Pair your accounts.

Export articles to Mendeley

Get article recommendations from ACS based on references in your Mendeley library.

Pair your accounts.

Export articles to Mendeley

Get article recommendations from ACS based on references in your Mendeley library.

You’ve supercharged your research process with ACS and Mendeley!

STEP 1:
Click to create an ACS ID

Please note: If you switch to a different device, you may be asked to login again with only your ACS ID.

Please note: If you switch to a different device, you may be asked to login again with only your ACS ID.

Please note: If you switch to a different device, you may be asked to login again with only your ACS ID.

MENDELEY PAIRING EXPIRED
Your Mendeley pairing has expired. Please reconnect
ACS Publications. Most Trusted. Most Cited. Most Read
CapsEnhancer: An Effective Computational Framework for Identifying Enhancers Based on Chaos Game Representation and Capsule Network
More

OR SEARCH CITATIONS

My Activity
Publications

Figure 1Loading Img
Download Hi-Res ImageDownload to MS-PowerPointCite This:J. Chem. Inf. Model. 2024, 64, 14, 5725-5736
  • Open Access
Bioinformatics

CapsEnhancer: An Effective Computational Framework for Identifying Enhancers Based on Chaos Game Representation and Capsule Network
Click to copy article linkArticle link copied!

  • Lantian Yao
    Lantian Yao
    Kobilka Institute of Innovative Drug Discovery, School of Medicine, The Chinese University of Hong Kong, Shenzhen 518172, China
    School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen 518172, China
    More by Lantian Yao
    Orcidhttps://orcid.org/0000-0003-4554-6827
  • Peilin Xie
    Peilin Xie
    Kobilka Institute of Innovative Drug Discovery, School of Medicine, The Chinese University of Hong Kong, Shenzhen 518172, China
    More by Peilin Xie
  • Jiahui Guan
    Jiahui Guan
    School of Medicine, The Chinese University of Hong Kong, Shenzhen 518172, China
    More by Jiahui Guan
  • Chia-Ru Chung
    Chia-Ru Chung
    Department of Computer Science and Information Engineering, National Central University, Taoyuan 320317, Taiwan
    More by Chia-Ru Chung
  • Yixian Huang
    Yixian Huang
    School of Medicine, The Chinese University of Hong Kong, Shenzhen 518172, China
    More by Yixian Huang
  • Yuxuan Pang
    Yuxuan Pang
    Division of Health Medical Intelligence, Human Genome Center, The Institute of Medical Science, The University of Tokyo, Tokyo 108-8639, Japan
    More by Yuxuan Pang
  • Huacong Wu
    Huacong Wu
    School of Medicine, The Chinese University of Hong Kong, Shenzhen 518172, China
    More by Huacong Wu
  • Ying-Chih Chiang*
    Ying-Chih Chiang
    Kobilka Institute of Innovative Drug Discovery, School of Medicine, The Chinese University of Hong Kong, Shenzhen 518172, China
    School of Medicine, The Chinese University of Hong Kong, Shenzhen 518172, China
    *Email: [email protected]
    More by Ying-Chih Chiang
  • Tzong-Yi Lee*
    Tzong-Yi Lee
    Institute of Bioinformatics and Systems Biology, National Yang Ming Chiao Tung University, Hsinchu 300093, Taiwan
    Center for Intelligent Drug Systems and Smart Bio-devices (IDS2B), National Yang Ming Chiao Tung University, Hsinchu 300093, Taiwan
    *Email: [email protected]
    More by Tzong-Yi Lee
Open PDFSupporting Information (1)

Journal of Chemical Information and Modeling

Cite this: J. Chem. Inf. Model. 2024, 64, 14, 5725–5736
Click to copy citationCitation copied!
https://doi.org/10.1021/acs.jcim.4c00546
Published June 30, 2024

Copyright © 2024 The Authors. Published by American Chemical Society. This publication is licensed under

CC-BY 4.0 .

Abstract

Click to copy section linkSection link copied!
High Resolution Image
Download MS PowerPoint Slide

Enhancers are a class of noncoding DNA, serving as crucial regulatory elements in governing gene expression by binding to transcription factors. The identification of enhancers holds paramount importance in the field of biology. However, traditional experimental methods for enhancer identification demand substantial human and material resources. Consequently, there is a growing interest in employing computational methods for enhancer prediction. In this study, we propose a two-stage framework based on deep learning, termed CapsEnhancer, for the identification of enhancers and their strengths. CapsEnhancer utilizes chaos game representation to encode DNA sequences into unique images and employs a capsule network to extract local and global features from sequence “images”. Experimental results demonstrate that CapsEnhancer achieves state-of-the-art performance in both stages. In the first and second stages, the accuracy surpasses the previous best methods by 8 and 3.5%, reaching accuracies of 94.5 and 95%, respectively. Notably, this study represents the pioneering application of computer vision methods to enhancer identification tasks. Our work not only contributes novel insights to enhancer identification but also provides a fresh perspective for other biological sequence analysis tasks.

This publication is licensed under

CC-BY 4.0 .
  • cc licence
  • by licence
Copyright © 2024 The Authors. Published by American Chemical Society

Subjects

what are subjects

Introduction

Click to copy section linkSection link copied!
Enhancers, short noncoding DNA sequences interspersed throughout the genome, play an indispensable role in the orchestration of gene expression and, by extension, every biological process in living organisms. (1−3) These unique genomic elements are known to amplify the transcription rate of their associated genes, acting as regulators in the vast genomic machinery. Enhancers facilitate the binding of proteins, such as transcription factors and coactivators, which modulate transcription initiation, thereby influencing many cellular activities such as differentiation, development, and responses to environmental stimuli. (4) Notably, enhancers can function from variable distances away from the genes they regulate and can be located upstream, downstream, or even within intronic regions of these genes. The flexible nature of their operation within the genomic landscape enhances their regulatory capacity, making them vital components in the architecture of life. (5−7)
The functional importance of enhancers extends beyond merely amplifying gene expression; they are critical in determining the spatiotemporal patterns of gene activity, thereby shaping the identity and function of each cell type. In essence, enhancers are at the heart of cellular diversity and organismal complexity. Their malfunctioning is associated with various genetic disorders, including cancer, highlighting their importance in maintaining cellular homeostasis. (8) Traditional enhancer identification methods, (9−11) such as ChIP-seq, although capable of identifying enhancers, also face challenges, including high costs, low throughput, and extensive starting material requirements. These challenges make them time-consuming, labor-intensive, and expensive. Moreover, the vastness and complexity of the human genome make the large-scale application of these experimental methods impractical. (12) As a result, the scientific community is increasingly utilizing computational methodologies to identify and classify enhancers. The advent of next-generation sequencing technologies has led to the development of numerous computational strategies aimed at distinguishing enhancers from other noncoding genomic regions, offering an economical and efficient alternative to traditional experimental approaches. (13)
Computational methods are generally classified into two categories: based on traditional machine learning classifiers and deep learning. Support vector machine (SVM) and random forest (RF) algorithms are frequently employed for enhancer classification. Liu et al. proposed a method called iEnhancer-2L, integrating the pseudo k-tuple nucleotide composition (PseKNC) of DNA sequences and utilizing SVM for enhancer identification. (14) Jia et al. introduced a tool named EnhancerPred, integrating biprofile bayes (BPB), nucleotide composition (NC), and PseNC, constructing a classifier using SVM. (15) Lim et al. developed a RF-based tool called iEnhancer-RF, integrating six kinds of features of DNA sequences. (16) Methods based on ensemble learning are also employed for the recognition of enhancers. In 2018, Liu et al. proposed an ensemble learning method named iEnhancer-EL, utilizing PseKNC, Kmer, and subsequence profile of sequences to predict enhancers. (17) Similarly, Wang et al. developed Enhancer-FRL, integrating ten kinds of features and employing five machine learning methods, including SVM and RF, to predict enhancers and their activities. (18) Gill et al. developed a deep forest-based tool, NEPERS, integrating four kinds of features to identify enhancers. (19)
In recent years, in addition to traditional machine learning algorithms, due to the development of deep learning, an increasing number of enhancer classifiers based on deep learning have emerged. Nguyen et al. developed iEnhancer-ECNN, utilizing one-hot encoding of sequences and convolutional neural network (CNN) to predict enhancers. (20) Le et al. introduced BERT-Enhancer, utilizing a BERT pretraining model to extract sequence encoding, followed by CNN to build a classifier. (21) Niu et al. developed a tool named iEnhancer-EBLSTM, using Kmers information from DNA sequences and employing bidirectional LSTM (BiLSTM) to construct an enhancer classifier. (22) A summary of relevant work is presented in Table 1.
Table 1. Summary of Existing Tools for Enhancer Identification
no.methodfeature encodingalgorithmyearreference
1iEnhancer-2LPseKNCSVM2016 (14)
2EnhancerPredBPB, NC, PseNCSVM2016 (15)
3iEnhancer-ELKmer, subsequence profile, PseKNCensemble learning2018 (17)
4iEnhancer-ECNNone-hot encoding, KmersCNN2019 (20)
5iEnhancer-XGk-spectrum profile, mismatch k-tuple, subsequence profile, PSSM, PseDNCbXGBoost2021 (23)
6BERT-EnhancerBERT encodingCNN2021 (21)
7iEnhancer-EBLSTMKmersBiLSTM2021 (22)
8iEnhancer-RFNBP, DBP, ANF, NCP, ENAC, XY K-GAPRF2021 (16)
9iEnhancer-RDKmers, PseKNC, KPCVDNN2021 (24)
10spEnhancerKmersBiLSTM2021 (25)
11Enhancer-FRLANF, CKSNAP, DAC, ENAC, Kmers, NCP,PseTIIP, SCPseDNC, SCPseTNC, TACCSVM, RF, KNN, naive Bayesian, LightGBM2022 (18)
12iEnhancer-BERTBERT encodingCNN2022 (26)
13iEnhancer-ELMKmer, BERT encodingMLP2023 (27)
14iEnhancer-DCSAWord2vecdual-scale CNN, spatial attention2023 (28)
15iEnhancer-SKNNKmer, PseDNC, PCPseDNC and Z-Curve9ensemble learning2023 (29)
16NEPERSPSTNPss, PSTNPdss, CKSNAP, NCPdeep forest2023 (19)
While the computational methodologies mentioned above exhibit promising outcomes, with each demonstrating distinct merits, additional investigation is warranted for the following reasons. Prior studies have demonstrated the discriminative capabilities of utilizing Kmers in distinguishing biosequences with diverse functionalities. (30−32) For example, Li et al. proposed a method called GCR-Net, which models Kmers of genomic sequences hierarchically to enhance the prediction of translation initiation sites. (33) Nevertheless, extant models, including CNNs, exhibit suboptimal performance in effectively learning the intricate associations between Kmers and their respective frequencies. An efficacious method for sequence encoding, chaos game representation (CGR), has the capability to transform biosequences into two-dimensional images. (34) By encoding Kmers frequencies as visual representations, employing computer vision methods to discern and acquire patterns inherent in CGR-encoded images is straightforward. (35) Moreover, although CNNs have achieved a series of successes in recent years, they are not without limitations. For instance, CNNs lack an understanding of the hierarchical structure of objects. Traditional CNNs struggle to capture the hierarchical structure and part-whole relationships of objects, leading to limitations in comprehending the relationship between the overall and local features of objects. (36,37) In recent years, researchers have been continuously developing novel methods to address the aforementioned limitations and applying them to various tasks. Guo et al. proposed a variational gated autoencoder-based feature extraction model to extract complex contextual features and infer disease-miRNA associations. (38) Additionally, a method called MCANet, which integrates multiscale convolution and self-attention mechanisms, adaptively reveals spatial-temporal contextual dependence to enhance Poly(A) signal prediction. (39) Wang et al. introduced a cross-feature enhancement module, which effectively reduces information redundancy and facilitates the integration and modeling of complementary features using attention mechanisms. (40)
To further address constraints of CNNs, a new generation of neural networks, known as capsule networks, has emerged. (41) Capsule networks introduce the concept of capsules to better capture the spatial hierarchical structure within objects. Each capsule represents a specific entity or part, and the relationships between capsules can be modeled. This contributes to an enhanced understanding of object hierarchical structures by the network. In recent years, capsule networks have been applied to tasks in bioinformatics, demonstrating satisfactory performance. (42−45)
In this study, we proposed a new scheme named CapsEnhancer, designed to achieve the identification of enhancers and their strength. The workflow of CapsEnhancer is shown in Figure 1. Experimental results demonstrate that CapsEnhancer achieves satisfactory performance on benchmark data sets. The main contributions of this study can be summarized as follows.
(1)

We designed a two-stage computational framework called CapsEnhancer to identify enhancers and their strengths. The first stage of CapsEnhancer focuses on enhancer recognition, distinguishing between enhancer and nonenhancer. The second stage involves predicting enhancer strength, specifically discerning between strong and weak enhancers.

(2)

CapsEnhancer uses CGR encoding to represent each DNA sequence as an image. Through this encoding method, it can effectively represent Kmers and their frequencies.

(3)

CapsEnhancer employs a capsule network-based architecture to learn local and global features from the “images” transformed from DNA sequences. CapsEnhancer represents the pioneering adoption of computer vision strategies for enhancer identification.

(4)

Experimental results demonstrate that CapsEnhancer attains state-of-the-art performance in the two-stage task. In comparison to previous methods, CapsEnhancer exhibits significant improvements, achieving an 8% increase in accuracy during the first stage and a 3.5% improvement in the second stage. Beyond providing a robust solution for enhancer identification, our framework introduces a novel perspective for other biological sequence analysis tasks.

Figure 1

Figure 1. Workflow of CapsEnhancer. First, we utilized benchmark data sets from previous studies. Subsequently, each DNA sequence was encoded using CGR encoding and represented as corresponding two-dimensional images. The model was then constructed using an architecture based on capsule networks. Hyperparameter adjustment was performed through fivefold cross-validation, and the model was evaluated using an independent test set, with the subsequent reporting of model performance metrics. The trained model was ultimately employed for enhancer identification, constituting a two-stage task. The first stage focused on discerning enhancers from nonenhancers, while the second stage aimed to predict enhancer strength, i.e., strong enhancers versus weak enhancers. The second stage employs the same FCGR images as the first stage to maintain consistency in the input representation. Capsule networks are used in both stages to build the models.

High Resolution Image
Download MS PowerPoint Slide

Materials and Methods

Click to copy section linkSection link copied!

Benchmark Data Set

In order to facilitate fair comparisons, we employed the data set constructed by Liu et al., (14,17) which has been widely used in enhancer prediction tasks. (15,16,18,20,29) The enhancers within this data set were derived from nine distinct cell lines, wherein they were isolated as DNA sequences from short 200 bp clips of uniform length. Subsequently, the CD-HIT software was employed to eliminate paired sequences exhibiting a similarity surpassing 20%.
The final data set can be represented as follows
{S=S+∪S−S+=SStrong+∪SWeak+
(1)
where the subset S+ comprises 1484 enhancer samples, and S− comprises 1484 nonenhancer samples, forming the first stage of the data set. In addition, SStrong+ consists of 742 strong enhancer samples, and SWeak+ comprises 742 weak enhancer samples, constituting the second stage of the data set. The independent test set is utilized for assessing the model’s performance, and it is sourced from the work of Liu et al., encompassing 100 strong enhancers, 100 weak enhancers, and 200 nonenhancers.
To facilitate a more comprehensive understanding of the distinctions between positive and negative samples, the GC content of data sets for two distinct stages was plotted, as presented in Figure S1. It is evident that enhancers exhibit a higher GC content compared to nonenhancers. Furthermore, strong enhancers also display a higher GC content when contrasted with weak enhancers.

Architecture Overview of CapsEnhancer

The model architecture of CapsEnhancer is illustrated in Figure 2. CapsEnhancer is a two-stage framework wherein the first stage aims to identify enhancer and nonenhancer, while the second stage focuses on distinguishing between strong enhancer and weak enhancer categories. Initially, DNA sequences are transformed into images through CGR encoding. Subsequently, a 2-dimensional convolutional neural network (Conv2D) is employed to preliminarily extract features from these images. The acquired preliminary features are then fed into a capsule network for further feature extraction and spatial modeling from the images. The main notations of this study are summarized in Table 2.

Figure 2

Figure 2. Architecture of CapsEnhancer. First, DNA sequences are encoded using CGR encoding and represented as 2D images. Subsequently, they are input into a Conv2D for preliminary feature extraction. Following this, the data is fed into a capsule network, which consists of a primary capsule layer and a type capsule layer. The primary capsule layer includes a Conv2D for further extracting local features. Then, a dynamic routing algorithm is utilized to capture the spatial relationships of features, resulting in the type capsule layer. As the task is a standard binary classification, the type capsule layer comprises two capsules, corresponding to the positive class and the negative class (stage 1: enhancer versus nonenhancer; stage 2: strong enhancer versus weak enhancer). Finally, the prediction probabilities for the two classes are obtained by calculating the lengths of the capsules in the type capsule layer.

High Resolution Image
Download MS PowerPoint Slide
Table 2. Main Notations and Descriptions
notationdescription
Nthe size of the FCGR images
mdimension of each primary capsule
ndimension of each type capsule
uiprimary capsules
Vjtype capsules
Wweight matrix in capsule networks
ci,jcoupling coefficients
pprediction probability

CGR Encoding

CGR is a mathematical method that employs iterated function systems to convert sequential data into a fractal depiction within a two-dimensional space. CGR is a milestone in graphical bioinformatics and is considered a powerful tool for feature encoding in biological sequences, including DNA, RNA, and protein sequences. (35,46)
We employed CGR encoding to encode DNA sequences in this study. Initially, allocate the four nucleotides (A, C, G, T) to the four vertices of a square. Figure 3A provides an illustrative example of encoding a sequence using CGR representation. For a DNA sequence s of length n, where s = s1, ..., si, ..., sn and si ∈ {A, C, G, T}, the coordinates of the new nucleotide si in the sequence are determined by the current amino acid type and the coordinates of the preceding nucleotide si–1. The position of si is located halfway along the line connecting the current position and the vertex associated with the nucleotide. CGR encoding of sequence s is a two-dimensional representation of ordered pairs (x1, y1) through (xi, yi) to (xn, yn), where (xi, yi) is defined as follows
(xi,yi)=12((xi−1,yi−1)+h(si))
(2)
where (x0, y0) = (0, 0) and
h(si)={(−1,1),si=A(−1,−1),si=C(1,−1),si=G(1,1),si=T
(3)

Figure 3

Figure 3. (A) Applying CGR encoding to an example sequence: CATG. (B) Dividing the CGR space during the iterative process.

High Resolution Image
Download MS PowerPoint Slide
Figure 3B illustrates the partitioning of the CGR space during the iterative process. Each subsquare within the CGR space holds distinctive significance. Upon dividing the CGR into four quadrants, the upper right corner encompasses points that symbolize subsequences terminating with the nucleotide T. This is attributed to the fact that the midpoint between any other point within the square and the corner T invariably resides within this quadrant. Upon subdividing this quadrant into four squares in a clockwise order, they, respectively, denote subsequences concluding with TT, GT, CT, and AT. This configuration facilitates the computation of 2-mer counts by tallying the points within these designated subsquares.
In contrast to the precise coordinate representation employed by the original CGR, a discretization method known as the frequency chaos game representation (FCGR) has been introduced to provide a coarser and less susceptible-to-noise abstraction for sequences. FCGR, an extension of CGR, involves a grid-based counting approach for determining the points within the CGR. The initial step of FCGR involves partitioning the CGR image into N × N regions. Subsequently, the point count within each region serves as the region’s frequency, enabling the compression of the CGR and resulting in an FCGR matrix with dimensions N × N applicable to input sequences of varying lengths. Therefore, the predefined grid values can serve as a representation of the frequency of Kmers. In this study, we opt for N = 64 as the parameter for generating R64×64 images corresponding to each DNA sequence.
Figure S2A,B, respectively, depict the FCGR images of an enhancer sequence and a nonenhancer sequence. Their FCGR images exhibit markedly distinct patterns, including highlighted regions in red and blue, which could potentially serve as discriminative signals for distinguishing between enhancer and nonenhancer sequences. These discriminative regions can be spatially modeled by the unique architecture of capsule networks, allowing the learning of their correlations to enhance classification accuracy.

Capsule Network

Over recent years, CNNs have surpassed numerous conventional models reliant on curated feature extraction, making substantial progress in various domains, including computer vision and bioinformatics. However, CNNs are constrained by their incapacity to comprehend spatial relationships between features and the loss of invariance due to pooling operations. Sabour et al. introduced a novel deep learning paradigm known as capsule network (CapsNet) to circumvent these limitations. (41)
Within the primary capsule layer resides a Conv2D layer, employed for further feature extraction. The outputs from this Conv2D layer are transformed into multiple m-dimensional vectors (the dimensionality m being a hyperparameter). These m-dimensional vectors undergo a nonlinear “squash” function that retains the direction of the vector while constraining its magnitude to a range between 0 and 1.
squash(s)=∥s∥21+∥s∥2s∥s∥
(4)
For the binary classification task, the type capsule layer encompasses two n-dimensional capsules: one positive capsule and one negative capsule. The length of each capsule in type capsule layer represents the probability of being predicted as a positive (or negative) sample. Figure 4 illustrates the computational process between the primary and type capsule layers.

Figure 4

Figure 4. Computational process between primary capsules and type capsules.

High Resolution Image
Download MS PowerPoint Slide
To derive the prediction vectors from capsule i to j, the outputs of the primary capsule layer ui are initially multiplied by a learnable weight matrix Wi,j. Subsequently, Sj is determined as the weighted sum of all computed u^j|i.
u^j|i=Wi,jui
(5)
Sj=∑i=1Lci,ju^j|i
(6)
where i and j, respectively, denote two capsules originating from the primary capsule layer and the type capsule layer, and L is the number of primary capsules. Here ci,j represents coupling coefficients, determined by the dynamic routing algorithm (see Algorithm S1 in Supporting Information), indicating the degree of coupling between the primary capsule i and the type capsule j. Sj is fed into the Squash function to produce an output vector Vj with a length between 0 and 1.
Based on the theory of capsule networks, the vector Vj is utilized to model positive and negative samples, specifically (enhancer versus nonenhancer or strong enhancer versus weak enhancer) in this task. Each element of Vj represents a feature of positive or negative samples, and the length of Vj signifies the probability of being predicted as a positive or negative sample. Hence, to derive the predicted probabilities, it is necessary, at the network’s terminus, to compute the length of Vj, as delineated by the following formula.
pj=∥Vj∥2
(7)
where pj, respectively, denote the model’s predictions for being in the positive class or negative class.

Performance Assessment

In this study, we used accuracy, sensitivity, specificity, and the Matthews correlation coefficient (MCC) as evaluation metrics for the two-stage task. Their definitions are as follows.
Accuracy=TP+TNTP+TN+FP+FN
(8)
Sensitivity=TPTP+FN
(9)
Specificity=TNTN+FP
(10)
MCC=TP×TN−FP×FN(TP+FP)(TP+FN)(TN+FP)(TN+FN)
(11)
where TP, TN, FP and FN represent the number of true positives, true negatives, false positives and false negatives, respectively.
The CGR encoding was implemented using the R package “Kaos”. (47) The model was trained for 100 epochs to ensure adequate fitting, employing the Adam optimizer (48) with an initial learning rate set to 0.1. Hyperparameter tuning was conducted using grid search and cross-validation techniques, with the specific search space outlined in Table S1. We iterate over all possible combinations of specified hyperparameter values and evaluate each combination using fivefold cross-validation to identify the best-performing hyperparameter combination. The pipeline for CapsEnhancer was established using PyTorch, (49) and the training process utilized 4 × Nvidia 2080 Ti GPUs.

Results and Discussion

Click to copy section linkSection link copied!

Performance Comparison with Existing Methods

First Stage: Enhancer Versus Nonenhancer

The capsule network-based with CGR encoding learning scheme was employed to adapt the task of standard binary classification between enhancer and nonenhancer sequences. We conducted a fair comparison with 13 currently existing tools, reporting performance on an independent test set as shown in Table 3. From Table 3, it is evident that CapsEnhancer achieves state-of-the-art performance compared to existing methods, exhibiting a substantial improvement in terms of accuracy, sensitivity, specificity, MCC, and AUC. In comparison to the second-ranked NEPERS method, CapsEnhancer exhibits an 8% improvement in ACC, reaching an accuracy of 94.5%, indicative of its precision in enhancer prediction. In terms of MCC, CapsEnhancer outperforms the second-ranked method by 0.16, reaching a value of 0.89. Furthermore, CapsEnhancer excels in sensitivity and specificity, surpassing the second-ranked method by 6 and 10%, reaching 93 and 96%, respectively. Notably, both sensitivity and specificity for CapsEnhancer exceed 90%, indicating its ability to provide more balanced predictions.
Table 3. Performance Comparison with Other Existing Methods on the Independent Test Set of the First Stage: Enhancer Versus Non-Enhancer
methodaccuracy (%)sensitivity (%)specificity (%)MCCAUC (%)
iEnhancer-2L73.071.075.00.46080.6
EnhancerPred74.073.574.50.48080.1
iEnhancer-EL74.871.078.50.49681.7
iEnhancer-ECNN76.978.575.20.53783.2
iEnhancer-XG75.874.077.50.515 
Enhancer-FRL78.080.575.50.56185.7
BERT-Enhancer75.680.071.20.514 
iEnhancer-EBLSTM77.275.579.50.53483.5
iEnhancer-RF79.878.581.00.59586.0
iEnhancer-RD78.881.076.50.57684.4
spEnhancer77.383.071.50.57982.4
iEnhancer-DCSA82.579.585.50.65185.6
NEPERS86.386.586.00.72594.8
CapsEnhancer (ours)94.593.096.00.89098.0
Moreover, a significant challenge for two-stage tasks is how to handle false positive samples from the first stage. These false positives will also undergo the prediction task in the second stage, thereby affecting the robustness of the model. As shown in Table 3, it is evident that CapsEnhancer has a very low false positive rate (1─specificity) of only 4%, which is more than a 10% reduction compared to other existing methods. This demonstrates that CapsEnhancer is more robust than other methods and better at avoiding false positives.

Second Stage: Strong Enhancer Versus Weak Enhancer

The second stage of the CapsEnhancer involves the task of predicting enhancer strength, specifically distinguishing between strong enhancers and weak enhancers. We compared the performance of the second stage with existing methods, and the results are presented in Table 4. CapsEnhancer continues to exhibit impeccable predictive performance in the second stage, showcasing a significant lead in metrics such as accuracy, sensitivity, specificity, MCC, and AUC. In terms of accuracy, CapsEnhancer outperforms the second-ranked iEnhancer-DCSA by 3.5%, achieving a remarkable accuracy of 95%. Furthermore, in terms of MCC, it surpasses the second position by 0.06, reaching a value of 0.903. CapsEnhancer demonstrates satisfactory performance in sensitivity and specificity, achieving 99 and 91%, respectively.
Table 4. Performance Comparison with Other Existing Methods on the Independent Test Set of the Second Stage: Strong Enhancer Versus Weak Enhancer
methodaccuracy (%)sensitivity (%)specificity (%)MCCAUC (%)
iEnhancer-2L60.547.074.00.21866.8
EnhancerPred55.045.065.00.10257.9
iEnhancer-EL61.054.068.00.22268.0
iEnhancer-ECNN67.879.156.40.36874.8
iEnhancer-XG63.570.057.00.272 
Enhancer-FRL73.598.049.00.53987.2
BERT-Enhancer     
iEnhancer-EBLSTM65.881.253.60.32468.8
iEnhancer-RF85.093.077.00.70997.0
iEnhancer-RD70.584.057.00.42679.2
spEnhancer62.091.033.00.37062.5
iEnhancer-DCSA91.598.085.00.83796.6
NEPERS89.094.084.00.78495.1
CapsEnhancer (ours)95.099.091.00.90399.2
The results above demonstrate that CapsEnhancer has achieved outstanding performance in both stages of the task, which can be attributed to several factors. First, the use of CGR encoding serves as an efficient method for converting DNA sequences into two-dimensional images, enabling the application of computer vision techniques to sequence-related problems. Importantly, CGR encoding excels in capturing the frequency of Kmers. (35) Prior literature has emphasized the significance of Kmers frequency as a critical feature in DNA sequence analysis. (46,50,51)
Second, owing to the architecture of the capsule network, the introduction of the capsule concept allows for effective spatial modeling of input images. Capsule networks overcome traditional CNN limitations, such as the inability to comprehend spatial relationships between features and the loss of invariance due to pooling operations. Consequently, in this context, capsule networks successfully learn the relationships between Kmers. Furthermore, owing to the aforementioned advantages of CapsEnhancer, we conducted a case study to illustrate its efficacy in managing sequencing errors and its capability to extend effectively to sequences of nonuniform lengths. Detailed information can be found in the case study section in Supporting Information. In conclusion, the synergistic combination of CGR encoding and capsule networks constitutes a pivotal factor in improving the performance of enhancer prediction tasks.

Effectiveness of the Capsule Network Architecture

The interaction between the primary capsule layer and the type capsule layer is at the core of the entire capsule network architecture. To visually demonstrate the superiority of the capsule network architecture, we extracted features from the samples in train sets at both the primary capsule layer and the class capsule layer in two distinct stages. Subsequently, utilizing the PCA dimensionality reduction technique, we reduced the extracted features to 2 dimensions, corresponding to the scatter plots shown in Figure 5. Red and blue points represent positive and negative samples, respectively (stage 1: enhancer and nonenhancer; stage 2: strong enhancer and weak enhancer).

Figure 5

Figure 5. Visualization of positive and negative samples of train set in primary capsule and type capsule layers of CapsEnhancer in two stages. (A) Primary capsule layer of stage 1. (B) Type capsule layer of stage 1. (C) Primary capsule layer of stage 2. (D) Type capsule layer of stage 2.

High Resolution Image
Download MS PowerPoint Slide
From Figure 5A,C, it can be observed that the points representing positive and negative samples are entangled, exhibiting similar distributions, making it challenging to distinguish between them. However, after undergoing the capsule network architecture, the red and blue points manifest clearly distinct distributions, facilitating easy differentiation. This indicates that the computational processes in the primary capsule layer and class capsule layer further refine the features. The processed features enable positive and negative samples to exhibit disparate distributions, thereby enhancing the predictive capabilities of the model.
This improvement is attributed to the dynamic routing algorithm of the capsule network, which allows information propagation and weight adjustments between different capsules. This dynamic routing mechanism is crucial for the capsule network’s ability to model spatial relationships among different features.

Ablation Experiment

Subsequently, we conducted ablation experiments to further validate the significance of the capsule network. We replaced the capsule network with a multilayer perceptron and performed experiments at both stages. The experimental results are presented in Table 5. It is evident that upon removing the capsule network, the model’s performance significantly deteriorated at both stages. In the first stage, the accuracy dropped by 17%, reaching only 77.3% compared to CapsEnhancer. In the second stage, the accuracy was lower by 9%, reaching only 86% compared to CapsEnhancer. In terms of MCC, the absence of the capsule network architecture resulted in a decrease of 0.345 and 0.161 at the two stages, achieving MCC values of 0.545 and 0.742, respectively.
Table 5. Performance Comparison with Other Existing Methods on the Independent Test Set of the Second Stage: Strong Enhancer Versus Weak Enhancer
stagemethodaccuracy (%)sensitivity (%)specificity (%)MCCAUC (%)
first stagewithout CapsNet77.379.075.50.5582.8
 CapsEnhancer94.593.096.00.8098.0
second stagewithout CapsNet86.098.074.00.7291.8
 CapsEnhancer95.099.091.00.9399.2
Furthermore, in the second stage of the model, without the capsule network, the sensitivity and specificity were 98 and 74%, respectively. This indicates that in the absence of the capsule network, the model not only fails to achieve precise predictions but also lacks the ability to achieve a balanced prediction.
In order to visually demonstrate the performance of the ablation experiments, we plotted the receiver operating characteristic (ROC) curves for CapsEnhancer and models without CapsNst, in two stages as illustrated in Figure 6. As evident from Figure 6, whether in stage 1 or stage 2, the ROC curve corresponding to CapsEnhancer consistently resides outside that of the model without CapsNst, achieving a higher AUC. In addition, we also plotted the precision-recall (PR) curves for both stages, as depicted in Figure S3. Similar to ROC curves, CapsEnhancer achieved a higher area under the PR compared to models without CapsNet. This further underscores the significance of the capsule network architecture in improving the predictive capabilities for the enhancer task.

Figure 6

Figure 6. ROC curves for CapsEnhancer and the model without capsule network in (A) stage 1 and (B) stage 2.

High Resolution Image
Download MS PowerPoint Slide

Feature Analysis

Within the domain of deep learning, discriminative features play a pivotal role in the development of robust classifiers. In contrast to existing methods, CapsEnhancer exhibits dual principal advantages: first, it leverages CGR encoding for the representation of DNA sequences, and second, it effectively learns from CGR images through the deployment of capsule network architecture. Consider the first stage, where the type capsule layer encompasses two capsules, each constituting a 32-dimensional vector corresponding to enhancer or nonenhancer categories. This configuration facilitates the construction of distinct features associated with enhancer and nonenhancer attributes. To underscore the discriminative efficacy of features extracted by CapsEnhancer concerning enhancers, 30 enhancers and 30 nonenhancers were randomly selected from the test set for comprehensive feature clustering analysis on the corresponding DNA sequences. In this analysis, we used hierarchical clustering with the complete linkage method. The resultant clustering patterns, illustrated in Figure 7, reveal two key observations: enhancers and nonenhancers distinctly cluster into separate subtrees, and DNA sequences of the same classification often exhibit analogous feature patterns. These findings provide compelling evidence that the features derived from the proposed CapsEnhancer method adeptly encapsulate traits pertinent to enhancers, offering further justification for the method’s effectiveness.

Figure 7

Figure 7. Clustering analysis map of latent features generated by CapsEnhancer on the independent test set in stage 1.

High Resolution Image
Download MS PowerPoint Slide

Conclusions

Click to copy section linkSection link copied!
Enhancers are a type of noncoding DNA element that can regulate gene expression. The identification of enhancers is crucial in the field of biology. First, it provides insights into the complex networks of gene regulation that govern various biological processes, such as development, differentiation, and response to environmental stimuli. By pinpointing enhancers associated with specific genes, researchers can unravel the molecular mechanisms underlying normal cellular functions and pathological conditions. Furthermore, the identification of enhancers has significant implications in the context of human health. Dysregulation of gene expression, often influenced by aberrant enhancer activity, is implicated in numerous diseases, including cancers and developmental disorders. Unraveling enhancer landscapes helps researchers identify potential therapeutic targets and develop strategies for precise intervention in gene expression patterns.
Traditional experimental methods, while effective in identifying enhancers, often demand substantial human and financial resources. In recent years, there has been an increasing emphasis on employing computational approaches for enhancer identification, driven by the rapid advancements in artificial intelligence methods. In this study, we propose a two-stage framework, CapsEnhancer, based on deep learning to efficiently predict enhancers and their strengths. The first stage focuses on identifying enhancers, while the second stage aims to predict strong and weak enhancers. Initially, we employ CGR encoding to represent each DNA sequence as an image, enabling the efficient representation of Kmers and frequencies. Furthermore, we utilize a capsule network-based architecture to extract local and global features of the images, overcoming the limitations of traditional CNNs and providing spatial modeling for features of these images. Experimental results demonstrate the outstanding predictive capabilities of our method in both stages, achieving state-of-the-art performance. This study employs computer vision methods to handle sequence data, and we believe that our research not only offers novel insights into enhancer identification but also provides a fresh perspective for other biological sequence analysis tasks.

Key Points

Click to copy section linkSection link copied!

  • We proposed a two-stage framework, CapsEnhancer, based on deep learning, for accurate prediction of enhancers and their strength.

  • CapsEnhancer employs CGR encoding to represent each DNA sequence as an image. Through this encoding methodology, it enables effective representation of Kmers and their frequencies.

  • CapsEnhancer utilizes an architecture based on capsule networks to learn both local and global features from DNA “images”. Capsule networks overcome the limitations of traditional CNNs by capturing spatial relationships among features in DNA “images”, thereby enhancing the model’s performance.

  • The framework proposed in our study employs computer vision strategies to process biosequence data, complemented by the integration of a next-generation neural network, the capsule network. This presents a novel approach and perspective for tasks of biosequence data analysis.

Data Availability

Click to copy section linkSection link copied!

Availability and implementationCapsEnhancer and data sets of this study are available at https://github.com/Cpillar/CapsEnhancer.

Supporting Information

Click to copy section linkSection link copied!

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jcim.4c00546.

  • GC content distribution of the data sets, FCGR images for two example sequences and PR curves for the ablation experiment, hyperparameters of CapsEnhancer, a case study, and description of the dynamic routing algorithm (PDF)

Terms & Conditions

Most electronic Supporting Information files are available without a subscription to ACS Web Editions. Such files may be downloaded by article for research use (if there is a public use license linked to the relevant article, that license may permit other uses). Permission may be obtained from ACS for other uses through requests via the RightsLink permission system: http://pubs.acs.org/page/copyright/permissions.html.

Author Information

Click to copy section linkSection link copied!
  • Corresponding Authors
    • Ying-Chih Chiang - Kobilka Institute of Innovative Drug Discovery, School of Medicine, The Chinese University of Hong Kong, Shenzhen 518172, China;  School of Medicine, The Chinese University of Hong Kong, Shenzhen 518172, China;  Email: [email protected]
    • Tzong-Yi Lee - Institute of Bioinformatics and Systems Biology, National Yang Ming Chiao Tung University, Hsinchu 300093, Taiwan;  Center for Intelligent Drug Systems and Smart Bio-devices (IDS2B), National Yang Ming Chiao Tung University, Hsinchu 300093, Taiwan;  Email: [email protected]
  • Authors
    • Lantian Yao - Kobilka Institute of Innovative Drug Discovery, School of Medicine, The Chinese University of Hong Kong, Shenzhen 518172, China;  School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen 518172, China;  Orcidhttps://orcid.org/0000-0003-4554-6827
    • Peilin Xie - Kobilka Institute of Innovative Drug Discovery, School of Medicine, The Chinese University of Hong Kong, Shenzhen 518172, China
    • Jiahui Guan - School of Medicine, The Chinese University of Hong Kong, Shenzhen 518172, China
    • Chia-Ru Chung - Department of Computer Science and Information Engineering, National Central University, Taoyuan 320317, Taiwan
    • Yixian Huang - School of Medicine, The Chinese University of Hong Kong, Shenzhen 518172, China
    • Yuxuan Pang - Division of Health Medical Intelligence, Human Genome Center, The Institute of Medical Science, The University of Tokyo, Tokyo 108-8639, Japan
    • Huacong Wu - School of Medicine, The Chinese University of Hong Kong, Shenzhen 518172, China
  • Author Contributions

    L.Y. and P.X. contributed equally to this work. L.Y. and T.-Y.L. presented the idea. L.Y., P.X., and J.G. implemented the framework. C.-R.C. collected the data. L.Y., Y.H, Y.P., and H.W. analyzed the results. T.-Y.L. and Y.-C.C. supervised the research project.

  • Funding

    This work was supported by National Natural Science Foundation of China (32070659), Shenzhen Science and Technology Innovation Commission (JCYJ20230807114206014), and the Kobilka Institute of Innovative Drug Discovery, The Chinese University of Hong Kong, Shenzhen, China. This work was also financially supported by the Center for Intelligent Drug Systems and Smart Biodevices (IDS2B) from The Featured Areas Research Center Program within the framework of the Higher Education Sprout Project and the Yushan Young Fellow Program (112C1N084C) by the Ministry of Education (MOE), National Science and Technology Council (NSTC 112-2321-B-A49-016 and 113-2634-F-039-001), and The National Health Research Institutes (NHRI-EX113-11320BI) in Taiwan.

  • Notes
    The authors declare no competing financial interest.

Acknowledgments

Click to copy section linkSection link copied!

The authors sincerely appreciate the Kobilka Institute of Innovative Drug Discovery, The Chinese University of Hong Kong (Shenzhen), and the “Center for intelligent Drug Systems and Smart Biodevices” from The Featured Areas Research Center Program within the framework of the Higher Education Sprout Project by the Ministry of Education in Taiwan. Y.-C.C. thanks The Royal Society for Newtown International Fellowship Alumni 2023 (AL\31027).

References

Click to copy section linkSection link copied!

This article references 51 other publications.

  1. 1
    Basith, S.; Hasan, M. M.; Lee, G.; Wei, L.; Manavalan, B. Integrative machine learning framework for the identification of cell-specific enhancers from the human genome. Briefings Bioinf. 2021, 22, bbab252,  DOI: 10.1093/bib/bbab252
    Google Scholar
    There is no corresponding record for this reference.
  2. 2
    Corradin, O.; Scacheri, P. Enhancer variants: evaluating functions in common disease. Genome Med. 2014, 6 (10), 85,  DOI: 10.1186/s13073-014-0085-3
    Google Scholar
    There is no corresponding record for this reference.
  3. 3
    Levine, M. Transcriptional enhancers in animal development and evolution. Curr. Biol. 2010, 20, R754– R763,  DOI: 10.1016/j.cub.2010.06.070
    Google Scholar
    There is no corresponding record for this reference.
  4. 4
    Zhang, L.; Yang, Y.; Chai, L.; Li, Q.; Liu, J.; Lin, H.; Liu, L. A deep learning model to identify gene expression level using cobinding transcription factor signals. Briefings Bioinf. 2022, 23, bbab501,  DOI: 10.1093/bib/bbab501
    Google Scholar
    There is no corresponding record for this reference.
  5. 5
    Heinz, S.; Romanoski, C. E.; Benner, C.; Glass, C. K. The selection and function of cell type-specific enhancers. Nat. Rev. Mol. Cell Biol. 2015, 16, 144– 154,  DOI: 10.1038/nrm3949
    Google Scholar
    There is no corresponding record for this reference.
  6. 6
    Furlong, E. E.; Levine, M. Developmental enhancers and chromosome topology. Science 2018, 361, 1341– 1345,  DOI: 10.1126/science.aau0320
    Google Scholar
    6
    Developmental enhancers and chromosome topology
    Furlong, Eileen E. M.; Levine, Michael
    Science (Washington, DC, United States) (2018), 361 (6409), 1341-1345CODEN: SCIEAS; ISSN:0036-8075. (American Association for the Advancement of Science)
    A review. Developmental enhancers mediate on/off patterns of gene expression in specific cell types at particular stages during metazoan embryogenesis. They typically integrate multiple signals and regulatory determinants to achieve precise spatiotemporal expression. Such enhancers can map quite far-one megabase or more-from the genes they regulate. How remote enhancers relay regulatory information to their target promoters is one of the central mysteries of genome organization and function. A variety of contrasting mechanisms have been proposed over the years, including enhancer tracking, linking, looping, and mobilization to transcription factories. We argue that extreme versions of these mechanisms cannot account for the transcriptional dynamics and precision seen in living cells, tissues, and embryos. We describe emerging evidence for dynamic three-dimensional hubs that combine different elements of the classical models.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhslOrurrJ&md5=c34d974a2b7dd2257a3c1c9546f7984d
  7. 7
    Schoenfelder, S.; Fraser, P. Long-range enhancer–promoter contacts in gene expression control. Nat. Rev. Genet. 2019, 20, 437– 455,  DOI: 10.1038/s41576-019-0128-0
    Google Scholar
    There is no corresponding record for this reference.
  8. 8
    Bauer, D. E.; Orkin, S. H. Hemoglobin switching’s surprise: the versatile transcription factor BCL11A is a master repressor of fetal hemoglobin. Curr. Opin. Genet. Dev. 2015, 33, 62– 70,  DOI: 10.1016/j.gde.2015.08.001
    Google Scholar
    There is no corresponding record for this reference.
  9. 9
    Chen, X.; Xu, H.; Yuan, P.; Fang, F.; Huss, M.; Vega, V. B.; Wong, E.; Orlov, Y. L.; Zhang, W.; Jiang, J. Integration of external signaling pathways with the core transcriptional network in embryonic stem cells. Cell 2008, 133, 1106– 1117,  DOI: 10.1016/j.cell.2008.04.043
    Google Scholar
    9
    Integration of external signaling pathways with the core transcriptional network in embryonic stem cells
    Chen, Xi; Xu, Han; Yuan, Ping; Fang, Fang; Huss, Mikael; Vega, Vinsensius B.; Wong, Eleanor; Orlov, Yuriy L.; Zhang, Weiwei; Jiang, Jianming; Loh, Yuin-Han; Yeo, Hock Chuan; Yeo, Zhen Xuan; Narang, Vipin; Govindarajan, Kunde Ramamoorthy; Leong, Bernard; Shahab, Atif; Ruan, Yijun; Bourque, Guillaume; Sung, Wing-Kin; Clarke, Neil D.; Wei, Chia-Lin; Ng, Huck-Hui
    Cell (Cambridge, MA, United States) (2008), 133 (6), 1106-1117CODEN: CELLB5; ISSN:0092-8674. (Cell Press)
    Transcription factors (TFs) and their specific interactions with targets are crucial for specifying gene-expression programs. To gain insights into the transcriptional regulatory networks in embryonic stem (ES) cells, we use chromatin immunopptn. coupled with ultra-high-throughput DNA sequencing (ChIP-seq) to map the locations of 13 sequence-specific TFs (Nanog, Oct4, STAT3, Smad1, Sox2, Zfx, c-Myc, n-Myc, Klf4, Esrrb, Tcfcp2l1, E2f1, and CTCF) and 2 transcription regulators (p300 and Suz12). These factors are known to play different roles in ES-cell biol. as components of the LIF and BMP signaling pathways, self-renewal regulators, and key reprogramming factors. Our study provides insights into the integration of the signaling pathways into the ES-cell-specific transcription circuitries. Intriguingly, we find specific genomic regions extensively targeted by different TFs. Collectively, the comprehensive mapping of TF-binding sites identifies important features of the transcriptional regulatory networks that define ES-cell identity.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BD1cXnsF2gurw%253D&md5=607c91e79f165fd417659c688a9725da
  10. 10
    May, D.; Blow, M. J.; Kaplan, T.; McCulley, D. J.; Jensen, B. C.; Akiyama, J. A.; Holt, A.; Plajzer-Frick, I.; Shoukry, M.; Wright, C. Large-scale discovery of enhancers from human heart tissue. Nat. Genet. 2012, 44, 89– 93,  DOI: 10.1038/ng.1006
    Google Scholar
    There is no corresponding record for this reference.
  11. 11
    Visel, A.; Blow, M. J.; Li, Z.; Zhang, T.; Akiyama, J. A.; Holt, A.; Plajzer-Frick, I.; Shoukry, M.; Wright, C.; Chen, F. ChIP-seq accurately predicts tissue-specific activity of enhancers. Nature 2009, 457, 854– 858,  DOI: 10.1038/nature07730
    Google Scholar
    11
    ChIP-seq accurately predicts tissue-specific activity of enhancers
    Visel, Axel; Blow, Matthew J.; Li, Zirong; Zhang, Tao; Akiyama, Jennifer A.; Holt, Amy; Plajzer-Frick, Ingrid; Shoukry, Malak; Wright, Crystal; Chen, Feng; Afzal, Veena; Ren, Bing; Rubin, Edward M.; Pennacchio, Len A.
    Nature (London, United Kingdom) (2009), 457 (7231), 854-858CODEN: NATUAS; ISSN:0028-0836. (Nature Publishing Group)
    A major yet unresolved quest in decoding the human genome is the identification of the regulatory sequences that control the spatial and temporal expression of genes. Distant-acting transcriptional enhancers are particularly challenging to uncover because they are scattered among the vast non-coding portion of the genome. Evolutionary sequence constraint can facilitate the discovery of enhancers, but fails to predict when and where they are active in vivo. Here we present the results of chromatin immunopptn. with the enhancer-assocd. protein p300 followed by massively parallel sequencing, and map several thousand in vivo binding sites of p300 in mouse embryonic forebrain, midbrain and limb tissue. We tested 86 of these sequences in a transgenic mouse assay, which in nearly all cases demonstrated reproducible enhancer activity in the tissues that were predicted by p300 binding. Our results indicate that in vivo mapping of p300 binding is a highly accurate means for identifying enhancers and their assocd. activities, and suggest that such data sets will be useful to study the role of tissue-specific enhancers in human biol. and disease on a genome-wide scale.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BD1MXhslCntLg%253D&md5=2780a5b9f819181972b13e0c8ecc7a4b
  12. 12
    Pennacchio, L. A.; Bickmore, W.; Dean, A.; Nobrega, M. A.; Bejerano, G. Enhancers: five essential questions. Nat. Rev. Genet. 2013, 14, 288– 295,  DOI: 10.1038/nrg3458
    Google Scholar
    12
    Enhancers: five essential questions
    Pennacchio, Len A.; Bickmore, Wendy; Dean, Ann; Nobrega, Marcelo A.; Bejerano, Gill
    Nature Reviews Genetics (2013), 14 (4), 288-295CODEN: NRGAAM; ISSN:1471-0056. (Nature Publishing Group)
    A review. Although enhancers are crucial and widespread gene-regulatory elements, we are far from a complete understanding of how they function or their importance in areas such as disease and evolution. Five prominent researchers discuss some of the key outstanding questions in enhancer biol.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3sXktVKhu7c%253D&md5=d86155e239743ebb1711130132dc521f
  13. 13
    Ku, C. S.; Naidoo, N.; Wu, M.; Soong, R. Studying the epigenome using next generation sequencing. J. Med. Genet. 2011, 48, 721– 730,  DOI: 10.1136/jmedgenet-2011-100242
    Google Scholar
    13
    Studying the epigenome using next generation sequencing
    Ku, Chee Seng; Naidoo, Nasheen; Wu, Mengchu; Soong, Richie
    Journal of Medical Genetics (2011), 48 (11), 721-730CODEN: JMDGAE; ISSN:0022-2593. (BMJ Publishing Group)
    A review. The advances in next generation sequencing (NGS) technologies have had a significant impact on epigenomic research. The arrival of NGS technologies has enabled a more powerful sequencing based method-i.e., ChlP-Seq-to interrogate whole genome histone modifications, improving on the conventional microarray based method (ChlP-chip). Similarly, the first human DNA methylome was mapped using NGS technologies. More importantly, studies of DNA methylation and histone modification using NGS technologies have yielded new discoveries and improved our knowledge of human biol. and diseases. The concept that cytosine methylation was restricted to CpG dinucleotides has only been recently challenged by new data generated from sequencing the DNA methylome. Approx. 25% of all cytosine methylation identified in stem cells was in a non-CG context. The non-CG methylation was more enriched in gene bodies and depleted in protein binding sites and enhancers. The recent developments of third generation sequencing technologies have shown promising results of directly sequencing methylated nucleotides and having the ability to differentiate between 5-methylcytosine and 5-hydroxymethylcytosine. The importance of 5-hydroxymethylcytosine remains largely unknown, but it has been found in various tissues. 5-Hydroxymethylcytosine was particularly enriched at promoters and in intragenic regions (gene bodies) but was largely absent from non-gene regions in DNA from human brain frontal lobe tissue. The presence of 5-hydroxymethylcytosine in gene bodies was more pos. correlated with gene expression levels. The importance of studying 5-methylcytosine and 5-hydroxymethylcytosine sep. for their biol. roles will become clearer when more efficient methods to distinguish them are available.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3MXhs1KktrzP&md5=aa3db24a99915dcb230fb5f340a2d778
  14. 14
    Liu, B.; Fang, L.; Long, R.; Lan, X.; Chou, K.-C. iEnhancer-2L: a two-layer predictor for identifying enhancers and their strength by pseudo k-tuple nucleotide composition. Bioinformatics 2016, 32, 362– 369,  DOI: 10.1093/bioinformatics/btv604
    Google Scholar
    There is no corresponding record for this reference.
  15. 15
    Jia, C.; He, W. EnhancerPred: a predictor for discovering enhancers based on the combination and selection of multiple features. Sci. Rep. 2016, 6, 38741,  DOI: 10.1038/srep38741
    Google Scholar
    There is no corresponding record for this reference.
  16. 16
    Lim, D. Y.; Khanal, J.; Tayara, H.; Chong, K. T. iEnhancer-RF: identifying enhancers and their strength by enhanced feature representation using random forest. Chemom. Intell. Lab. Syst. 2021, 212, 104284,  DOI: 10.1016/j.chemolab.2021.104284
    Google Scholar
    There is no corresponding record for this reference.
  17. 17
    Liu, B.; Li, K.; Huang, D.-S.; Chou, K.-C. iEnhancer-EL identifying enhancers and their strength with ensemble learning approach. Bioinformatics 2018, 34, 3835– 3842,  DOI: 10.1093/bioinformatics/bty458
    Google Scholar
    There is no corresponding record for this reference.
  18. 18
    Wang, C.; Zou, Q.; Ju, Y.; Shi, H. Enhancer-FRL: improved and robust identification of enhancers and their activities using feature representation learning. IEEE/ACM Trans. Comput. Biol. Bioinform. 2023, 20, 967– 975,  DOI: 10.1109/TCBB.2022.3204365
    Google Scholar
    There is no corresponding record for this reference.
  19. 19
    Gill, M.; Ahmed, S.; Kabir, M.; Hayat, M. A novel predictor for the analysis and prediction of enhancers and their strength via multi-view features and deep forest. Information 2023, 14, 636,  DOI: 10.3390/info14120636
    Google Scholar
    There is no corresponding record for this reference.
  20. 20
    Nguyen, Q. H.; Nguyen-Vo, T.-H.; Le, N. Q. K.; Do, T. T.; Rahardja, S.; Nguyen, B. P. iEnhancer-ECNN: identifying enhancers and their strength using ensembles of convolutional neural networks. BMC Genom. 2019, 20, 951,  DOI: 10.1186/s12864-019-6336-3
    Google Scholar
    There is no corresponding record for this reference.
  21. 21
    Le, N. Q. K.; Ho, Q.-T.; Nguyen, T.-T.-D.; Ou, Y.-Y. A transformer architecture based on BERT and 2D convolutional neural network to identify DNA enhancers from sequence information. Briefings Bioinf. 2021, 22, bbab005,  DOI: 10.1093/bib/bbab005
    Google Scholar
    There is no corresponding record for this reference.
  22. 22
    Niu, K.; Luo, X.; Zhang, S.; Teng, Z.; Zhang, T.; Zhao, Y. iEnhancer-EBLSTM: identifying enhancers and strengths by ensembles of bidirectional long short-term memory. Front. Genet. 2021, 12, 665498,  DOI: 10.3389/fgene.2021.665498
    Google Scholar
    There is no corresponding record for this reference.
  23. 23
    Cai, L.; Ren, X.; Fu, X.; Peng, L.; Gao, M.; Zeng, X. iEnhancer-XG: interpretable sequence-based enhancers and their strength predictor. Bioinformatics 2021, 37, 1060– 1067,  DOI: 10.1093/bioinformatics/btaa914
    Google Scholar
    There is no corresponding record for this reference.
  24. 24
    Yang, H.; Wang, S.; Xia, X. iEnhancer-RD: identification of enhancers and their strength using RKPK features and deep neural networks. Anal. Biochem. 2021, 630, 114318,  DOI: 10.1016/j.ab.2021.114318
    Google Scholar
    There is no corresponding record for this reference.
  25. 25
    Mu, X.; Wang, Y.; Duan, M.; Liu, S.; Li, F.; Wang, X.; Zhang, K.; Huang, L.; Zhou, F. A novel position-specific encoding algorithm (SeqPose) of nucleotide sequences and its application for detecting enhancers. Int. J. Mol. Sci. 2021, 22, 3079,  DOI: 10.3390/ijms22063079
    Google Scholar
    There is no corresponding record for this reference.
  26. 26
    Luo, H.; Chen, C.; Shan, W.; Ding, P.; Luo, L. iEnhancer-BERT: a novel transfer learning architecture based on DNA-Language model for identifying enhancers and their strength. In International Conference on Intelligent Computing , 2022; pp 153– 165. DOI: 10.1007/978-3-031-13829-4_13 .
    Google Scholar
    There is no corresponding record for this reference.
  27. 27
    Li, J.; Wu, Z.; Lin, W.; Luo, J.; Zhang, J.; Chen, Q.; Chen, J. iEnhancer-ELM: improve enhancer identification by extracting position-related multiscale contextual information based on enhancer language models. Bioinform. Adv. 2023, 3, vbad043,  DOI: 10.1093/bioadv/vbad043
    Google Scholar
    There is no corresponding record for this reference.
  28. 28
    Wang, W.; Wu, Q.; Li, C. iEnhancer-DCSA: identifying enhancers via dual-scale convolution and spatial attention. BMC Genom. 2023, 24, 393,  DOI: 10.1186/s12864-023-09468-1
    Google Scholar
    There is no corresponding record for this reference.
  29. 29
    Wu, H.; Liu, M.; Zhang, P.; Zhang, H. iEnhancer-SKNN: a stacking ensemble learning-based method for enhancer identification and classification using sequence information. Briefings Funct. Genomics 2023, 22, 302– 311,  DOI: 10.1093/bfgp/elac057
    Google Scholar
    There is no corresponding record for this reference.
  30. 30
    Ng, P. dna2vec: consistent vector representations of variable-length k-mers. arXiv 2017, arXiv:1701.06279
    Google Scholar
    There is no corresponding record for this reference.
  31. 31
    Iuchi, H.; Matsutani, T.; Yamada, K.; Iwano, N.; Sumi, S.; Hosoda, S.; Zhao, S.; Fukunaga, T.; Hamada, M. Representation learning applications in biological sequence analysis. Comput. Struct. Biotechnol. J. 2021, 19, 3198– 3208,  DOI: 10.1016/j.csbj.2021.05.039
    Google Scholar
    31
    Representation learning applications in biological sequence analysis
    Iuchi, Hitoshi; Matsutani, Taro; Yamada, Keisuke; Iwano, Natsuki; Sumi, Shunsuke; Hosoda, Shion; Zhao, Shitao; Fukunaga, Tsukasa; Hamada, Michiaki
    Computational and Structural Biotechnology Journal (2021), 19 (), 3198-3208CODEN: CSBJAC; ISSN:2001-0370. (Elsevier B.V.)
    A review. Although remarkable advances have been reported in high-throughput sequencing, the ability to aptly analyze a substantial amt. of rapidly generated biol. (DNA/RNA/protein) sequencing data remains a crit. hurdle. To tackle this issue, the application of natural language processing (NLP) to biol. sequence anal. has received increased attention. In this method, biol. sequences are regarded as sentences while the single nucleic acids/amino acids or k-mers in these sequences represent the words. Embedding is an essential step in NLP, which performs the conversion of these words into vectors. Specifically, representation learning is an approach used for this transformation process, which can be applied to biol. sequences. Vectorized biol. sequences can then be applied for function and structure estn., or as input for other probabilistic models. Considering the importance and growing trend for the application of representation learning to biol. research, in the present study, we have reviewed the existing knowledge in representation learning for biol. sequence anal.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXitFKisrjL&md5=3fc5035b7e5c9ac8df53cc3b7f4d538c
  32. 32
    Wen, J.; Liu, Y.; Shi, Y.; Huang, H.; Deng, B.; Xiao, X. A classification model for lncRNA and mRNA based on k-mers and a convolutional neural network. BMC Bioinf. 2019, 20, 469,  DOI: 10.1186/s12859-019-3039-3
    Google Scholar
    There is no corresponding record for this reference.
  33. 33
    Li, W.; Guo, Y.; Wang, B.; Yang, B. Learning spatiotemporal embedding with gated convolutional recurrent networks for translation initiation site prediction. Pattern Recogn. 2023, 136, 109234,  DOI: 10.1016/j.patcog.2022.109234
    Google Scholar
    There is no corresponding record for this reference.
  34. 34
    Jeffrey, H. J. Chaos game representation of gene structure. Nucleic Acids Res. 1990, 18, 2163– 2170,  DOI: 10.1093/nar/18.8.2163
    Google Scholar
    34
    Chaos game representation of gene structure
    Jeffrey, H. Joel
    Nucleic Acids Research (1990), 18 (8), 2163-70CODEN: NARHAD; ISSN:0305-1048.
    This paper presents a new method for representing DNA sequences. It permits the representation and investigation of patterns in sequences, visually revealing previously unknown structures. Based on a technique from chaotic dynamics, the method produces a picture of a gene sequence which displays both local and global patterns. The pictures have a complex structure which varies depending on the sequence. The method is termed Chaos Game Representation (CGR). CGR raises a new set of questions about the structure of DNA sequences, and is a new tool for investigating gene structure.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADyaK3cXksFeltrk%253D&md5=2832019a98de2fa2987a2f70e5f32340
  35. 35
    Löchel, H. F.; Eger, D.; Sperlea, T.; Heider, D. Deep learning on chaos game representation for proteins. Bioinformatics 2020, 36, 272– 279,  DOI: 10.1093/bioinformatics/btz493
    Google Scholar
    There is no corresponding record for this reference.
  36. 36
    LaLonde, R.; Bagci, U. Capsules for object segmentation. arXiv 2018, arXiv:1804.04241
    Google Scholar
    There is no corresponding record for this reference.
  37. 37
    Dong, Z.; Lin, S. Research on image classification based on capsnet. In 2019 IEEE 4th Advanced Information Technology, Electronic and Automation Control Conference (IAEAC) , 2019; pp 1023– 1026.
    Google Scholar
    There is no corresponding record for this reference.
  38. 38
    Guo, Y.; Zhou, D.; Ruan, X.; Cao, J. Variational gated autoencoder-based feature extraction model for inferring disease-miRNA associations based on multiview features. Neural Network. 2023, 165, 491– 505,  DOI: 10.1016/j.neunet.2023.05.052
    Google Scholar
    There is no corresponding record for this reference.
  39. 39
    Guo, Y.; Zhou, D.; Li, P.; Li, C.; Cao, J. Context-aware poly (a) signal prediction model via deep spatial–temporal neural networks. IEEE Trans. Neural Netw. Learn. Syst. 2024, 35, 8241– 8253,  DOI: 10.1109/tnnls.2022.3226301
    Google Scholar
    There is no corresponding record for this reference.
  40. 40
    Wang, X.; Guan, Z.; Qian, W.; Cao, J.; Wang, C.; Ma, R. STFuse: infrared and visible image fusion via semisupervised transfer learning. IEEE Trans. Neural Netw. Learn. Syst. 2024, 1– 14,  DOI: 10.1109/tnnls.2023.3328060
    Google Scholar
    There is no corresponding record for this reference.
  41. 41
    Sabour, S.; Frosst, N.; Hinton, G. E. Dynamic routing between capsules. In Advances in Neural Information Processing Systems , 2017; Vol. 30.
    Google Scholar
    There is no corresponding record for this reference.
  42. 42
    Yao, L.; Pang, Y.; Wan, J.; Chung, C.-R.; Yu, J.; Guan, J.; Leung, C.; Chiang, Y.-C.; Lee, T.-Y. ABPCaps: a novel capsule network-based method for the prediction of antibacterial peptides. Appl. Sci. 2023, 13, 6965,  DOI: 10.3390/app13126965
    Google Scholar
    There is no corresponding record for this reference.
  43. 43
    Huang, Y.; Huang, H.-Y.; Chen, Y.; Lin, Y.-C.-D.; Yao, L.; Lin, T.; Leng, J.; Chang, Y.; Zhang, Y.; Zhu, Z. A robust drug–target interaction prediction framework with capsule network and transfer learning. Int. J. Mol. Sci. 2023, 24, 14061,  DOI: 10.3390/ijms241814061
    Google Scholar
    There is no corresponding record for this reference.
  44. 44
    Wang, D.; Liang, Y.; Xu, D. Capsule network for protein post-translational modification site prediction. Bioinformatics 2019, 35, 2386– 2394,  DOI: 10.1093/bioinformatics/bty977
    Google Scholar
    44
    Capsule network for protein post-translational modification site prediction
    Wang, Duolin; Liang, Yanchun; Xu, Dong
    Bioinformatics (2019), 35 (14), 2386-2394CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
    Motivation: Computational methods for protein post-translational modification (PTM) site prediction provide a useful approach for studying protein functions. The prediction accuracy of the existing methods has significant room for improvement. A recent deep-learning architecture, Capsule Network (CapsNet), which can characterize the internal hierarchical representation of input data, presents a great opportunity to solve this problem, esp. using small training data. Results: We proposed a CapsNet for predicting protein PTM sites, including phosphorylation, Nlinked glycosylation, N6-acetyllysine, methyl-arginine, S-palmitoyl-cysteine, pyrrolidonecarboxylic- acid and SUMOylation sites. The CapsNet outperformed the baseline convolutional neural network architecture MusiteDeep and other well-known tools in most cases and provided promising results for practical use, esp. in learning from small training data. The capsule length also gives an accurate est. for the confidence of the PTM prediction. We further demonstrated that the internal capsule features could be trained as a motif detector of phosphorylation sites when no kinase-specific phosphorylation labels were provided. In addn., CapsNet generates robust representations that have strong discriminant power in distinguishing kinase substrates from different kinase families. Our study sheds some light on the recognition mechanism of PTMs and applications of CapsNet on other bioinformatic problems.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXjvF2lsbc%253D&md5=824e5d94dc1f2ddb865ac97e1fc57be1
  45. 45
    Khanal, J.; Tayara, H.; Zou, Q.; To Chong, K. DeepCap-Kcr: accurate identification and investigation of protein lysine crotonylation sites based on capsule network. Briefings Bioinf. 2022, 23, bbab492,  DOI: 10.1093/bib/bbab492
    Google Scholar
    There is no corresponding record for this reference.
  46. 46
    Shang, J.; Peng, C.; Tang, X.; Sun, Y. PhaVIP: Phage VIrion protein classification based on chaos game representation and vision transformer. arXiv 2023, arXiv:2301.12422
    Google Scholar
    There is no corresponding record for this reference.
  47. 47
    Löchel, H. F.; Heider, D. Chaos game representation and its applications in bioinformatics. Comput. Struct. Biotechnol. J. 2021, 19, 6263– 6271,  DOI: 10.1016/j.csbj.2021.11.008
    Google Scholar
    47
    Chaos game representation and its applications in bioinformatics
    Lochel Hannah Franziska; Heider Dominik
    Computational and structural biotechnology journal (2021), 19 (), 6263-6271 ISSN:2001-0370.
    Chaos game representation (CGR), a milestone in graphical bioinformatics, has become a powerful tool regarding alignment-free sequence comparison and feature encoding for machine learning. The algorithm maps a sequence to 2-dimensional space, while an extension of the CGR, the so-called frequency matrix representation (FCGR), transforms sequences of different lengths into equal-sized images or matrices. The CGR is a generalized Markov chain and includes various properties, which allow a unique representation of a sequence. Therefore, it has a broad spectrum of applications in bioinformatics, such as sequence comparison and phylogenetic analysis and as an encoding of sequences for machine learning. This review introduces the construction of CGRs and FCGRs, their applications on DNA and proteins, and gives an overview of recent applications and progress in bioinformatics.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BB2cblt1Ojsg%253D%253D&md5=8e8167b2552e037e8c046d3dcf24e119
  48. 48
    Kingma, D. P.; Ba, J. Adam: a method for stochastic optimization. arXiv 2014, arXiv:1412.6980
    Google Scholar
    There is no corresponding record for this reference.
  49. 49
    Paszke, A.; Gross, S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T.; Lin, Z.; Gimelshein, N.; Antiga, L.; Pytorch: an imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems , 2019; Vol. 32.
    Google Scholar
    There is no corresponding record for this reference.
  50. 50
    Kishk, A.; Elzizy, A.; Galal, D.; Razek, E. A.; Fawzy, E.; Ahmed, G.; Gawish, M.; Hamad, S.; El-Hadidi, M. A hybrid machine learning approach for the phenotypic classification of metagenomic colon cancer reads based on kmer frequency and biomarker profiling. In 2018 9th Cairo International Biomedical Engineering Conference (CIBEC) , 2018; pp 118– 121.
    Google Scholar
    There is no corresponding record for this reference.
  51. 51
    Yin, B.; Balvert, M.; Zambrano, D.; Schönhuth, A.; Bohte, S. An image representation based convolutional network for DNA classification. arXiv 2018, arXiv:1806.04931
    Google Scholar
    There is no corresponding record for this reference.

Cited By

Click to copy section linkSection link copied!

This article has not yet been cited by other publications.

Download PDF
Open PDF
Go to Journal of Chemical Information and Modeling
Get e-Alerts
Get e-Alerts

Journal of Chemical Information and Modeling

Cite this: J. Chem. Inf. Model. 2024, 64, 14, 5725–5736
Click to copy citationCitation copied!
https://doi.org/10.1021/acs.jcim.4c00546
Published June 30, 2024

Copyright © 2024 The Authors. Published by American Chemical Society. This publication is licensed under

CC-BY 4.0 .

Article Views

728

Altmetric

-

Citations

-
Learn about these metrics

Article Views are the COUNTER-compliant sum of full text article downloads since November 2008 (both PDF and HTML) across all institutions and individuals. These metrics are regularly updated to reflect usage leading up to the last few days.

Citations are the number of other articles citing this article, calculated by Crossref and updated daily. Find more information about Crossref citation counts.

The Altmetric Attention Score is a quantitative measure of the attention that a research article has received online. Clicking on the donut icon will load a page at altmetric.com with additional details about the score and the social media presence for the given article. Find more information on the Altmetric Attention Score and how the score is calculated.

  • Abstract

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 1

    Figure 1. Workflow of CapsEnhancer. First, we utilized benchmark data sets from previous studies. Subsequently, each DNA sequence was encoded using CGR encoding and represented as corresponding two-dimensional images. The model was then constructed using an architecture based on capsule networks. Hyperparameter adjustment was performed through fivefold cross-validation, and the model was evaluated using an independent test set, with the subsequent reporting of model performance metrics. The trained model was ultimately employed for enhancer identification, constituting a two-stage task. The first stage focused on discerning enhancers from nonenhancers, while the second stage aimed to predict enhancer strength, i.e., strong enhancers versus weak enhancers. The second stage employs the same FCGR images as the first stage to maintain consistency in the input representation. Capsule networks are used in both stages to build the models.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 2

    Figure 2. Architecture of CapsEnhancer. First, DNA sequences are encoded using CGR encoding and represented as 2D images. Subsequently, they are input into a Conv2D for preliminary feature extraction. Following this, the data is fed into a capsule network, which consists of a primary capsule layer and a type capsule layer. The primary capsule layer includes a Conv2D for further extracting local features. Then, a dynamic routing algorithm is utilized to capture the spatial relationships of features, resulting in the type capsule layer. As the task is a standard binary classification, the type capsule layer comprises two capsules, corresponding to the positive class and the negative class (stage 1: enhancer versus nonenhancer; stage 2: strong enhancer versus weak enhancer). Finally, the prediction probabilities for the two classes are obtained by calculating the lengths of the capsules in the type capsule layer.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 3

    Figure 3. (A) Applying CGR encoding to an example sequence: CATG. (B) Dividing the CGR space during the iterative process.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 4

    Figure 4. Computational process between primary capsules and type capsules.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 5

    Figure 5. Visualization of positive and negative samples of train set in primary capsule and type capsule layers of CapsEnhancer in two stages. (A) Primary capsule layer of stage 1. (B) Type capsule layer of stage 1. (C) Primary capsule layer of stage 2. (D) Type capsule layer of stage 2.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 6

    Figure 6. ROC curves for CapsEnhancer and the model without capsule network in (A) stage 1 and (B) stage 2.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 7

    Figure 7. Clustering analysis map of latent features generated by CapsEnhancer on the independent test set in stage 1.

    High Resolution Image
    Download MS PowerPoint Slide

  • References

    This article references 51 other publications.

    1. 1
      Basith, S.; Hasan, M. M.; Lee, G.; Wei, L.; Manavalan, B. Integrative machine learning framework for the identification of cell-specific enhancers from the human genome. Briefings Bioinf. 2021, 22, bbab252,  DOI: 10.1093/bib/bbab252
      There is no corresponding record for this reference.
    2. 2
      Corradin, O.; Scacheri, P. Enhancer variants: evaluating functions in common disease. Genome Med. 2014, 6 (10), 85,  DOI: 10.1186/s13073-014-0085-3
      There is no corresponding record for this reference.
    3. 3
      Levine, M. Transcriptional enhancers in animal development and evolution. Curr. Biol. 2010, 20, R754– R763,  DOI: 10.1016/j.cub.2010.06.070
      There is no corresponding record for this reference.
    4. 4
      Zhang, L.; Yang, Y.; Chai, L.; Li, Q.; Liu, J.; Lin, H.; Liu, L. A deep learning model to identify gene expression level using cobinding transcription factor signals. Briefings Bioinf. 2022, 23, bbab501,  DOI: 10.1093/bib/bbab501
      There is no corresponding record for this reference.
    5. 5
      Heinz, S.; Romanoski, C. E.; Benner, C.; Glass, C. K. The selection and function of cell type-specific enhancers. Nat. Rev. Mol. Cell Biol. 2015, 16, 144– 154,  DOI: 10.1038/nrm3949
      There is no corresponding record for this reference.
    6. 6
      Furlong, E. E.; Levine, M. Developmental enhancers and chromosome topology. Science 2018, 361, 1341– 1345,  DOI: 10.1126/science.aau0320
      6
      Developmental enhancers and chromosome topology
      Furlong, Eileen E. M.; Levine, Michael
      Science (Washington, DC, United States) (2018), 361 (6409), 1341-1345CODEN: SCIEAS; ISSN:0036-8075. (American Association for the Advancement of Science)
      A review. Developmental enhancers mediate on/off patterns of gene expression in specific cell types at particular stages during metazoan embryogenesis. They typically integrate multiple signals and regulatory determinants to achieve precise spatiotemporal expression. Such enhancers can map quite far-one megabase or more-from the genes they regulate. How remote enhancers relay regulatory information to their target promoters is one of the central mysteries of genome organization and function. A variety of contrasting mechanisms have been proposed over the years, including enhancer tracking, linking, looping, and mobilization to transcription factories. We argue that extreme versions of these mechanisms cannot account for the transcriptional dynamics and precision seen in living cells, tissues, and embryos. We describe emerging evidence for dynamic three-dimensional hubs that combine different elements of the classical models.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhslOrurrJ&md5=c34d974a2b7dd2257a3c1c9546f7984d
    7. 7
      Schoenfelder, S.; Fraser, P. Long-range enhancer–promoter contacts in gene expression control. Nat. Rev. Genet. 2019, 20, 437– 455,  DOI: 10.1038/s41576-019-0128-0
      There is no corresponding record for this reference.
    8. 8
      Bauer, D. E.; Orkin, S. H. Hemoglobin switching’s surprise: the versatile transcription factor BCL11A is a master repressor of fetal hemoglobin. Curr. Opin. Genet. Dev. 2015, 33, 62– 70,  DOI: 10.1016/j.gde.2015.08.001
      There is no corresponding record for this reference.
    9. 9
      Chen, X.; Xu, H.; Yuan, P.; Fang, F.; Huss, M.; Vega, V. B.; Wong, E.; Orlov, Y. L.; Zhang, W.; Jiang, J. Integration of external signaling pathways with the core transcriptional network in embryonic stem cells. Cell 2008, 133, 1106– 1117,  DOI: 10.1016/j.cell.2008.04.043
      9
      Integration of external signaling pathways with the core transcriptional network in embryonic stem cells
      Chen, Xi; Xu, Han; Yuan, Ping; Fang, Fang; Huss, Mikael; Vega, Vinsensius B.; Wong, Eleanor; Orlov, Yuriy L.; Zhang, Weiwei; Jiang, Jianming; Loh, Yuin-Han; Yeo, Hock Chuan; Yeo, Zhen Xuan; Narang, Vipin; Govindarajan, Kunde Ramamoorthy; Leong, Bernard; Shahab, Atif; Ruan, Yijun; Bourque, Guillaume; Sung, Wing-Kin; Clarke, Neil D.; Wei, Chia-Lin; Ng, Huck-Hui
      Cell (Cambridge, MA, United States) (2008), 133 (6), 1106-1117CODEN: CELLB5; ISSN:0092-8674. (Cell Press)
      Transcription factors (TFs) and their specific interactions with targets are crucial for specifying gene-expression programs. To gain insights into the transcriptional regulatory networks in embryonic stem (ES) cells, we use chromatin immunopptn. coupled with ultra-high-throughput DNA sequencing (ChIP-seq) to map the locations of 13 sequence-specific TFs (Nanog, Oct4, STAT3, Smad1, Sox2, Zfx, c-Myc, n-Myc, Klf4, Esrrb, Tcfcp2l1, E2f1, and CTCF) and 2 transcription regulators (p300 and Suz12). These factors are known to play different roles in ES-cell biol. as components of the LIF and BMP signaling pathways, self-renewal regulators, and key reprogramming factors. Our study provides insights into the integration of the signaling pathways into the ES-cell-specific transcription circuitries. Intriguingly, we find specific genomic regions extensively targeted by different TFs. Collectively, the comprehensive mapping of TF-binding sites identifies important features of the transcriptional regulatory networks that define ES-cell identity.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BD1cXnsF2gurw%253D&md5=607c91e79f165fd417659c688a9725da
    10. 10
      May, D.; Blow, M. J.; Kaplan, T.; McCulley, D. J.; Jensen, B. C.; Akiyama, J. A.; Holt, A.; Plajzer-Frick, I.; Shoukry, M.; Wright, C. Large-scale discovery of enhancers from human heart tissue. Nat. Genet. 2012, 44, 89– 93,  DOI: 10.1038/ng.1006
      There is no corresponding record for this reference.
    11. 11
      Visel, A.; Blow, M. J.; Li, Z.; Zhang, T.; Akiyama, J. A.; Holt, A.; Plajzer-Frick, I.; Shoukry, M.; Wright, C.; Chen, F. ChIP-seq accurately predicts tissue-specific activity of enhancers. Nature 2009, 457, 854– 858,  DOI: 10.1038/nature07730
      11
      ChIP-seq accurately predicts tissue-specific activity of enhancers
      Visel, Axel; Blow, Matthew J.; Li, Zirong; Zhang, Tao; Akiyama, Jennifer A.; Holt, Amy; Plajzer-Frick, Ingrid; Shoukry, Malak; Wright, Crystal; Chen, Feng; Afzal, Veena; Ren, Bing; Rubin, Edward M.; Pennacchio, Len A.
      Nature (London, United Kingdom) (2009), 457 (7231), 854-858CODEN: NATUAS; ISSN:0028-0836. (Nature Publishing Group)
      A major yet unresolved quest in decoding the human genome is the identification of the regulatory sequences that control the spatial and temporal expression of genes. Distant-acting transcriptional enhancers are particularly challenging to uncover because they are scattered among the vast non-coding portion of the genome. Evolutionary sequence constraint can facilitate the discovery of enhancers, but fails to predict when and where they are active in vivo. Here we present the results of chromatin immunopptn. with the enhancer-assocd. protein p300 followed by massively parallel sequencing, and map several thousand in vivo binding sites of p300 in mouse embryonic forebrain, midbrain and limb tissue. We tested 86 of these sequences in a transgenic mouse assay, which in nearly all cases demonstrated reproducible enhancer activity in the tissues that were predicted by p300 binding. Our results indicate that in vivo mapping of p300 binding is a highly accurate means for identifying enhancers and their assocd. activities, and suggest that such data sets will be useful to study the role of tissue-specific enhancers in human biol. and disease on a genome-wide scale.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BD1MXhslCntLg%253D&md5=2780a5b9f819181972b13e0c8ecc7a4b
    12. 12
      Pennacchio, L. A.; Bickmore, W.; Dean, A.; Nobrega, M. A.; Bejerano, G. Enhancers: five essential questions. Nat. Rev. Genet. 2013, 14, 288– 295,  DOI: 10.1038/nrg3458
      12
      Enhancers: five essential questions
      Pennacchio, Len A.; Bickmore, Wendy; Dean, Ann; Nobrega, Marcelo A.; Bejerano, Gill
      Nature Reviews Genetics (2013), 14 (4), 288-295CODEN: NRGAAM; ISSN:1471-0056. (Nature Publishing Group)
      A review. Although enhancers are crucial and widespread gene-regulatory elements, we are far from a complete understanding of how they function or their importance in areas such as disease and evolution. Five prominent researchers discuss some of the key outstanding questions in enhancer biol.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3sXktVKhu7c%253D&md5=d86155e239743ebb1711130132dc521f
    13. 13
      Ku, C. S.; Naidoo, N.; Wu, M.; Soong, R. Studying the epigenome using next generation sequencing. J. Med. Genet. 2011, 48, 721– 730,  DOI: 10.1136/jmedgenet-2011-100242
      13
      Studying the epigenome using next generation sequencing
      Ku, Chee Seng; Naidoo, Nasheen; Wu, Mengchu; Soong, Richie
      Journal of Medical Genetics (2011), 48 (11), 721-730CODEN: JMDGAE; ISSN:0022-2593. (BMJ Publishing Group)
      A review. The advances in next generation sequencing (NGS) technologies have had a significant impact on epigenomic research. The arrival of NGS technologies has enabled a more powerful sequencing based method-i.e., ChlP-Seq-to interrogate whole genome histone modifications, improving on the conventional microarray based method (ChlP-chip). Similarly, the first human DNA methylome was mapped using NGS technologies. More importantly, studies of DNA methylation and histone modification using NGS technologies have yielded new discoveries and improved our knowledge of human biol. and diseases. The concept that cytosine methylation was restricted to CpG dinucleotides has only been recently challenged by new data generated from sequencing the DNA methylome. Approx. 25% of all cytosine methylation identified in stem cells was in a non-CG context. The non-CG methylation was more enriched in gene bodies and depleted in protein binding sites and enhancers. The recent developments of third generation sequencing technologies have shown promising results of directly sequencing methylated nucleotides and having the ability to differentiate between 5-methylcytosine and 5-hydroxymethylcytosine. The importance of 5-hydroxymethylcytosine remains largely unknown, but it has been found in various tissues. 5-Hydroxymethylcytosine was particularly enriched at promoters and in intragenic regions (gene bodies) but was largely absent from non-gene regions in DNA from human brain frontal lobe tissue. The presence of 5-hydroxymethylcytosine in gene bodies was more pos. correlated with gene expression levels. The importance of studying 5-methylcytosine and 5-hydroxymethylcytosine sep. for their biol. roles will become clearer when more efficient methods to distinguish them are available.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3MXhs1KktrzP&md5=aa3db24a99915dcb230fb5f340a2d778
    14. 14
      Liu, B.; Fang, L.; Long, R.; Lan, X.; Chou, K.-C. iEnhancer-2L: a two-layer predictor for identifying enhancers and their strength by pseudo k-tuple nucleotide composition. Bioinformatics 2016, 32, 362– 369,  DOI: 10.1093/bioinformatics/btv604
      There is no corresponding record for this reference.
    15. 15
      Jia, C.; He, W. EnhancerPred: a predictor for discovering enhancers based on the combination and selection of multiple features. Sci. Rep. 2016, 6, 38741,  DOI: 10.1038/srep38741
      There is no corresponding record for this reference.
    16. 16
      Lim, D. Y.; Khanal, J.; Tayara, H.; Chong, K. T. iEnhancer-RF: identifying enhancers and their strength by enhanced feature representation using random forest. Chemom. Intell. Lab. Syst. 2021, 212, 104284,  DOI: 10.1016/j.chemolab.2021.104284
      There is no corresponding record for this reference.
    17. 17
      Liu, B.; Li, K.; Huang, D.-S.; Chou, K.-C. iEnhancer-EL identifying enhancers and their strength with ensemble learning approach. Bioinformatics 2018, 34, 3835– 3842,  DOI: 10.1093/bioinformatics/bty458
      There is no corresponding record for this reference.
    18. 18
      Wang, C.; Zou, Q.; Ju, Y.; Shi, H. Enhancer-FRL: improved and robust identification of enhancers and their activities using feature representation learning. IEEE/ACM Trans. Comput. Biol. Bioinform. 2023, 20, 967– 975,  DOI: 10.1109/TCBB.2022.3204365
      There is no corresponding record for this reference.
    19. 19
      Gill, M.; Ahmed, S.; Kabir, M.; Hayat, M. A novel predictor for the analysis and prediction of enhancers and their strength via multi-view features and deep forest. Information 2023, 14, 636,  DOI: 10.3390/info14120636
      There is no corresponding record for this reference.
    20. 20
      Nguyen, Q. H.; Nguyen-Vo, T.-H.; Le, N. Q. K.; Do, T. T.; Rahardja, S.; Nguyen, B. P. iEnhancer-ECNN: identifying enhancers and their strength using ensembles of convolutional neural networks. BMC Genom. 2019, 20, 951,  DOI: 10.1186/s12864-019-6336-3
      There is no corresponding record for this reference.
    21. 21
      Le, N. Q. K.; Ho, Q.-T.; Nguyen, T.-T.-D.; Ou, Y.-Y. A transformer architecture based on BERT and 2D convolutional neural network to identify DNA enhancers from sequence information. Briefings Bioinf. 2021, 22, bbab005,  DOI: 10.1093/bib/bbab005
      There is no corresponding record for this reference.
    22. 22
      Niu, K.; Luo, X.; Zhang, S.; Teng, Z.; Zhang, T.; Zhao, Y. iEnhancer-EBLSTM: identifying enhancers and strengths by ensembles of bidirectional long short-term memory. Front. Genet. 2021, 12, 665498,  DOI: 10.3389/fgene.2021.665498
      There is no corresponding record for this reference.
    23. 23
      Cai, L.; Ren, X.; Fu, X.; Peng, L.; Gao, M.; Zeng, X. iEnhancer-XG: interpretable sequence-based enhancers and their strength predictor. Bioinformatics 2021, 37, 1060– 1067,  DOI: 10.1093/bioinformatics/btaa914
      There is no corresponding record for this reference.
    24. 24
      Yang, H.; Wang, S.; Xia, X. iEnhancer-RD: identification of enhancers and their strength using RKPK features and deep neural networks. Anal. Biochem. 2021, 630, 114318,  DOI: 10.1016/j.ab.2021.114318
      There is no corresponding record for this reference.
    25. 25
      Mu, X.; Wang, Y.; Duan, M.; Liu, S.; Li, F.; Wang, X.; Zhang, K.; Huang, L.; Zhou, F. A novel position-specific encoding algorithm (SeqPose) of nucleotide sequences and its application for detecting enhancers. Int. J. Mol. Sci. 2021, 22, 3079,  DOI: 10.3390/ijms22063079
      There is no corresponding record for this reference.
    26. 26
      Luo, H.; Chen, C.; Shan, W.; Ding, P.; Luo, L. iEnhancer-BERT: a novel transfer learning architecture based on DNA-Language model for identifying enhancers and their strength. In International Conference on Intelligent Computing , 2022; pp 153– 165. DOI: 10.1007/978-3-031-13829-4_13 .
      There is no corresponding record for this reference.
    27. 27
      Li, J.; Wu, Z.; Lin, W.; Luo, J.; Zhang, J.; Chen, Q.; Chen, J. iEnhancer-ELM: improve enhancer identification by extracting position-related multiscale contextual information based on enhancer language models. Bioinform. Adv. 2023, 3, vbad043,  DOI: 10.1093/bioadv/vbad043
      There is no corresponding record for this reference.
    28. 28
      Wang, W.; Wu, Q.; Li, C. iEnhancer-DCSA: identifying enhancers via dual-scale convolution and spatial attention. BMC Genom. 2023, 24, 393,  DOI: 10.1186/s12864-023-09468-1
      There is no corresponding record for this reference.
    29. 29
      Wu, H.; Liu, M.; Zhang, P.; Zhang, H. iEnhancer-SKNN: a stacking ensemble learning-based method for enhancer identification and classification using sequence information. Briefings Funct. Genomics 2023, 22, 302– 311,  DOI: 10.1093/bfgp/elac057
      There is no corresponding record for this reference.
    30. 30
      Ng, P. dna2vec: consistent vector representations of variable-length k-mers. arXiv 2017, arXiv:1701.06279
      There is no corresponding record for this reference.
    31. 31
      Iuchi, H.; Matsutani, T.; Yamada, K.; Iwano, N.; Sumi, S.; Hosoda, S.; Zhao, S.; Fukunaga, T.; Hamada, M. Representation learning applications in biological sequence analysis. Comput. Struct. Biotechnol. J. 2021, 19, 3198– 3208,  DOI: 10.1016/j.csbj.2021.05.039
      31
      Representation learning applications in biological sequence analysis
      Iuchi, Hitoshi; Matsutani, Taro; Yamada, Keisuke; Iwano, Natsuki; Sumi, Shunsuke; Hosoda, Shion; Zhao, Shitao; Fukunaga, Tsukasa; Hamada, Michiaki
      Computational and Structural Biotechnology Journal (2021), 19 (), 3198-3208CODEN: CSBJAC; ISSN:2001-0370. (Elsevier B.V.)
      A review. Although remarkable advances have been reported in high-throughput sequencing, the ability to aptly analyze a substantial amt. of rapidly generated biol. (DNA/RNA/protein) sequencing data remains a crit. hurdle. To tackle this issue, the application of natural language processing (NLP) to biol. sequence anal. has received increased attention. In this method, biol. sequences are regarded as sentences while the single nucleic acids/amino acids or k-mers in these sequences represent the words. Embedding is an essential step in NLP, which performs the conversion of these words into vectors. Specifically, representation learning is an approach used for this transformation process, which can be applied to biol. sequences. Vectorized biol. sequences can then be applied for function and structure estn., or as input for other probabilistic models. Considering the importance and growing trend for the application of representation learning to biol. research, in the present study, we have reviewed the existing knowledge in representation learning for biol. sequence anal.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXitFKisrjL&md5=3fc5035b7e5c9ac8df53cc3b7f4d538c
    32. 32
      Wen, J.; Liu, Y.; Shi, Y.; Huang, H.; Deng, B.; Xiao, X. A classification model for lncRNA and mRNA based on k-mers and a convolutional neural network. BMC Bioinf. 2019, 20, 469,  DOI: 10.1186/s12859-019-3039-3
      There is no corresponding record for this reference.
    33. 33
      Li, W.; Guo, Y.; Wang, B.; Yang, B. Learning spatiotemporal embedding with gated convolutional recurrent networks for translation initiation site prediction. Pattern Recogn. 2023, 136, 109234,  DOI: 10.1016/j.patcog.2022.109234
      There is no corresponding record for this reference.
    34. 34
      Jeffrey, H. J. Chaos game representation of gene structure. Nucleic Acids Res. 1990, 18, 2163– 2170,  DOI: 10.1093/nar/18.8.2163
      34
      Chaos game representation of gene structure
      Jeffrey, H. Joel
      Nucleic Acids Research (1990), 18 (8), 2163-70CODEN: NARHAD; ISSN:0305-1048.
      This paper presents a new method for representing DNA sequences. It permits the representation and investigation of patterns in sequences, visually revealing previously unknown structures. Based on a technique from chaotic dynamics, the method produces a picture of a gene sequence which displays both local and global patterns. The pictures have a complex structure which varies depending on the sequence. The method is termed Chaos Game Representation (CGR). CGR raises a new set of questions about the structure of DNA sequences, and is a new tool for investigating gene structure.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADyaK3cXksFeltrk%253D&md5=2832019a98de2fa2987a2f70e5f32340
    35. 35
      Löchel, H. F.; Eger, D.; Sperlea, T.; Heider, D. Deep learning on chaos game representation for proteins. Bioinformatics 2020, 36, 272– 279,  DOI: 10.1093/bioinformatics/btz493
      There is no corresponding record for this reference.
    36. 36
      LaLonde, R.; Bagci, U. Capsules for object segmentation. arXiv 2018, arXiv:1804.04241
      There is no corresponding record for this reference.
    37. 37
      Dong, Z.; Lin, S. Research on image classification based on capsnet. In 2019 IEEE 4th Advanced Information Technology, Electronic and Automation Control Conference (IAEAC) , 2019; pp 1023– 1026.
      There is no corresponding record for this reference.
    38. 38
      Guo, Y.; Zhou, D.; Ruan, X.; Cao, J. Variational gated autoencoder-based feature extraction model for inferring disease-miRNA associations based on multiview features. Neural Network. 2023, 165, 491– 505,  DOI: 10.1016/j.neunet.2023.05.052
      There is no corresponding record for this reference.
    39. 39
      Guo, Y.; Zhou, D.; Li, P.; Li, C.; Cao, J. Context-aware poly (a) signal prediction model via deep spatial–temporal neural networks. IEEE Trans. Neural Netw. Learn. Syst. 2024, 35, 8241– 8253,  DOI: 10.1109/tnnls.2022.3226301
      There is no corresponding record for this reference.
    40. 40
      Wang, X.; Guan, Z.; Qian, W.; Cao, J.; Wang, C.; Ma, R. STFuse: infrared and visible image fusion via semisupervised transfer learning. IEEE Trans. Neural Netw. Learn. Syst. 2024, 1– 14,  DOI: 10.1109/tnnls.2023.3328060
      There is no corresponding record for this reference.
    41. 41
      Sabour, S.; Frosst, N.; Hinton, G. E. Dynamic routing between capsules. In Advances in Neural Information Processing Systems , 2017; Vol. 30.
      There is no corresponding record for this reference.
    42. 42
      Yao, L.; Pang, Y.; Wan, J.; Chung, C.-R.; Yu, J.; Guan, J.; Leung, C.; Chiang, Y.-C.; Lee, T.-Y. ABPCaps: a novel capsule network-based method for the prediction of antibacterial peptides. Appl. Sci. 2023, 13, 6965,  DOI: 10.3390/app13126965
      There is no corresponding record for this reference.
    43. 43
      Huang, Y.; Huang, H.-Y.; Chen, Y.; Lin, Y.-C.-D.; Yao, L.; Lin, T.; Leng, J.; Chang, Y.; Zhang, Y.; Zhu, Z. A robust drug–target interaction prediction framework with capsule network and transfer learning. Int. J. Mol. Sci. 2023, 24, 14061,  DOI: 10.3390/ijms241814061
      There is no corresponding record for this reference.
    44. 44
      Wang, D.; Liang, Y.; Xu, D. Capsule network for protein post-translational modification site prediction. Bioinformatics 2019, 35, 2386– 2394,  DOI: 10.1093/bioinformatics/bty977
      44
      Capsule network for protein post-translational modification site prediction
      Wang, Duolin; Liang, Yanchun; Xu, Dong
      Bioinformatics (2019), 35 (14), 2386-2394CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
      Motivation: Computational methods for protein post-translational modification (PTM) site prediction provide a useful approach for studying protein functions. The prediction accuracy of the existing methods has significant room for improvement. A recent deep-learning architecture, Capsule Network (CapsNet), which can characterize the internal hierarchical representation of input data, presents a great opportunity to solve this problem, esp. using small training data. Results: We proposed a CapsNet for predicting protein PTM sites, including phosphorylation, Nlinked glycosylation, N6-acetyllysine, methyl-arginine, S-palmitoyl-cysteine, pyrrolidonecarboxylic- acid and SUMOylation sites. The CapsNet outperformed the baseline convolutional neural network architecture MusiteDeep and other well-known tools in most cases and provided promising results for practical use, esp. in learning from small training data. The capsule length also gives an accurate est. for the confidence of the PTM prediction. We further demonstrated that the internal capsule features could be trained as a motif detector of phosphorylation sites when no kinase-specific phosphorylation labels were provided. In addn., CapsNet generates robust representations that have strong discriminant power in distinguishing kinase substrates from different kinase families. Our study sheds some light on the recognition mechanism of PTMs and applications of CapsNet on other bioinformatic problems.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXjvF2lsbc%253D&md5=824e5d94dc1f2ddb865ac97e1fc57be1
    45. 45
      Khanal, J.; Tayara, H.; Zou, Q.; To Chong, K. DeepCap-Kcr: accurate identification and investigation of protein lysine crotonylation sites based on capsule network. Briefings Bioinf. 2022, 23, bbab492,  DOI: 10.1093/bib/bbab492
      There is no corresponding record for this reference.
    46. 46
      Shang, J.; Peng, C.; Tang, X.; Sun, Y. PhaVIP: Phage VIrion protein classification based on chaos game representation and vision transformer. arXiv 2023, arXiv:2301.12422
      There is no corresponding record for this reference.
    47. 47
      Löchel, H. F.; Heider, D. Chaos game representation and its applications in bioinformatics. Comput. Struct. Biotechnol. J. 2021, 19, 6263– 6271,  DOI: 10.1016/j.csbj.2021.11.008
      47
      Chaos game representation and its applications in bioinformatics
      Lochel Hannah Franziska; Heider Dominik
      Computational and structural biotechnology journal (2021), 19 (), 6263-6271 ISSN:2001-0370.
      Chaos game representation (CGR), a milestone in graphical bioinformatics, has become a powerful tool regarding alignment-free sequence comparison and feature encoding for machine learning. The algorithm maps a sequence to 2-dimensional space, while an extension of the CGR, the so-called frequency matrix representation (FCGR), transforms sequences of different lengths into equal-sized images or matrices. The CGR is a generalized Markov chain and includes various properties, which allow a unique representation of a sequence. Therefore, it has a broad spectrum of applications in bioinformatics, such as sequence comparison and phylogenetic analysis and as an encoding of sequences for machine learning. This review introduces the construction of CGRs and FCGRs, their applications on DNA and proteins, and gives an overview of recent applications and progress in bioinformatics.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BB2cblt1Ojsg%253D%253D&md5=8e8167b2552e037e8c046d3dcf24e119
    48. 48
      Kingma, D. P.; Ba, J. Adam: a method for stochastic optimization. arXiv 2014, arXiv:1412.6980
      There is no corresponding record for this reference.
    49. 49
      Paszke, A.; Gross, S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T.; Lin, Z.; Gimelshein, N.; Antiga, L.; Pytorch: an imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems , 2019; Vol. 32.
      There is no corresponding record for this reference.
    50. 50
      Kishk, A.; Elzizy, A.; Galal, D.; Razek, E. A.; Fawzy, E.; Ahmed, G.; Gawish, M.; Hamad, S.; El-Hadidi, M. A hybrid machine learning approach for the phenotypic classification of metagenomic colon cancer reads based on kmer frequency and biomarker profiling. In 2018 9th Cairo International Biomedical Engineering Conference (CIBEC) , 2018; pp 118– 121.
      There is no corresponding record for this reference.
    51. 51
      Yin, B.; Balvert, M.; Zambrano, D.; Schönhuth, A.; Bohte, S. An image representation based convolutional network for DNA classification. arXiv 2018, arXiv:1806.04931
      There is no corresponding record for this reference.

  • Supporting Information

    Supporting Information

    The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jcim.4c00546.

    • GC content distribution of the data sets, FCGR images for two example sequences and PR curves for the ablation experiment, hyperparameters of CapsEnhancer, a case study, and description of the dynamic routing algorithm (PDF)


    Terms & Conditions

    Most electronic Supporting Information files are available without a subscription to ACS Web Editions. Such files may be downloaded by article for research use (if there is a public use license linked to the relevant article, that license may permit other uses). Permission may be obtained from ACS for other uses through requests via the RightsLink permission system: http://pubs.acs.org/page/copyright/permissions.html.