Discriminative motif optimization based on perceptron training

Weakly-Supervised Convolutional Neural Network Architecture for Predicting Protein-DNA Binding

Although convolutional neural networks (CNN) have outperformed conventional methods in predicting the sequence specificities of protein-DNA binding in recent years, they do not take full advantage of the intrinsic weakly-supervised information of DNA sequences that a bound sequence may contain multiple TFBS(s). Here, we propose a weakly-supervised convolutional neural network architecture (WSCNN), combining multiple-instance learning (MIL) with CNN, to further boost the performance of predicting protein-DNA binding. WSCNN first divides each DNA sequence into multiple overlapping subsequences (instances) with a sliding window, and then separately models each instance using CNN, and finally fuses the predicted scores of all instances in the same bag using four fusion methods, including Max, Average, Linear Regression, and Top-Bottom Instances. The experimental results on in vivo and in vitro datasets illustrate the performance of the proposed approach. Moreover, models built on in vitro data using WSCNN can predict in vivo protein-DNA binding with good accuracy. In addition, we give a quantitative analysis of the importance of the reverse-complement mode in predicting in vivo protein-DNA binding, and explain why not directly use advanced pooling layers to combine MIL with CNN, through a series of experiments.

Sharing DNA-binding information across structurally similar proteins enables accurate specificity determination

We are now in an era where protein-DNA interactions have been experimentally assayed for thousands of DNA-binding proteins. In order to infer DNA-binding specificities from these data, numerous sophisticated computational methods have been developed. These approaches typically infer DNA-binding specificities by considering interactions for each protein independently, ignoring related and potentially valuable interaction information across other proteins that bind DNA via the same structural domain. Here we introduce a framework for inferring DNA-binding specificities by considering protein-DNA interactions for entire groups of structurally similar proteins simultaneously. We devise both constrained optimization and label propagation algorithms for this task, each balancing observations at the individual protein level against dataset-wide consistency of interaction preferences. We test our approaches on two large, independent Cys2His2 zinc finger protein-DNA interaction datasets. We demonstrate that jointly inferring specificities within each dataset individually dramatically improves accuracy, leading to increased agreement both between these two datasets and with a fixed external standard. Overall, our results suggest that sharing protein-DNA interaction information across structurally similar proteins is a powerful means to enable accurate inference of DNA-binding specificities.

A map of direct TF-DNA interactions in the human genome

Chromatin immunoprecipitation followed by sequencing (ChIP-seq) is the most popular assay to identify genomic regions, called ChIP-seq peaks, that are bound in vivo by transcription factors (TFs). These regions are derived from direct TF-DNA interactions, indirect binding of the TF to the DNA (through a co-binding partner), nonspecific binding to the DNA, and noise/bias/artifacts. Delineating the bona fide direct TF-DNA interactions within the ChIP-seq peaks remains challenging. We developed a dedicated software, ChIP-eat, that combines computational TF binding models and ChIP-seq peaks to automatically predict direct TF-DNA interactions. Our work culminated with predicted interactions covering >4% of the human genome, obtained by uniformly processing 1983 ChIP-seq peak data sets from the ReMap database for 232 unique TFs. The predictions were a posteriori assessed using protein binding microarray and ChIP-exo data, and were predominantly found in high quality ChIP-seq peaks. The set of predicted direct TF-DNA interactions suggested that high-occupancy target regions are likely not derived from direct binding of the TFs to the DNA. Our predictions derived co-binding TFs supported by protein-protein interaction data and defined cis-regulatory modules enriched for disease- and trait-associated SNPs. We provide this collection of direct TF-DNA interactions and cis-regulatory modules through the UniBind web-interface (http://unibind.uio.no).

DiscMLA: An Efficient Discriminative Motif Learning Algorithm over High-Throughput Datasets

The transcription factors (TFs) can activate or suppress gene expression by binding to specific sites, hence are crucial regulatory elements for transcription. Recently, series of discriminative motif finders have been tailored to offering promising strategy for harnessing the power of large quantities of accumulated high-throughput experimental data. However, in order to achieve high speed, these algorithms have to sacrifice accuracy by employing simplified statistical models during the searching process. In this paper, we propose a novel approach named Discriminative Motif Learning via AUC (DiscMLA) to discover motifs on high-throughput datasets. Unlike previous approaches, DiscMLA tries to optimize with a more comprehensive criterion (AUC) during motifs searching. In addition, based on an experimental observation of motif identification on large-scale datasets, some novel procedures are designed to accelerate DiscMLA. The experimental results on 52 real-world datasets demonstrate that our approach substantially outperforms previous methods on discriminative motif learning problems. DiscMLA' stability, discriminability, and validity will help to exploit high-throughput datasets and answer many fundamental biological questions.

A novel k-mer set memory (KSM) motif representation improves regulatory variant prediction

The representation and discovery of transcription factor (TF) sequence binding specificities is critical for understanding gene regulatory networks and interpreting the impact of disease-associated noncoding genetic variants. We present a novel TF binding motif representation, the k-mer set memory (KSM), which consists of a set of aligned k-mers that are overrepresented at TF binding sites, and a new method called KMAC for de novo discovery of KSMs. We find that KSMs more accurately predict in vivo binding sites than position weight matrix (PWM) models and other more complex motif models across a large set of ChIP-seq experiments. Furthermore, KSMs outperform PWMs and more complex motif models in predicting in vitro binding sites. KMAC also identifies correct motifs in more experiments than five state-of-the-art motif discovery methods. In addition, KSM-derived features outperform both PWM and deep learning model derived sequence features in predicting differential regulatory activities of expression quantitative trait loci (eQTL) alleles. Finally, we have applied KMAC to 1600 ENCODE TF ChIP-seq data sets and created a public resource of KSM and PWM motifs. We expect that the KSM representation and KMAC method will be valuable in characterizing TF binding specificities and in interpreting the effects of noncoding genetic variations.

Direct AUC optimization of regulatory motifs

Motivation

The discovery of transcription factor binding site (TFBS) motifs is essential for untangling the complex mechanism of genetic variation under different developmental and environmental conditions. Among the huge amount of computational approaches for de novo identification of TFBS motifs, discriminative motif learning (DML) methods have been proven to be promising for harnessing the discovery power of accumulated huge amount of high-throughput binding data. However, they have to sacrifice accuracy for speed and could fail to fully utilize the information of the input sequences.

Results

We propose a novel algorithm called CDAUC for optimizing DML-learned motifs based on the area under the receiver-operating characteristic curve (AUC) criterion, which has been widely used in the literature to evaluate the significance of extracted motifs. We show that when the considered AUC loss function is optimized in a coordinate-wise manner, the cost function of each resultant sub-problem is a piece-wise constant function, whose optimal value can be found exactly and efficiently. Further, a key step of each iteration of CDAUC can be efficiently solved as a computational geometry problem. Experimental results on real world high-throughput datasets illustrate that CDAUC outperforms competing methods for refining DML motifs, while being one order of magnitude faster. Meanwhile, preliminary results also show that CDAUC may also be useful for improving the interpretability of convolutional kernels generated by the emerging deep learning approaches for predicting TF sequences specificities.

Availability and Implementation

CDAUC is available at: https://drive.google.com/drive/folders/0BxOW5MtIZbJjNFpCeHlBVWJHeW8.

Supplementary information

Supplementary data are available at Bioinformatics online.

LMMO: A Large Margin Approach for Refining Regulatory Motifs

Although discriminative motif discovery (DMD) methods are promising for eliciting motifs from high-throughput experimental data, they usually have to sacrifice accuracy and may fail to fully leverage the potential of large datasets. Recently, it has been demonstrated that the motifs identified by DMDs can be significantly improved by maximizing the receiver-operating characteristic curve (AUC) metric, which has been widely used in the literature to rank the performance of elicited motifs. However, existing approaches for motif refinement choose to directly maximize the non-convex and discontinuous AUC itself, which is known to be difficult and may lead to suboptimal solutions. In this paper, we propose Large Margin Motif Optimizer (LMMO), a large-margin-type algorithm for refining regulatory motifs. By relaxing the AUC cost function with the surrogate convex hinge loss, we show that the resultant learning problem can be cast as an instance of difference-of-convex (DC) programs, and solve it iteratively using constrained concave-convex procedure (CCCP). To further save computational time, we combine LMMO with existing techniques for improving the scalability of large-margin-type algorithms, such as cutting plane method. Experimental evaluations on synthetic and real data illustrate the performance of the proposed approach. The code of LMMO is freely available at: https://github.com/ekffar/LMMO.

BEESEM: estimation of binding energy models using HT-SELEX data

Motivation

Characterizing the binding specificities of transcription factors (TFs) is crucial to the study of gene expression regulation. Recently developed high-throughput experimental methods, including protein binding microarrays (PBM) and high-throughput SELEX (HT-SELEX), have enabled rapid measurements of the specificities for hundreds of TFs. However, few studies have developed efficient algorithms for estimating binding motifs based on HT-SELEX data. Also the simple method of constructing a position weight matrix (PWM) by comparing the frequency of the preferred sequence with single-nucleotide variants has the risk of generating motifs with higher information content than the true binding specificity.

Results

We developed an algorithm called BEESEM that builds on a comprehensive biophysical model of protein-DNA interactions, which is trained using the expectation maximization method. BEESEM is capable of selecting the optimal motif length and calculating the confidence intervals of estimated parameters. By comparing BEESEM with the published motifs estimated using the same HT-SELEX data, we demonstrate that BEESEM provides significant improvements. We also evaluate several motif discovery algorithms on independent PBM and ChIP-seq data. BEESEM provides significantly better fits to in vitro data, but its performance is similar to some other methods on in vivo data under the criterion of the area under the receiver operating characteristic curve (AUROC). This highlights the limitations of the purely rank-based AUROC criterion. Using quantitative binding data to assess models, however, demonstrates that BEESEM improves on prior models.

Availability and Implementation

Freely available on the web at http://stormo.wustl.edu/resources.html .

Contact

stormo@wustl.edu.

Supplementary information

Supplementary data are available at Bioinformatics online.

Comparison of discriminative motif optimization using matrix and DNA shape-based models

BACKGROUND

Transcription factor (TF) binding site specificity is commonly represented by some form of matrix model in which the positions in the binding site are assumed to contribute independently to the site's activity. The independence assumption is known to be an approximation, often a good one but sometimes poor. Alternative approaches have been developed that use k-mers (DNA "words" of length k) to account for the non-independence, and more recently DNA structural parameters have been incorporated into the models. ChIP-seq data are often used to assess the discriminatory power of motifs and to compare different models. However, to measure the improvement due to using more complex models, one must compare to optimized matrix models.

RESULTS

We describe a program "Discriminative Additive Model Optimization" (DAMO) that uses positive and negative examples, as in ChIP-seq data, and finds the additive position weight matrix (PWM) that maximizes the Area Under the Receiver Operating Characteristic Curve (AUROC). We compare to a recent study where structural parameters, serving as features in a gradient boosting classifier algorithm, are shown to improve the AUROC over JASPAR position frequency matrices (PFMs). In agreement with the previous results, we find that adding structural parameters gives the largest improvement, but most of the gain can be obtained by an optimized PWM and nearly all of the gain can be obtained with a di-nucleotide extension to the PWM.

CONCLUSION

To appropriately compare different models for TF bind sites, optimized models must be used. PWMs and their extensions are good representations of binding specificity for most TFs, and more complex models, including the incorporation of DNA shape features and gradient boosting classifiers, provide only moderate improvements for a few TFs.

WSMD: weakly-supervised motif discovery in transcription factor ChIP-seq data

Although discriminative motif discovery (DMD) methods are promising for eliciting motifs from high-throughput experimental data, due to consideration of computational expense, most of existing DMD methods have to choose approximate schemes that greatly restrict the search space, leading to significant loss of predictive accuracy. In this paper, we propose Weakly-Supervised Motif Discovery (WSMD) to discover motifs from ChIP-seq datasets. In contrast to the learning strategies adopted by previous DMD methods, WSMD allows a "global" optimization scheme of the motif parameters in continuous space, thereby reducing the information loss of model representation and improving the quality of resultant motifs. Meanwhile, by exploiting the connection between DMD framework and existing weakly supervised learning (WSL) technologies, we also present highly scalable learning strategies for the proposed method. The experimental results on both real ChIP-seq datasets and synthetic datasets show that WSMD substantially outperforms former DMD methods (including DREME, HOMER, XXmotif, motifRG and DECOD) in terms of predictive accuracy, while also achieving a competitive computational speed.

New BAR tools for mining expression data and exploring Cis-elements in Arabidopsis thaliana

Identifying sets of genes that are specifically expressed in certain tissues or in response to an environmental stimulus is useful for designing reporter constructs, generating gene expression markers, or for understanding gene regulatory networks. We have developed an easy-to-use online tool for defining a desired expression profile (a modification of our Expression Angler program), which can then be used to identify genes exhibiting patterns of expression that match this profile as closely as possible. Further, we have developed another online tool, Cistome, for predicting or exploring cis-elements in the promoters of sets of co-expressed genes identified by such a method, or by other methods. We present two use cases for these tools, which are freely available on the Bio-Analytic Resource at http://BAR.utoronto.ca.

DNA Shape Features Improve Transcription Factor Binding Site Predictions In Vivo

Interactions of transcription factors (TFs) with DNA comprise a complex interplay between base-specific amino acid contacts and readout of DNA structure. Recent studies have highlighted the complementarity of DNA sequence and shape in modeling TF binding in vitro. Here, we have provided a comprehensive evaluation of in vivo datasets to assess the predictive power obtained by augmenting various DNA sequence-based models of TF binding sites (TFBSs) with DNA shape features (helix twist, minor groove width, propeller twist, and roll). Results from 400 human ChIP-seq datasets for 76 TFs show that combining DNA shape features with position-specific scoring matrix (PSSM) scores improves TFBS predictions. Improvement has also been observed using TF flexible models and a machine-learning approach using a binary encoding of nucleotides in lieu of PSSMs. Incorporating DNA shape information is most beneficial for E2F and MADS-domain TF families. Our findings indicate that incorporating DNA sequence and shape information benefits the modeling of TF binding under complex in vivo conditions.

Identification of Predictive Cis-Regulatory Elements Using a Discriminative Objective Function and a Dynamic Search Space

The generation of genomic binding or accessibility data from massively parallel sequencing technologies such as ChIP-seq and DNase-seq continues to accelerate. Yet state-of-the-art computational approaches for the identification of DNA binding motifs often yield motifs of weak predictive power. Here we present a novel computational algorithm called MotifSpec, designed to find predictive motifs, in contrast to over-represented sequence elements. The key distinguishing feature of this algorithm is that it uses a dynamic search space and a learned threshold to find discriminative motifs in combination with the modeling of motifs using a full PWM (position weight matrix) rather than k-mer words or regular expressions. We demonstrate that our approach finds motifs corresponding to known binding specificities in several mammalian ChIP-seq datasets, and that our PWMs classify the ChIP-seq signals with accuracy comparable to, or marginally better than motifs from the best existing algorithms. In other datasets, our algorithm identifies novel motifs where other methods fail. Finally, we apply this algorithm to detect motifs from expression datasets in C. elegans using a dynamic expression similarity metric rather than fixed expression clusters, and find novel predictive motifs.

SeAMotE: a method for high-throughput motif discovery in nucleic acid sequences

Background

The large amount of data produced by high-throughput sequencing poses new computational challenges. In the last decade, several tools have been developed for the identification of transcription and splicing factor binding sites.

Results

Here, we introduce the SeAMotE (Sequence Analysis of Motifs Enrichment) algorithm for discovery of regulatory regions in nucleic acid sequences. SeAMotE provides (i) a robust analysis of high-throughput sequence sets, (ii) a motif search based on pattern occurrences and (iii) an easy-to-use web-server interface. We applied our method to recently published data including 351 chromatin immunoprecipitation (ChIP) and 13 crosslinking immunoprecipitation (CLIP) experiments and compared our results with those of other well-established motif discovery tools. SeAMotE shows an average accuracy of 80% in finding discriminative motifs and outperforms other methods available in literature.

Conclusions

SeAMotE is a fast, accurate and flexible algorithm for the identification of sequence patterns involved in protein-DNA and protein-RNA recognition. The server can be freely accessed at http://s.tartaglialab.com/new_submission/seamote.

Electronic supplementary material

The online version of this article (doi:10.1186/1471-2164-15-925) contains supplementary material, which is available to authorized users.