Predicting the sequence specificities of DNA- and RNA-binding proteins by deep learning

Abstract

Knowing the sequence specificities of DNA-and RNA-binding proteins is essential for developing models of the regulatory processes in biological systems and for identifying causal disease variants. Here we show that sequence specificities can be ascertained from experimental data with 'deep learning' techniques, which offer a scalable, flexible and unified computational approach for pattern discovery. Using a diverse array of experimental data and evaluation metrics, we find that deep learning outperforms other state-of-the-art methods, even when training on in vitro data and testing on in vivo data. We call this approach DeepBind and have built a stand-alone software tool that is fully automatic and handles millions of sequences per experiment. Specificities determined by DeepBind are readily visualized as a weighted ensemble of position weight matrices or as a 'mutation map' that indicates how variations affect binding within a specific sequence.DNA-and RNA-binding proteins play a central role in gene regulation, including transcription and alternative splicing. The sequence specificities of a protein are most commonly characterized using position weight matrices 1 (PWMs), which are easy to interpret and can be scanned over a genomic sequence to detect potential binding sites. However, growing evidence indicates that sequence specificities can be more accurately captured by more complex techniques 2-5 . Recently, 'deep learning' has achieved record-breaking performance in a variety of information technology applications 6,7 . We adapted deep learning methods to the task of predicting sequence specificities and found that they compete favorably with the state of the art. Our approach, called DeepBind, is based on deep convolutional neural networks and can discover new patterns even when the locations of patterns within sequences are unknown-a task for which traditional neural networks require an exorbitant amount of training data.There are several challenging aspects in learning models of sequence specificity using modern high-throughput technologies. First, the data come in qualitatively different forms. Protein binding microarrays (PBMs) 8 and RNAcompete assays 9 provide a specificity coefficient for each probe sequence, whereas chromatin immunoprecipitation (ChIP)-seq 10 provides a ranked list of putatively bound sequences of varying length, and HT-SELEX 11 generates a set of very high affinity sequences. Second, the quantity of data is large. A typical high-throughput experiment measures between 10,000 and 100,000 sequences, and it is computationally demanding to incorporate them all. Third, each data acquisition technology has its own artifacts, biases and limitations, and we must discover the pertinent specificities despite these unwanted effects. For example, ChIP-seq reads often localize to "hyper-ChIPable" regions of the genome near highly expressed genes 12 .DeepBind (Fig. 1) addresses the above challenges. (i) It can be applied to both microarray and sequencing data; (ii) it can learn from millions of...

Variant Interpretation: Functional Assays to the Rescue

“…For example, DeepBind is an algorithm that uses deep learning to predict the sequence specificity of RNA-and DNA-binding proteins. 66 DeepBind is trained on MAVE data, including DNA-binding assays such as SELEX, protein-binding microarrays, and RNA-binding assays such as RNAcompete. [67][68][69] MAVE data can also be useful in evaluating new predictive tools, as was done for EVmutation, which predicts variant effects in proteins from co-variation in multiple-sequence alignments.…”

Insight into GATA1 transcriptional activity through interrogation ofciselements disrupted in human erythroid disorders

“…This difficulty is most clearly reflected in the distribution of mutations listed in databases of Mendelian disorders, such as the Human Gene Mutation Database, where most mutations are found within coding regions (86%) or at intronic splice sites (11%), with only a small fraction (3%) identified in regulatory regions (14). Newer methods for annotating and predicting the impact of noncoding (NC) variants have provided substantial improvements (15,16), but experimental validation of the presumed effects remains critical for the determination of pathogenicity and elucidation of the mechanism of action (13,17).…”

“…First, we trained a gapped k-mer support vector machine (gkmer-SVM) on EB and K562 open chromatin data and used delta-SVM to predict single nucleotide effects (16). We then used already-trained models for the TFs GATA1, TAL1, KLF1, and NFE2 from DeepBind, a convolutional neural network approach (15). For each CRE proximal to the 20 MED genes, we created a "mutation map" of the predicted effects of all possible single nucleotide changes (15).…”

“…We then used already-trained models for the TFs GATA1, TAL1, KLF1, and NFE2 from DeepBind, a convolutional neural network approach (15). For each CRE proximal to the 20 MED genes, we created a "mutation map" of the predicted effects of all possible single nucleotide changes (15).…”

Stacked Autoencoder Method for Fabric Defect Detection

“…On the robotic area, Ian Lenz, Honglak Lee and Ashutosh Saxena are using deep learning methods in order to solve the problem of detecting robotic grasps in an RGB-D view of a scene containing objects [15]. Bioinformatics area, Deep learning using for recognizing disease from DNA sequences [16] or predicting protein secondary structure [17]. Another state-of-art study is on largescale video classification with Convolutional Neural Networks (CNN) using a new dataset of 1 million YouTube videos belonging to 487 classes [18].…”