Source-linked AI summary
DeeperBind: Enhancing Prediction of Sequence Specificities of DNA Binding Proteins
Hamid Reza Hassanzadeh, May D. Wang
TL;DR
High-throughput binding assays yield noisy measurements, while accurate localization of transcription-factor binding sites matters for biological and drug-related applications. DeeperBind combines convolutional feature extraction with LSTMs to model positional dynamics in DNA probes, including varying-length sequences. The authors report that it outperforms available prediction tools, often by a large margin.
Problem
High-throughput techniques do not provide sufficiently accurate binding-affinity predictions and motif characterization, so downstream analysis is needed to locate protein binding sites.
Source: §I · p. 1Method
DeeperBind combines convolutional feature extraction with LSTM recurrence to model positional dynamics and sub-region contributions in DNA probes, including varying-length sequences.
Source: C. DeeperBind · p. 3, §IV · p. 6Results
DeeperBind consistently outperforms other methods by a large margin on measured-versus-predicted probe intensities, and shows better generalization behavior despite higher model complexity.
Source: §III · p. 5Takeaways & Limitations
The authors conclude that DeeperBind predicts binding specificities from high-throughput data while retaining positional information and handling varying-length sequences.
Source: §IV · pp. 5–6
Abstract
from arXiv · showhide
Transcription factors (TFs) are macromolecules that bind to \textit{cis}-regulatory specific sub-regions of DNA promoters and initiate transcription. Finding the exact location of these binding sites (aka motifs) is important in a variety of domains such as drug design and development. To address this need, several \textit{in vivo} and \textit{in vitro} techniques have been developed so far that try to characterize and predict the binding specificity of a protein to different DNA loci. The major problem with these techniques is that they are not accurate enough in prediction of the binding affinity and characterization of the corresponding motifs. As a result, downstream analysis is required to uncover the locations where proteins of interest bind. Here, we propose DeeperBind, a long short term recurrent convolutional network for prediction of protein binding specificities with respect to DNA probes. DeeperBind can model the positional dynamics of probe sequences and hence reckons with the contributions made by individual sub-regions in DNA sequences, in an effective way. Moreover, it can be trained and tested on datasets containing varying-length sequences. We apply our pipeline to the datasets derived from protein binding microarrays (PBMs), an in-vitro high-throughput technology for quantification of protein-DNA binding preferences, and present promising results. To the best of our knowledge, this is the most accurate pipeline that can predict binding specificities of DNA sequences from the data produced by high-throughput technologies through utilization of the power of deep learning for feature generation and positional dynamics modeling.
I. INTRODUCTION
Locating transcription-factor binding sites can inform gene regulation, disease-variant analysis, and drug design, but high-throughput measurements are noisy and require downstream analysis. DeeperBind is proposed to predict DNA-probe binding preferences and is reported to surpass DeepBind on popular metrics.
I. INTRODUCTION
- Binding-site identification can inform gene regulation, gene-function annotation, regulatory-process analysis, causal disease-variant identification, and drug design.Transcription-factor genes are estimated to comprise 5–10% of eukaryotic genes.Source: §I · p. 1
- High-throughput in-vivo and in-vitro assays produce noisy, erroneous, low-resolution measurements that require downstream processing to elicit useful information.The study evaluates its method on protein binding microarray experiments, which measure relative DNA-probe binding preferences through spot intensities.Source: §I · p. 1
- DeeperBind is reported to surpass DeepBind on popular metrics for predicting transcription-factor sequence specificities.The introduction also reports published performance measures for RankMotif++, PERGO, KmerHMM, and Seed & Wobble.Source: §I · pp. 1–2
II. MATERIALS AND METHODS
The section motivates treating DNA sequences as one-dimensional image-like inputs for deep convolutional models. It connects local nucleotide patterns and convolutional filters to motif recognition.
II. MATERIALS AND METHODS
- DNA loci with specific functions are treated as one-dimensional images because neighboring nucleotides exhibit nonrandom patterns.The passage compares this local dependence with neighboring-pixel dependence in images.Source: §II · p. 2
- Stationary patterns among nearby nucleotides motivate convolutional models that delay fully connected layers until later pipeline stages.The passage presents this as a way to avoid using fully connected layers until the final stages.Source: §II · p. 2
- Convolution resembles scanning DNA sequences with position weight matrix filters, which were used to characterize motifs.Source: §II · p. 2
A. DeepBind
DeepBind predicts probe intensity using convolutional motif detectors, pooling, and fully connected layers. Its global max-pooling supports varying-length inputs but removes positional information, and its single-motif assumption can misrepresent probes with multiple contributing motifs.
A. DeepBind
- DeepBind estimates probe intensity through convolution, nonlinear thresholding, max-pooling, and one or two fully connected layers.Its convolutional filters scan sequences like position weight matrices, without requiring log-odds coefficients.Source: A. DeepBind · p. 2
- Global max-pooling gives DeepBind consistent-size inputs for varying-length sequences but discards positional information from convolutional feature maps.Source: A. DeepBind · p. 2
- DeepBind implicitly assumes at most one motif per probe, which can mislead training when several moderately good motifs jointly yield high affinity.In that case, the model may penalize all but one detector and overweight the remaining one.Source: A. DeepBind · p. 2
- The model also lacks long- and short-range dependency modeling, although motif-level and nucleotide-level dependencies may both affect binding affinity.The passage notes that motif position near probe ends versus its middle can correspond to different affinity because of technological limitations.Source: A. DeepBind · p. 2
B. Recurrent Neural Networks
LSTMs use gates to address the difficulty regular recurrent networks have propagating gradients across long sequences. The described architecture can stack LSTM layers to model sequential signals.
B. Recurrent Neural Networks
- Regular recurrent networks can struggle to capture long-range dependencies because gradients become prohibitively small when propagated toward distant time points.Source: B. Recurrent Neural Networks · p. 2
- LSTMs address vanishing gradients through differentiable gates that let memory units decide what information to retain or forget.The passage attributes the LSTM proposal to Hochreiter et al..Source: B. Recurrent Neural Networks · p. 2
- In the described LSTM, sigmoid gates act on the current input and previous hidden state to selectively forget, emphasize, or add cell-state information.The gate parameters are tied across the unrolled units.Source: B. Recurrent Neural Networks · p. 3
- Stacked LSTM layers pass the preceding layer’s hidden state upward, and Figure 1 depicts both a memory cell and a deep LSTM network.Source: B. Recurrent Neural Networks · p. 3, Figure 1 · p. 3
C. DeeperBind
DeeperBind adds recurrence to DeepBind’s convolutional design to model positional dependencies in DNA probes. The combined model also supports variable-length inputs and is end-to-end trainable.
C. DeeperBind
- CNN features feed LSTMs that model long-term motif and short-term nucleotide dependencies, including positional patterns such as indels.The recurrent structure is intended to capture positional dynamics that convolutional and PWM approaches assuming adjacent-position independence may miss.Source: C. DeeperBind · p. 3
- The recurrent architecture supports variable-length inputs, allowing data from different platforms to be used in a unified framework.The design is end-to-end trainable and can exploit GPUs for speed gains.Source: C. DeeperBind · p. 3
- Probes are one-hot encoded, convolved at stride one without pooling, and passed through LSTM and fully connected layers to predict binding preference.Omitting pooling preserves positional information; dropout regularization is used in the fully connected layers.Source: C. DeeperBind · p. 3, Figure 2 · p. 4
III. RESULTS
On the PBM test arrays, DeeperBind consistently outperformed the comparison methods in Spearman rank correlation and classification. Training-array rank plots also show its predictions concentrated toward the top for positive probes.
III. RESULTS
- Both models were trained on array #1, split into 70% training and 30% validation, and evaluated on the independent array #2.Settings were selected by validation-set Spearman coefficients; the PBM datasets covered CEH-22 and Oct-1.Source: §III · p. 4
- For the top 100 positive probes in array #1, DeeperBind’s predicted ranks are concentrated toward the top of the charts.The paper interprets this training-array pattern as evidence that the model learned discriminating patterns predictive of measured intensities.Source: Figure 3 · p. 4, §III · p. 4
- DeeperBind consistently outperformed the other methods by a large margin in Spearman rank correlation, with DeepBind the runner-up.The reported correlations are for array #2 and include shallow learners as well as DeepBind.Source: Table II · p. 5, §III · p. 5
- DeeperBind surpassed the other models in true positive rate at 1% false positive rate for both datasets.The paper also reports better generalization behavior in ROC analysis despite DeeperBind’s higher model complexity.Source: Table III · p. 5, §III · p. 5
IV. CONCLUSION
DeeperBind combines CNNs and LSTMs to predict DNA-binding affinity, retaining positional information and supporting variable-length sequences. The authors report that it beats the most accurate prediction tools they know, while identifying motif discovery and heterogeneous-data benchmarking as future work.
IV. CONCLUSION
- DeeperBind beats the most accurate prediction tools the authors know, often by a large margin.Source: §IV · pp. 5–6
- LSTM layers capture positional dynamics lost through pooling in the baseline and allow DeeperBind to handle variable-length sequences.Source: §IV · p. 6
- The authors identify motif discovery and comprehensive benchmarking on heterogeneous in-vivo and in-vitro data as future directions.Source: §IV · p. 6