specific sets of genes. Superenhancers are one of the most well-studied examples of such modules. They are generally defined as clusters of enhancers with exceptionally high densities of transcription factors, mediator complexes, and active histone marks that regulate a set of target genes [74]. These regions are typically characterized by strong enhancer-enhancer interactions and coordinated regulation of their target genes. In this study, the CRM-TI framework focuses on inferring regulatory relationships between CRMs and genes, which can further serve as a fundamental component for studying superenhancer functions. Superenhancers can be viewed as networks of interacting CRMs coupled with coordinated regulation of target genes. Therefore, by combining CRM-TI predictions with tools that characterize CRM-CRM interaction networks, it would be possible to investigate the regulatory architecture of CRM clusters and their cooperative regulation of target genes. Systematic investigation of superenhancer structures and their regulatory interactions represents an important potential biological application of CRM-TI.

Tool usage of CRM-TI

CRM-TI requires predefined CRM regions as inputs. Therefore, CRMs must first be identified before performing CRM-gene interaction prediction. In practice, candidate CRMs can be obtained using existing genome-wide CRM discovery tools such as regCNN [33], which was previously developed for identifying regulatory regions in the Drosophila genome. Once CRM regions of interest are available, CRM-TI can be applied to evaluate their potential regulatory relationships with given genes. To facilitate large-scale analyses and reduce the need for manual specification of individual CRM-gene pairs, besides CRM-gene pairs, the system also allows users to provide a list of CRMs together with a list of genes to obtain predictions for all possible CRM-gene pairs between the two lists. In addition, users may provide a list of CRMs alone, in which case the system automatically evaluates their potential regulatory relationships with all genes in the genome. The download link of CRM-TI can be found in the “Availability of data and materials” subsection.

Conclusions

In this study, we introduced CRM-TI, a pipeline for CRM-target gene assignment. CRM-TI integrates both CRM-gene interaction recognition and CRM ncRNA-transcription-related gene regulation identification to achieve more accurate CRM-target gene assignments. Overall, CRM-TI demonstrated superior performance compared with existing tools and baseline methods. We also showed that relying solely on CRM-gene interaction recognition produces noisy assignments, whereas combining interaction recognition with ncRNA-transcription-related gene regulation identification substantially improves CRM-target gene assignment accuracy. This finding highlighted the cooperative roles of CRM-transcribed ncRNAs and chromatin loop formation. The development of CRM-TI thus provided new insights and testable predictions for research into modular transcriptional regulation in metazoan species.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

YHY and GLH analyzed the dataset and implemented the CRM-gene assignment model. YHY and THY visualized the results and were the major contributors in writing the manuscript. All authors read and approved the final manuscript.

Funding

This work was supported by National Cheng Kung University and National Science and Technology Council, Taiwan [NSTC 114-2628-E-006-007-MY3, NSTC 112-2221-E-006-129-MY2, MOST 110-2222-E-006-017, MOST 111-2221-E-006-231]. The study was also supported in part by the Higher Education Sprout Project, Ministry of Education to the Headquarters of University Advancement at National Cheng Kung University, and Miin Wu School of Computing, National Cheng Kung University, Taiwan. The authors confirmed that these institutes have no role in the data collection, analysis, and interpretation processes.

Data availability

No experimental datasets were generated or analysed during the current study. The source code of CRM-TI and the prepared CRM-target gene pairs used in this study are available at https://github.com/cobisLab/CRM-TI.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.