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 matrices1 (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 techniques2,3,4,5. Recently, 'deep learning' has achieved record-breaking performance in a variety of information technology applications6,7. 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 assays9 provide a specificity coefficient for each probe sequence, whereas chromatin immunoprecipitation (ChIP)-seq10 provides a ranked list of putatively bound sequences of varying length, and HT-SELEX11 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 genes12.
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 sequences through parallel implementation on a graphics processing unit (GPU); (iii) it generalizes well across technologies, even without correcting for technology-specific biases; (iv) it can tolerate a moderate degree of noise and mislabeled training data; and (v) it can train predictive models fully automatically, alleviating the need for careful and time-consuming hand-tuning. Importantly, a trained model can be applied and visualized in ways that are familiar to users of PWMs. We explored two downstream applications: uncovering the regulatory role of RNA binding proteins (RBPs) in alternative splicing, and analyzing disease-associated genetic variants that can affect transcription factor binding and gene expression.
Training DeepBind and scoring sequences
For training, DeepBind uses a set of sequences and, for each sequence, an experimentally determined binding score. Sequences can have varying lengths (14–101 nt in our experiments), and binding scores can be real-valued measurements or binary class labels. For a sequence s, DeepBind computes a binding score f(s) using four stages:
The convolution stage (convM) scans a set of motif detectors with parameters M across the sequence. Motif detector Mk is a 4 × m matrix, much like a PWM of length m but without requiring coefficients to be probabilities or log odds ratios. The rectification stage isolates positions with a good pattern match by shifting the response of detector Mk by bk and clamping all negative values to zero. The pooling stage computes the maximum and average of each motif detector's rectified response across the sequence; maximizing helps to identify the presence of longer motifs, whereas averaging helps to identify cumulative effects of short motifs, and the contribution of each is determined automatically by learning. These values are fed into a nonlinear neural network with weights W, which combines the responses to produce a score (Fig. 2a, Supplementary Fig. 1 and Supplementary Notes, sec. 1.1).
We use deep learning techniques13,14,15,16 to infer model parameters and to optimize algorithm settings. Our training pipeline (Fig. 2b) alleviates the need for hand tuning, by automatically adjusting many calibration parameters, such as the learning rate, the degree of momentum14, the mini-batch size, the strength of parameter regularization, and the dropout probability15.
To obtain the results reported below, we trained DeepBind models on a combined 12 terabases of sequence data, spanning thousands of public PBM, RNAcompete, ChIP-seq and HT-SELEX experiments. We provide the source code for DeepBind together with an online repository (http://tools.genes.toronto.edu/deepbind/) of 927 DeepBind models representing 538 distinct transcription factors and 194 distinct RBPs, each of which was trained on high-quality data and can be applied to score new sequences using an easily installed executable file with no hardware or software requirements.
Ascertaining DNA sequence specificities
To evaluate DeepBind's ability to characterize DNA-binding protein specificity, we used PBM data from the revised DREAM5 TF-DNA Motif Recognition Challenge by Weirauch et al.17. The PBM data represent 86 different mouse transcription factors, each measured using two independent array designs. Both designs contain ∼40,000 probes that cover all possible 10-mers, and all nonpalindromic 8-mers, 32 times. Participating teams were asked to train on the probe intensities using one array design and to predict the intensities of the held-out array design, which was not made available to participants.
Weirauch et al.17 evaluated 26 algorithms that can be trained on PBM measurements, including FeatureREDUCE17, BEEML-PBM18, MatrixREDUCE19, RankMotif++20 and Seed-and-Wobble21. For each individual algorithm, they optimized the data preprocessing steps to attain best test performance. Methods were evaluated using the Pearson correlation between the predicted and actual probe intensities, and values from the area under the receiver operating characteristic (ROC) curve (AUC) computed by setting high-intensity probes as positives and the remaining probes as negatives17. To the best of our knowledge, this is the largest independent evaluation of this type. When we tested DeepBind under the same conditions, it outperformed all 26 methods (Fig. 3a). DeepBind also ranked first among 15 teams when we submitted it to the online DREAM5 evaluation script (Supplementary Table 1).
To assess the ability of DeepBind models trained using in vitro PBM data to predict sequence specificities measured using in vivo ChIP-seq data, we followed the method described by Weirauch et al.17. Predicting transcription factor binding in vivo is more difficult because it is affected by other proteins, the chromatin state and the physical accessibility of the binding site. We found that DeepBind also achieves the highest score when applied to the in vivo ChIP-seq data (Fig. 3b and Supplementary Fig. 2). The best method reported in the original evaluation (Team_D, a k-mer-based model) and the best reported in the revised evaluation (FeatureREDUCE, a hybrid PWM/k-mer model) both had reasonable, but not the best, performance on in vivo data, which might be due to overfitting to PBM noise17. DeepBind motifs for DREAM5 data and the data sets that follow are available in the online repository.
Ascertaining RNA sequence specificities
We next evaluated the ability of DeepBind to identify RNA binding sites. For experimental training data, we used previously published binding data for 207 distinct RBPs from 24 eukaryotes, including 85 distinct human RBPs22, which were generated using the in vitro RNAcompete system9. These RBPs span different structural families, such as the RNA recognition motif, hnRNP K-homology domain and multiple zinc finger families. The arrays contain ∼240,000 short single-stranded probes (30–41 nt) that cover all 9-mers at least 16 times and all 7-mers at least 155 times. The probes were randomly split into two similarly sized sets called SetA and SetB. Note that RNAcompete probes were designed to have only weak secondary structure and may not fully reflect the specificities of structure-selective RBPs such as Vtsp1 (ref. 23).
We trained DeepBind on RNAcompete SetA sequences and evaluated its predictive accuracy on held-out SetB sequences. DeepBind consistently had better overall performance than MartixREDUCE19 and Ray et al.'s9 PWM_align_Z method (Fig. 3c). This result was consistent across a panel of evaluation methods, including Pearson and Spearman correlations between the predicted and measured probe intensities, Pearson and Spearman correlations between all 7-mer E-scores (rank-based) and Z-scores (intensity-based), and AUCs computed from predicted and actual probe intensities (Supplementary Notes, sec. 4.2, Supplementary Table 2 and Supplementary Fig. 3). E-score correlation is robust to outliers and array biases21. Correlations of 7-mer scores can be computed using measured data, providing upper bounds on performance. Strikingly, DeepBind performs as well as experimental E-scores (Fig. 3c).
We also applied models trained on in vitro RNAcompete data to in vivo CLIP-seq (crosslinking and immunoprecipitation followed by high-throughput sequencing) and RIP-seq (RNA-immunoprecipitation sequencing) data22. DeepBind was comparable or substantially better than published PFMs (position frequency matrices) on the in vivo data sets we considered, in particular TARDBP, PTBP1 and hnRNP A1 (Fig. 3d; Supplementary Notes, sec. 4.3, Supplementary Table 3 and Supplementary Fig. 4). The ROC curves for Figure 3d are in Supplementary Figure 5.
We found that it is more challenging to train models to predict RBP sequence specificities than transcription factor specificities. Even though the DREAM5 transcription factor training and testing data sets were collected using different array designs, the Spearman correlations between predicted and measured probe intensities were significantly higher for DREAM5 transcription factors than for RNAcompete RBPs (P = 1.8 × 10−30; one-sided Mann-Whitney U test; 0.758 average, n = 66 and 0.439 average, n = 244, respectively). There are several potential reasons. First, RBPs usually bind to single-stranded RNA22, which is more flexible than DNA and can fold into a relatively stable secondary structure. Second, RNA recognition motif, the most abundant RNA binding domain, is highly flexible, and usually the assembly of multiple domains are needed for proper binding24. Third, “indirect readout”5 in which the structure of nucleotides guide the binding process, is more prominent in RBPs; Gupta and Gribskov25 analyzed 211 protein-RNA structural complexes and reported that nearly three-quarters of protein-RNA interactions involve the RNA backbone, not just the bases.
In vitro models accurately identify in vivo bound sequences
We next evaluated DeepBind's performance using 506 in vivo ENCODE ChIP-seq data sets, which were preprocessed to remove protocol and laboratory biases26 (Supplementary Table 4). Unlike experiments with in vitro data, these experiments were influenced by cell type–specific effects, transcription factor-nucleosome interactions, cooperation and competition between transcription factors and other cofactors, and pioneer transcription factors that can remodel chromatin and facilitate the binding of other transcription factors27. To train DeepBind, we used as positives the 101-bp sequences centered at the point source called for each peak, and we used shuffled positive sequences with matching dinucleotide composition as negatives (same as ENCODE's analysis27; Supplementary Notes, sec. 3).
For computational reasons, most existing methods analyze only the top few hundred peaks from among tens of thousands of peaks (the median number of peaks for ENCODE is ∼17,000). However, Wang et al.27 found that, for example, ∼16,000 of the top ∼20,000 SPI1 peaks contain the SPI1 motif. For each data set, DeepBind was able to incorporate all ChIP-seq training peaks, and our results suggest that these extra peaks provide useful information. Specifically, we used the top 500 even-numbered peaks as held-out test data and trained models on either all other peaks (DeepBind*) or just the top 500 odd-numbered peaks (DeepBind) (Supplementary Notes, sec. 6.1 and Supplementary Fig. 6). For comparison, we also applied MEME-ChIP28 to the same top 500 odd-numbered training peaks, derived five PWMs for each experiment using the ENCODE settings27, and scored test sequences using either the top PWM (MEME-M1) or the sum of scores for all five PWMs (MEME-SUM). MEME-ChIP could not be trained using all possible training peaks due to computational limitations.
The 506 ChIP-seq experiments represent 137 unique transcription factors, so we computed one AUC for each transcription factor by taking its median test AUC across all associated experiments (antibodies, cell lines, laboratories). DeepBind trained on the top 500 odd peaks achieved higher test AUC (0.85 average) than both MEME-SUM (0.82 average; P = 2 × 10−15, Wilcoxon one-sided signed-rank test, n = 137) and MEME-M1 (0.78 average; P = 1 × 10−23). By training DeepBind* on all peaks not held out we achieved significantly higher test AUC (0.90 average) than training on the top 500 odd peaks either DeepBind (P = 2.4 × 10−20) or MEME-SUM (P = 7 × 10−23) (Fig. 3e and Supplementary Table 5). We additionally tested the DeepBind FoxA2 ChIP-seq model using 64 'electrophoretic mobility shift assay' (EMSA)-measured binding affinities29 and found that it achieves the highest Spearman correlation among published methods, including JASPAR PWMs30, TRANSFAC PWMs31, and other models trained on ChIP-seq data (Supplementary Notes, sec. 6.3 and Supplementary Fig. 7).
We also trained DeepBind on Jolma et al.'s in vitro HT-SELEX data for 303 human DNA-binding domains, 84 mouse DNA-binding domains and 151 full-length human transcription factors, representing 411 distinct transcription factors32 (Supplementary Table 6). DeepBind achieved higher test AUC (0.71 average) than the PWMs discovered by Jolma et al.'s semi-automatic algorithm32 (0.63 average; P = 2.8 × 10−94, Wilcoxon one-sided signed-rank test, n = 566) (Fig. 3e and Supplementary Notes, sec. 7.2, and Supplementary Table 7). We next evaluated the performance of several HT-SELEX models for which there exists corresponding in vivo ChIP-seq data from ENCODE. DeepBind achieved higher in vivo AUC (0.85 average) than both the best-possible choice of Jolma et al.'s PWMs32 (0.82 average; P = 0.001 Wilcoxon one-sided signed-rank test, n = 35) and the best MEME-ChIP configuration MEME-SUM-5k (0.82 average; P = 0.002), suggesting that DeepBind can generalize from HT-SELEX to other data acquisition technologies despite being based on a general-purpose machine learning framework (Fig. 3f and Supplementary Notes, sec. 7.3, and Supplementary Table 8). The ROC curves for Figure 3f are in Supplementary Figure 8.
Identifying and visualizing damaging genetic variants
Genetic variants that create or abrogate binding sites can alter gene expression patterns and potentially lead to disease33. A promising direction in precision medicine is to use binding models to identify, group and visualize variants that potentially change protein binding. To explore the effects of genetic variations using DeepBind, we developed a visualization called a 'mutation map' (Supplementary Notes, sec. 10.1), which illustrates the effect that every possible point mutation in a sequence may have on binding affinity. A mutation map conveys two types of information. First, for a given sequence, the mutation map shows how important each base is for the DeepBind analysis by the height of the base letter. Second, the mutation map includes a heat map of size 4 by n, where n is the sequence length, indicating how much each possible mutation will increase or decrease the binding score.
We examined variants within promoters using nearly 600 DeepBind models trained using ChIP-seq and HT-SELEX data. Figure 4a shows how several Human Gene Mutation Database (HGMD)34 single-nucleotide variants (SNVs) disrupted an SP1 binding site in the LDL-R promoter, leading to familial hypercholesterolemia35. It can be seen in the wild-type mutation map that all HGMD SNVs falling in the SP1 binding site decrease the score. Figure 4b depicts how a cancer risk variant (rs6983267) in a MYC enhancer36 weakens a TCF7L2 binding site. Figure 4c illustrates how an inherited gain-of-function SNP creates a GATA1 binding site in the globin cluster, disrupting the original globin cluster promoters37. GATA4 is known to regulate Bcl-2 (B-cell CLL/lymphoma 2) expression in ovarian granulosa cell tumors38. We analyzed all COSMIC 69 (ref. 39) SNVs located in the Bcl-2 promoters and found two that damage a single GATA4 binding site (Fig. 4d).
One notable example involves tandem RFX3 binding sites in the GPR56 promoter. A 15-bp deletion of a highly conserved segment deletes one of the binding sites, resulting in abnormal cortical development40. DeepBind analysis discovered a third RFX3 binding site located on the opposite strand, overlapping both of the known tandem binding sites (Fig. 4e), and so the 15-bp deletion might destroy this third RFX binding site as well.
Two of the highest-scoring SNV-ridden promoters in HGMD belong to HBB (hemoglobin beta chain) and F7 (coagulation factor VII). Disruption of these two genes results in β-thalassemia and hemophilia, respectively. Figure 4f,g shows how numerous HGMD SNVs might damage transcription factor binding sites in these promoters.
Lastly, it was shown that mutations in the TERT (telomerase reverse transcriptase) promoter are linked to many types of aggressive cancer41, mostly by gain-of-function mutations creating ETS family42,43 or GABP41 transcription factor binding sites. We analyzed the TERT promoter using DeepBind ETS (ELK1/ELK4) and GABP-α models and the resulting mutation maps all show potential gain-of-function mutations corresponding to the literature. In Figure 4h, all COSMIC 69 SNVs observed in the analyzed regions are shown. It can be seen that the two highly recurrent somatic G→A SNVs (G228A and G250A) create the strongest putative GABP-α binding sites, which have each been confirmed to increase expression of TERT41. In addition to these highly recurrent mutations, DeepBind also identified a familial T→G mutation that is known to increase TERT expression, potentially by introducing a GABP-α binding site42.
DeepBind models identify deleterious genomic variants
We investigated whether DeepBind can be used to predict deleterious SNVs in promoters, by training a deep neural network to discriminate between high-frequency derived alleles (neutral or negative) and simulated variants (putatively deleterious, or positive) from the CADD framework44. The scores of ∼600 DeepBind transcription factor models for the wild type and mutant sequences were used as inputs (∼1,200 inputs; Supplementary Fig. 9). The rationale is that a true transcription factor binding site is likely to be located with other transcription factor binding sites, and so these additional scores collectively provide context. When evaluated using held-out test data, the neural network, called DeepFind, achieved an AUC of 0.71, which increased to 0.73 when we included as inputs the distance to the closest transcription start site and a transversion/transition flag. When we included nine conservation features, the AUC increased to 0.76. Supplementary Figure 10 shows that distribution of DeepFind scores for derived alleles and simulated variants. For comparison, the published genome-wide CADD scores achieved an AUC of 0.64, even though they were computed using hundreds of features, including overlapping ChIP-seq peaks, transcription factor binding sites, DNA accessibility profiles and chromatin state, among others (Supplementary Notes, sec. 9).
In vitro models are consistent with known splicing patterns
Alternative splicing, through which multiple transcripts are produced from a single gene, is responsible for generating significant transcriptional diversity in metazoa, to the degree that more than 95% of multi-exon human genes are alternatively spliced45. RBPs play a crucial role in regulating splicing, having an impact on a wide variety of developmental stages such as stem cell differentiation46 and tissue development47. We predicted binding scores at junctions near exons that are putatively regulated by known splicing regulators that exhibit large changes in splicing when knocked down, including: Nova (neuro-oncological ventral antigen 1 and 2), PTBP1 (polypyrimidine tract binding protein 1), Mbnl1/2 (muscleblind-like protein), hnRNP C (heterogeneous nuclear ribonucleoprotein C) and cytotoxic TIA (T-cell intracellular antigen-like 1).
Predictions were consistent with experimental CLIP-seq data and known binding profiles of studied RBPs (Fig. 5 and Supplementary Notes, sec. 8, and Supplementary Table 9). For example, exons known to be downregulated by Nova had higher Nova scores in their upstream introns, and exons known to be upregulated by Nova had higher Nova scores in their downstream intron48. Similarly, TIA has been shown to upregulate exons when bound to the downstream intron49, and PTBP1 has been shown to suppress exon inclusion when bound to upstream introns of weaker splice sites50.
Though there is no single agreed upon metric for evaluating the quality of sequence specificity predictions17, we found that DeepBind surpasses the state of the art across a wide variety of data sets and evaluation metrics. Importantly, our results show that DeepBind models trained in vitro work well at scoring in vivo data, suggesting an ability to capture genuine properties of nucleic acid binding interactions. DeepBind scales well to large data sets and, for both ChIP-seq and HT-SELEX, we found that there was valuable information to be learned from sequences that other techniques discard for computational reasons.
A frequent concern with large, deep models is that they can overfit the training data. To address this, we incorporated several regularizers developed in the deep learning community, including dropout, weight decay and early stopping (Supplementary Notes, sec. 2). Indeed, we found that the calibration phase of DeepBind training was crucial for obtaining models that generated robust predictions on held-out test data; our experiments indicate that at least 30 calibration settings should be evaluated to obtain reliable training settings across all data sets (Supplementary Notes, sec. 2.2). Automatic calibration and training (Fig. 2b) are computationally demanding, so we implemented these stages on powerful GPUs for a 10–70× acceleration. Once a model was trained, we extracted its parameters and provided a CPU implementation of the model for easy application by downstream users. Users can browse the available RBP and transcription factor models using our online repository, which displays a familiar PWM-like representation of each model (Fig. 6).
DeepBind is based on deep learning, a scalable and modular pattern discovery method, and does not rely on common application-specific heuristics such as 'seed finding'. Deep learning furthermore has an extremely active research community that is garnering huge investment from academia and industry; we believe current and future insights from the deep learning community will lead to enhancements to DeepBind and to sequence analysis in general.
Deep learning techniques.
We implemented several of the simplest and most effective techniques now practiced within the deep learning community: mini-batch stochastic gradient descent (SGD), Nesterov momentum14, rectified linear units6 (ReLUs), dropout regularization15, automatic model selection through calibration16 ('hyper-parameter search') and, finally, GPU acceleration to facilitate fast training on large data sets. When combined, these simple techniques have achieved breakthrough performance in computer vision6, speech recognition7 and more. Supplementary Notes, sec. 1 and 2 explain the role of each technique in more detail.
Given a specific convolutional network architecture (number of motif detectors, lengths of motif detectors, width of pooling, number of layers, number of parameters), the training procedure for DeepBind is conceptually straightforward and can be outlined as follows. At the outset of training, all parameters are initialized to random values (Supplementary Notes, sec. 2.1). A 'mini-batch' of N randomly selected input sequences is then fed through the network, generating N (initially random) score predictions. The discrepancy between each prediction and its corresponding target is used to improve the performance of the network through a step known as 'back-propagation' (Fig. 2 and Supplementary Notes, sec. 1.2). Training then continues with a new mini-batch of N randomly selected input sequences. The number of iterations (mini-batches) used in a final training run is determined by the automatic calibration phase.
Back propagation is merely an efficient way to compute the partial derivative of each parameter of a deep model with respect to a training objective. Given a random mini-batch of training sequences s(1.N) and corresponding target scores t(1.N), the performance of the network is improved by approximately minimizing the training objective
where denotes the L1 norm (sum of absolute values), weight decay coefficients β1, β2 ≥ 0, and LOSS(p,t) is a function such as the squared error (p − t)2 or negative log-likelihood (Supplementary Notes, sec. 1.3).
Deep neural networks typically require human expertise and repeated manual attempts to train correctly. The reason is that the networks themselves, and their training procedure, are sensitive to many calibration parameters, also known as 'hyper-parameters' or 'nuisance parameters' (Supplementary Notes, sec. 2.1). Calibration parameters must be chosen to avoid under- or overfitting to the training data, but are otherwise irrelevant to the end-user. For neural networks, typical parameters include 'learning rates', 'network depth', 'initial weights' and 'weight decay'. The complete list of hyperparameters and the way they are sampled are summarized in Supplementary Table 10. Similarly, for support vector machines one must likewise choose a kernel function and weight decay. For random forests one must choose a tree depth and an ensemble size.
For each DeepBind model to be trained, we randomly sampled 30 complete sets of calibration parameters16, and for each fold of a three-way random split of the training data we trained 30 models in parallel on a single GPU. We rate each calibration by its average validation performance across those folds (threefold cross-validation AUC or mean squared error). The best calibration was then used to train a new model on all the training data (Supplementary Notes, sec. 2.2). The code for performing the experiments is available for download in the Supplementary Software.
We are grateful to K.B. Cook, Q.D. Morris and T.R. Hughes for helpful discussions. This work was supported by a grant from the Canadian Institutes of Health Research (OGP-106690) to B.J.F., a John C. Polanyi Fellowship Grant to B.J.F., and funding from the Canadian Institutes for Advanced Research to B.J.F. and M.T.W. B.A. was supported by a joint Autism Research Training and NeuroDevNet Fellowship. A.D. was supported by a Fellowship from the Natural Science and Engineering Research Council of Canada.
Integrated supplementary information
This code download is distributed as part of the Nature Biotechnology supplementary software release for DeepBind. Users of DeepBind are encouraged to instead use the latest source code and binaries for scoring sequences at http://tools.genes.toronto.edu/deepbind/Your access to and use of the downloadable code (the “Code”) contained in this Supplementary Software is subject to a non-exclusive, revocable, non-transferable, and limited right to use the Code for the exclusive purpose of undertaking academic, governmental, or not-for-profit research. Use of the Code or any part thereof for commercial or clinical purposes is strictly prohibited in the absence of a Commercial License Agreement from Deep Genomics. (firstname.lastname@example.org)