Characterizing DNA recognition preferences of transcription factors using global couplings and high-throughput sequencing

Abstract

DNA–transcription factor (TF) interactions are essential for gene regulation. Fully characterizing TF recognition specificities and identifying their genomic binding targets are important to understand TF function and regulatory networks. Recently, high-throughput sequencing technology HT-SELEX (high-throughput systematic evolution of ligands by exponential enrichment) has been used to measure hundreds of TFs, providing massive datasets that comprise TF binding preferences. However, there is a need to develop comprehensive computational modeling to fully extract and characterize critical TF binding preferences and fail to distinguish genome-wide binding targets. In this study, we developed a global pairwise model called DCA-Scapes trained with experimental HT-SELEX data. Our approach uncovered high-resolution TF recognition specificity landscapes, enabled the prediction of in vivo binding sequences, and was validated with ChIP-seq (ChIP sequencing) data. In addition, the DCA-Scapes model was utilized to refine the locations of binding regions and accurately identify the binding sites within the ChIP-seq enriched peaks. Moreover, we extended our model to cover the entire human genome, uncovering potential TF target sites that exhibit tissue-specific TF recognition across various cellular environments.

Graphical Abstract

Graphical Abstract.

Introduction

The interactions between transcription factors (TFs) and DNA play key roles in transcription regulation, which directly interprets the genome and guides its expression [1]. TFs can bind to different DNA sequences with various binding specificities [2], interact with specific DNA target sequences in the genome, and consequently activate or repress transcription of nearby genes [1]. Characterizing DNA recognition properties of TFs and identifying TF binding sites can provide a gateway to further understanding TF function, reveal the potential gene regulation networks, and decode functional properties of the genome.

Over the last decade, diverse methods have been developed to experimentally derive and measure TF–DNA binding preferences from in vivo and in vitro perspectives. The most prevalent in vivo method is ChIP sequencing (ChIP-seq) [3], which combines chromatin immunoprecipitation and massively parallel sequencing. ChIP-based assays allow measurement of genomic DNA sequences that are bound to TFs, but the methodology is unsuitable for many lesser studied TFs. ChIP-seq-based experiments are highly dependent on antibody quality; however, many TFs lack available ChIP-grade antibodies [4]. In addition, ChIP-seq data have broad signal footprints with hundreds of nucleotides in length, which makes it hard to determine the exact TF binding site [5, 6]. ChIP-seq measurements are also influenced by environmental biases in different cell lines or tissues, including protein–protein interactions of other proteins with TFs and TF accessibility to genomic sequences [1,7]. To fully understand the patterns of TF–DNA recognition, many tissue samples are required to generate a comprehensive dataset, making it prohibitive for some rare cell types.

In vitro experiments were established to provide a more general unbiased estimation for TF binding preferences. Starting with low-throughput sequencing-based methods, such as systematic evolution of ligands by exponential enrichment (SELEX) [8], electrophoretic mobility shift assay [9], and nuclease footprinting [10], to high-throughput methods, including protein binding microarrays [11] and high-throughput SELEX (HT-SELEX) [12], numerous attempts were initiated to produce DNA data that systematically characterize TF specificities. High-throughput sequencing methods allow the processing of protein binding measurement of thousands to millions of DNA sequences and provide massive datasets that comprehensively comprise TF binding preferences. Due to the vast amount of data, computational modeling was required to accurately extract and infer the binding specificities from these high-throughput experimental data.

Position weight matrix (PWM) [13, 14] is the most used model to represent the DNA-binding preferences of TFs. PWM is a matrix derived from position frequency matrices [15], with a probability score for each nucleotide at each position. These probabilities can be added to estimate the overall binding affinity of DNA elements. All the most commonly used TF-binding profile databases, including JASPAR [16, 17], TRANSFAC [18], and CIS-BP [19], collect sequencing data and use PWM-based methods to generate and store binding motif patterns. However, PWM motifs were built under the assumption that nucleotide positions are independent. Other features like DNA shape that might be required for TF recognition specificity [20] cannot be captured by PWM. In addition, the PWM motifs only reflect one or several strong recognition preference patterns, while neglecting weak preferences, which reduces the accuracy in identifying genomic binding sites. All these complexities justify a more comprehensive computational model for studying TF interactions.

We have built a global pairwise DCA-Scapes model in our previous work [21] that captured the sequence and structure specificity requirements of protein–RNA interactions from experimental data generated by in vitro selection and high-throughput sequencing of RNAs (SEQRS) [22]. We have successfully characterized core features of protein recognition, identified binding elements from their genomic context, and accurately predicted binding affinities for RNA variants. The parallels between HT-SELEX and SEQRS provide the possibility to utilize our statistical model to infer TF–DNA interaction preferences from HT-SELEX data. Here, we took the most comprehensive HT-SELEX dataset of mammalian TFs [23, 24], applied our global modeling to uncover the TF–DNA interaction landscapes, predicted genomic binding sites, and incorporated the genomic binding targets with ChIP-seq data to reveal tissue-specific TF recognition preferences.

Materials and methods

Data processing

The HT-SELEX sequencing data were downloaded from the European Nucleotide Archive (ENA; http://www.ebi.ac.uk/ena) under study identifier PRJEB14744, which increased the sequencing depth [24] of previous existing HT‐SELEX data [23]. This dataset includes 548 experiments covering 410 human and mouse proteins and was filtered by rigorous quality control criteria [24], leaving 215 high-quality TF data with sufficient library complexity and read counts. Here, we focused on the HT-SELEX experiments, beginning with the 20 oligonucleotide DNA, that include 147 human TFs and 37 mouse TFs (Supplementary Table S1).

DCA-Scapes and Hamiltonian scores

As described in our previous study, we have built a comprehensive DCA-Scapes model that adapted coevolutionary modeling via direct coupling analysis (DCA) [25] to analyze protein–RNA recognition from in vitro selection and high-throughput sequencing of RNAs (SEQRS) [21]. Here, rather than studying protein–RNA interactions, we used our model to study TFs and DNA interactions. For each TF protein, our model was trained using the reads from the last cycle of selection (round 4) of the HT-SELEX experiment, with the initial nonselected sequence pool as a background control. These sequence arrays were translated to numeric arrays using the following notation:

(1)

where FR represents the last round sequences array, BG is the background control nonselected sequence array, L is the length of oligonucleotides (L = 20), and M is the number of sequences in each sequence array. The four types of nucleotides (A, C, G, T) were translated into consecutive numbers 1, 2, 3, and 4.

For the training data (FR) and background data (BG), we separately infer the joint probability distribution of 20-mer DNA sequences with two types of parameters: pairwise couplings .

(2)

During the inferring process, to reduce the freedom from independent constraints, the final couplings and local fields are measured as an average of the four-nucleotide gauged state [21]. These two parameters for the training and background data are collectively interpreted as the fitness function score, known as the Hamiltonian score. For a given DNA sequence x, the Hamiltonian score aims to quantitatively predict the likelihood of it being a target for the TFs. A more negative Hamiltonian score indicates a greater likelihood of favorable TF recognition.

(3)

In vivo sequence prediction validation

To test the accuracy of our model in predicting in vivo binding sites, we downloaded the ChIP-seq data from the ENCODE project [26] for all TFs included in our study. We found 66 TFs with at least one ChIP-seq experiment, and 240 ChIP-seq experiments in total. The hg38 (GRCh38) genomic coordinates of the top 500 peaks in each ChIP-seq experiment were used to extract the sequences from the human genome. Same-length genomic regions in both upstream and downstream were selected as the negative controls. Upstream and downstream sequences of a peak that overlapped with another peak from the same TF were removed from negative controls to preserve distinct classification of positives and negatives. To distinguish the peaks from two nearby sequence regions, our Hamiltonian scores were calculated with the sliding window size of 20. The three most negative Hamiltonian scores were used to represent the TF recognition specificity of each region. The scores of peaks and two negative regions were used to calculate a receiver operating characteristic (ROC) curve. The prediction performance was evaluated as the area under the ROC curve (AUC) (Supplementary Table S2).

Null model’s construction with random sequences

To further understand binding strength as predicted from Hamiltonian score and apply Hamiltonian scores in identifying the binding targets from human genome sequences, we created a random sequence dataset with 1 million random 20-mer DNA sequences. These random sequences contain a similar nucleotide distribution to the human genome (GRCh38). For each TF, a null model was constructed by calculating the Hamiltonian scores for the random sequences based on the parameters: couplings that were learned from the HT-SELEX sequencing array data. With this null model score distribution, the P-value for a given sequence could be estimated. Since nonfunctional random sequences are expected to have higher Hamiltonian score than functional one, we use a one-tailed z-test to evaluate significance; the average Hamiltonian score and standard deviation along the random 20-mers are included in Supplementary Table S3.

Binding site discovery in the human genome

We further applied our model to discover binding sites in the human genome. We obtained the genome sequence hg38 (GRCh38) from UCSC Genome Browser [27] and quantified the binding affinity for each position in the whole genome with Hamiltonian score. With a sliding window size of 20, from the first nucleotide to the end of the genome, we built the genome sequence library with 3 209 277 460 20-mer sequences. To reduce memory usage and increase the speed of computing, we split this human genome library into 500 pieces, each subset including around 6 418 600 20-mer sequences (the last subset contains 6 396 060 20-mer sequences). We parallelly computed the Hamiltonian scores for all 20-mer sequences in each subset, and only the sequences with scores better than the null model score range were kept. All selected sequences from subsets were sorted based on Hamiltonian score to identify the most likely binding site within the genome.

Visualization of the structural complex of MAX TF bound to a double-stranded DNA

To visualize the interaction between the MAX TF and the selected DNA sequence, we simulated the interaction between the MAX TF and a 20-bp DNA segment with the highest-ranking HT-SELEX DNA sequence (DNA sequence: GATACTGACAAGCCGTGAAC). To accurately capture the interactions between DNA and proteins, as well as sufficient sampling of their conformations, we adopted a residue-resolution coarse-grained protein–DNA model. In this model, each protein residue was represented by its C atom, and protein–protein interactions were modeled using the structure-based model [28, 29]. For DNA, each nucleotide was modeled by three coarse-grained sites, representing phosphate, sugar, and base. Interactions with the DNA were modeled using the 3SPN.2C model [30]. The protein–protein interactions were scaled by a factor of 3.75 to avoid protein unfolding at 300 K. The protein–DNA interactions were simulated with Debye–Hückel screening interaction [31] at a monovalent ion concentration of 100 mM. This model has been validated to characterize the higher-order structures of chromatin and protein–DNA interactions accurately [32–35].

To ensure that our sampling was sufficiently equilibrated, we employed umbrella sampling [28,36] to sample the association of MAX with DNA, using the center-of-mass distance between MAX and DNA as the umbrella coordinate. We then used the weighted histogram analysis method algorithm [28,37] to construct the reweighted free energy profile. From this profile, we selected the most populated structural ensemble with a MAX–DNA center-of-mass distance ranging from 28.0 to 34.0 Å. Those selected structures were then clustered based on the root-mean-square deviation between all pairs of structures, following structural alignment. The clustering was performed using the single-linkage algorithm implemented in the Gromacs software [38], with a distance cutoff of 6 Å.

Results

DCA-Scapes TF–DNA interaction model construction using HT-SELEX datasets

We selected 184 high-quality HT-SELEX datasets [23, 24] (147 human TFs and 37 mouse TFs) to build the integrated TF–DNA interaction model DCA-Scapes (Fig. 1). For each TF protein, we trained our model with the TF-bound DNA sequences in round 4 of the HT-SELEX experiment and the initial nonselected sequence pool in round 0 as the background set. With DNA sequences selected from HT-SELEX, our model first estimated a joint probability distribution of these sequences with two types of recognition parameters: pairwise couplings . For any given DNA sequence, these recognition parameters can be collectively interpreted as the Hamiltonian score (see the “Materials and methods” section), which quantitatively predicts how likely this DNA sequence binds to a TF. A random sequence distribution null model that included 1 million random sequences with similar genomic nucleotide bias was generated to further quantify the probability of interaction (Supplementary Fig. S1). Here, we applied this global comprehensive model in exploring the TF–DNA interaction with in vivo binding site prediction and validated our prediction with ChIP-seq experiments.

Figure 1.

Schematic representation of DCA-Scapes modeling in TF–DNA interaction. (A) A HT-SELEX dataset including 184 TFs was used as the input of DCA-Scapes. For each TF data, a global integrated DNA interaction model was built from the sequence data that allowed a high-resolution recognition of binding patterns through pairwise couplings and local fields . The binding likelihood of a sequence can be quantified through a Hamiltonian score. This score was used for (B) in vivo binding sequence prediction and validation with ChIP-seq data and to (C) identify the TF binding sites in the entire human genome.

In vivo sequence recognition prediction with ChIP-seq data

With the model trained from in vitro experimental sequence data, we further tested our model for in vivo binding site prediction using ChIP-seq data. We selected 240 ChIP-seq experiments that represent 66 TFs (Fig. 2A). Of the 66 TFs, 36 TFs have one ChIP-seq experiment, while the others have at least two ChIP-seq experiments. Three TFs have more than 20 ChIP-seq experiments (MAX:22; CEBPB:28; NR3C1:29). For each ChIP-seq experiment, we predicted the binding of the top 500 most enriched peak sequences and used sequences of the same length from their unbound upstream and downstream regions as sequence feature controls (Fig. 2B). To evaluate the performance, we plotted ROC curve and computed the AUC score for each ChIP-seq experiment. The average AUC for 22 MAX protein ChIP-seq experiments is 0.83 (Fig. 2C), 0.86 for CEBPB, and 0.72 for NR3C1 (Supplementary Fig. S2). The overall average AUC across 240 ChIP-seq experiments is 0.705 ± 0.163 (Fig. 2D), indicating the model’s capability to distinguish in vivo TF binding sites from genomic background. The observed variability in performance likely reflects differences in ChIP-seq data quality and incomplete coverage. To ensure a comprehensive and unbiased evaluation, all available datasets were incorporated into the analysis. Collectively, these results demonstrate that the model performs robustly across a broad range of TFs and experimental conditions.

Figure 2.

Binding prediction evaluation with in vivo ChIP-seq experiment data. (A) The frequency distribution of numbers of ChIP-seq experiments (x-axis). Three TFs contain more than 20 ChIP-seq experiment datasets, and 36 TFs have one ChIP-seq experiment. (B) Top 500 most enriched peaks in ChIP-seq used as input of DCA-Scapes. Sequences were extracted from the top 500 enriched peaks with the same length unbound upstream and downstream sequence as the control. (C) ROC for MAX protein binding prediction in 22 ChIP-seq experiments. The AUC was calculated for each experiment. (D) Average AUC for all the 240 ChIP-seq experiments (mean = 0.7045, SD = ±0.1630).

ChIP-seq has been widely applied to inferring genomic protein–DNA interacting sites. The recognition regions could be estimated by identifying the peak areas that are enriched with bound short sequence reads. However, the peak regions could be hundreds of nucleotides, and determining the exact binding sites within these regions is challenging. Here, we developed the DCA-Scapes model to refine the binding region further and obtain a precise identification of binding sites inside the ChIP-seq enriched peaks. We first selected the two most enriched peaks in CEBPB (ENCODE file ID: ENCFF064WDQ; Fig. 3A) and MAX experiments (ENCODE file ID: ENCFF266YHW; Fig. 3B) and predicted the binding specificities with Hamiltonian scores for every position in the peak regions as well as upstream and downstream controls. Utilizing random score distributions for MAX and CEBPB as null model (Supplementary Fig. S2), sequences with significantly more favorable Hamiltonian scores were illustrated in orange color (Fig. 3). Interestingly, for both MAX and CEBPB proteins, two sequence clusters were found to contain the lowest Hamiltonian scores. All these most favorable Hamiltonian score sequences are in the peak regions instead of control regions, suggesting the presence of TF recognition target sites within those peak sequences.

Figure 3.

Comparison between DCA-Scapes model predictions and ChIP-seq enriched peaks. For each genomic region, the Hamiltonian scores predicted by the DCA-Scapes model are plotted along the chromosomal coordinates. Random-like binding scores are indicative of non-specific or background regions. In contrast, sequences with lower Hamiltonian scores suggest more favorable and specific binding potential; these are highlighted and shown in orange. ChIP-seq peak regions, derived from experimental data, are shown as orange rectangles and aligned with their corresponding genomic positions for direct visual comparison. (A) Example for the CEBPB TF. (B) Example for the MAX TF.

To further test and understand this observation, we illustrated all related peaks from the ChIP-seq experiments in each TF as orange rectangles under the Hamiltonian score plot where their positions and lengths are plotted according to their genomic coordinates. These enriched peaks from different tissues/experiments vary in length and cover similar but not identical chromosome locations; for instance, peaks in ChIP-seq experiments like ENCFF592ZPR (Fig. 3A) and ENCFF493DCP (Fig. 3B) are very short and less than half the length as other peaks. Variability in ChIP-seq peak coverage can strongly impact AUC values. As illustrated in Fig. 3B, the TF MAX appears to bind at two distinct sites within a region. Some ChIP-seq experiments capture peaks covering both sites, while others detect only one. If validation relies on a dataset that captures only a single site, the other may be incorrectly labeled as a false positive. In reality, both sites represent true binding events, but incomplete or inconsistent ChIP-seq coverage can obscure this. Interestingly, our most favorable Hamiltonian score sequence clusters agree well with the enriched peak coverages of these ChIP-seq experiments, indicating the precise binding site was identified. We expanded our score calculation to the top 5 peaks and observed the same trends (Supplementary Fig. S3 for CEBPB, Supplementary Fig. S4 for MAX). These results highlight the ability of our model for refining the binding region in ChIP-seq experiments and identify the exact TF binding site.

Genomic recognition target prediction

With the ability to accurately predict whether in vivo sequences bind to TF, we expanded our search to the entire human genome. For each human TF protein, we predicted the recognition likelihood as Hamiltonian scores in all human genome sequences and the top 50 best-scoring binding targets were reported. These TFs are separated into two groups: validated TF models (validated using ChIP-seq data with high accuracy: average AUC > 0.7, N = 18) and not-validated TF models (TFs do not have ChIP-seq data to compare, N = 81, or get low AUC score, N = 48).

To further understand our top TF binding targets and their binding specificities in different biological environments, we compared our top 50 targets with ChIP-seq enriched peaks in different tissues/cell lines. In most of the ChIP-seq experiments of CEBPB, most of our top targets were identified to bind to TF (Fig. 4A). Interestingly, we also observed some top targets exhibit tissue-specific recognition, only being able to bind to TFs in certain types of tissues. For instance, our second target that is related to the PIGL (phosphatidylinositol glycan anchor biosynthesis class L) gene only showed TF recognition specificities in lung fibroblast isolated IMR-90 cells. The PIGL gene is not very well studied yet, but its glycosylated function (GPI anchor) has been identified to be important for normal development of lung tissues [39] and fibroblast migration [40]. The 32nd target site that encoded the regulation target of GXYLT1 (glucoside xylosyltransferase 1) showed TF recognition specificities in both lung carcinoma epithelial cells (A549) and lung fibroblast isolated IMR-90 cells. GXYLT1 has been shown to be an immunogenic antigen of lung cancer [41, 42]. These observations support the CEBPB protein having an important role in lung development [43] through regulating lung-specific gene expression. The MAX protein top targets also show strong tissue specificity of TF recognition (Fig. 4B). Only two target sites (3rd, 17th) were consistently found to bind to TF MAX across various tissues. In contrast, a greater number of targets were found to bind to TF only in ChIP-seq experiment environments related to leukemia (K562 and NB4) cell lines. Some of these target genes have already been identified as key markers in leukemia. The first genomic target of MAX is an encoded lncRNA LINC00683, which is involved in many cancer-related pathways such as the Wnt pathway [44], and the under-expression of LINC00649 has been found to be an adverse prognostic marker in acute myeloid leukemia [45]. Other top targets, including the 3^rd target NPM1 [46, 47] gene, 7^th target PRKCB [48] gene, and 10^th target CRPPA [49] gene, are essential for leukemia development. The strong interactions between the MAX protein and these genes as predicted from our model indicate a possible MAX regulation pathway in leukemia and some potential therapeutic strategy through MAX–DNA interaction.

Figure 4.

TF tissue-specific recognition in top 50 genomic recognition targets. The top 50 most favorable Hamiltonian score targets in the human genome (x-axis) were compared to ChIP-seq enriched peaks. In ChIP-seq experiments with specific tissues (y-axis), genomic targets that were found to be enriched are highlighted. Low-quality ChIP-seq experiments of tissues labeled by ENCODE as insufficient read depth are excluded [26] (Supplementary Fig. S5). Two protein examples are shown as (A) CEBPB protein and (B) MAX protein.

High-resolution TF recognition specificity patterns DCA-Scapes

In addition to predicting TF recognition, a significant advantage of DCA-Scapes is that it enables precise, high-resolution identification of TF recognition patterns. Employing the MAX protein as an example, by scoring the HT-SELEX sequences and using the null model as control (Fig. 5A), one sequence with the most favorable recognition Hamiltonian score (P-value = 2.7223e−39) was predicted to have strong binding specificity to MAX. In previous studies, the MAX homodimers were found to recognize the core sequence to CACGTG [50]. Two repeats of this core motif were identified in our most favorable DNA element. More interestingly, our coupling landscapes demonstrate two types of strong residue interactions (Fig. 5C). One is the short-range interaction among nearby residues, one and two positions away, which provides the sequence specificities like the core motif for protein recognition, and the other is a strong long-range interaction between regions in residues 4–9 CACGTG and 15–20 CACGTG. From the protein–DNA complex structure, we have identified that the presence of these two well-established MAX binding motifs (highlighted in yellow) is essential for creating the binding site for the protein (Fig. 5B). This facilitates robust interactions with two MAX protein loops. These discoveries indicate that the significance of protein recognition extends beyond the motif sequence; it also encompasses the spatial requirements defined by the interactions between motif repeats for MAX protein recognition.

Figure 5.

MAX protein high-resolution recognition pattern. (A) MAX protein Hamiltonian score distribution for HT-SELEX sequencing data (left) and simulated random data (right). (B) A representative TF–DNA complex structure with the most favorable sequence Hamiltonian score simulated by a residue-resolution coarse-grained protein–DNA model. (C) The MAX–DNA interaction landscape. Pairwise coupling values for all the possible residue pairs are plotted according to the color scale. Short-range interactions between nearby positions are circled above, as well as long-range interactions between positions distant by 11 sites.

Discussion

Our study comprehensively explored the TF binding specificities from a large number of HT-SELEX experiments and generated high-resolution binding patterns that uncovered both sequence and structure requirements for TF recognition. This model was applied to refine the binding region identified by ChIP-seq enriched peaks and predict genomic binding targets. These targets reported could be applied to better understand regulation networks for TFs. In addition to the sequence in the reference genome (hg38), Hamiltonian scores can also predict the binding likelihood for any given sequence [21], including variants of the genomic targets, providing tools to analyze the recognition of TF to disease-associated genetic variants.

In this model, we included 147 human TFs and 37 mouse TFs with high-quality HT-SELEX data. According to recent studies, there are ∼1500 TFs identified in human genome [51–53], and around 941 are quantitatively identified in the mouse genome [54]. More TFs could be modeled using our methodology once the experimental tests expand. In the meantime, using the computational model to predict these untested TFs could be feasible. Co-evolutionary modeling has been used to engineer protein interactions based on protein sequences [55, 56]. A more comprehensive co-evolutionally interaction network could be built by incorporating the TF protein sequence with the HT-SELEX binding specificities data.

By integrating our TF binding predictions with ChIP-seq data, we identified clear patterns of tissue-specific TF recognition, reflecting how transcriptional regulation varies across cellular environments. For example, we identified MAX binding targets that are only enriched in leukemia cell environments. These binding targets can be further investigated to understand the MAX regulation pathway and discover some potential therapeutic strategies. TF recognition specificity differences between tissues could be influenced by the DNA accessibilities in different tissue environments or by proteins that interact with TFs, such as how the MYC protein can form a heterodimer with MAX and increase binding to the MAX core binding motif. For cell lines that have not been tested with ChIP-seq, incorporating our model with data from open chromatin identification assays such as ATAC-seq [57, 58] that determine chromatin accessibility across the genome could be used to predict the TF tissue-specific binding for some rare cell types.

We acknowledge that the current ChIP-seq datasets are limited in both cell-type diversity and experimental consistency. For instance, CEBPB datasets are largely derived from a single cell type, whereas MAX is profiled across heterogeneous tissues with fewer replicates. These limitations highlight a key strength of the DCA-Scapes model: its ability to generalize TF recognition beyond the constraints of existing data and uncover context-dependent regulatory interactions that may not yet be experimentally observed. Several predicted targets were not detected in available ChIP-seq datasets, suggesting that additional tissue-specific binding events remain to be discovered and experimentally validated.

DCA-Scapes not only recognizes exact patterns in data to give them favorable scores, e.g. pattern matching, but also assesses the possibility of a much larger set of combinations that might occur in nature. These combinations might not be observed in experimental data, and yet they comply with the coupling patterns and individual site propensities. Although we acknowledge that we will never be able to cover the complete space of recognition with such smaller experimental sample, we can expand this space much more than just the one provided by consensus sequences. Finally, the inclusion of pairwise statistics not only can provide an identification of functional motifs, but also increases the explanatory power of the model. An example of this is the discussion around the MAX recognition (Fig. 5), where the long-range interactions found on the coupling matrix provide an insight into the interaction between the TF and the nucleotide sequence, which is not available in local models.

Supplementary data

Supplementary data is available at NAR online.

Conflict of interest

None declared.

Funding

We acknowledge support from the National Institutes of Health (NIH R35GM133631 to F.M. and A.D.S.). F.M. acknowledges support from the National Science Foundation CAREER award (MCB-1943442). We also acknowledge startup funding from North Carolina State University. This research was also funded, in part, by the North Carolina State Genetics and Genomics Academy and Comparative Medicine Institute. Funding to pay the Open Access publication charges for this article was provided by National Science Foundation (MCB-1943442).

Data availability

The code used to compute DCA-Scapes has been adapted from its previous implementation and can be found at https://github.com/morcoslab/DCAscapes. Code and estimated parameters are also provided as a datafile in online supplementary material. We developed an online tool published at https://dcascapes.org to analyze input nucleotide sequences from the user against the 184 TFs in this study (Supplementary Fig. S6). Training and testing data can be accessed from (i) the European Nucleotide Archive (ENA; http://www.ebi.ac.uk/ena) under study identifier PRJEB14744, (ii) the ENCODE project https://www.encodeproject.org/about/data-access/ with a list of file IDs provided as (Supplementary Tables S3 and S4), and (iii) UCSC Genome Browser identifier hg38 (GRCh38) at https://genome.ucsc.edu/cgi-bin/hgGateway.