Abstract
Accurately predicting Transcription Factor Binding Sites (TFBS) is essential for understanding gene regulation. Traditional models often rely on Position Weight Matrices (PWMs) that assume positional independence, failing to capture complex biological dependencies. We present CNN_TF, a dual-path Convolutional Neural Network framework that integrates raw DNA sequences with histone modification data to capture both low-order and higher-order position dependencies. Our results demonstrate that CNN_TF outperforms state-of-the-art methods by at least 5.3% in AUPR, highlighting the critical role of higher-order dependencies and the complementarity of sequence and chromatin features in TFBS prediction.
Introduction
Transcription Factors (TFs) are regulatory proteins that bind to specific DNA sequences to control gene expression. Mapping these binding sites (TFBS) is a fundamental task in genomics. Historically, Position Weight Matrices (PWMs) have been the standard for TFBS prediction. However, PWMs operate under the "low-order" assumption that each position in a binding site is independent. In reality, biological binding often involves "higher-order" dependencies, where nucleotides at distant positions influence binding affinity. While deep learning models like DeepBind and DeepSEA have made strides, they often struggle to extract these long-range dependencies from raw sequence data alone.
Methodology
The proposed CNN_TF architecture utilizes a dual-path approach to capture diverse genomic features.
Dual-Path Architecture
The model consists of two parallel Convolutional Neural Network (CNN) arms designed to extract complementary information:
- Sequence Arm (Low-Order): This path processes one-hot encoded 2,000-bp DNA segments. It focuses on local motifs and immediate neighbor dependencies.
- Histone Arm (Higher-Order): This path processes histone modification data, which naturally spans multiple positions and reflects the broader chromatin environment. By applying a CNN to these features (binned into 100-bp windows), the model extracts large-scale, higher-order dependencies.
The features extracted from both arms are concatenated into a "Combined features" vector and passed through a softmax classifier to predict TFBS presence.
Datasets and Benchmarking
The model was evaluated using mouse embryonic stem (mES) cells (13 TFs) and human cell lines (5 TFs across 5 cell types). CNN_TF was benchmarked against traditional classifiers (SVM, Random Forest), DNA shape-based models, and deep learning architectures like DeepSEA and DanQ.
Results
Performance Gains
CNN_TF demonstrated a significant performance boost, outperforming existing state-of-the-art methods by at least 5.3% in AUPR. In human cell-type tests, the model showed an average AUPR increase of 0.152 over DNA shape-based models. For specific TFs like CTCF, the model achieved a Positive Predictive Value (PPV) of 98.6%, a substantial improvement over older HMM-based methods.
Complementarity of Dependencies
A key finding of this study is the complementarity between low-order (sequence) and higher-order (histone) dependencies. In 12 out of 13 TFs tested in mouse cells, the combined model yielded superior results compared to using either feature set in isolation.
As shown in the figure above, the combined model consistently maintains higher AUPR even as the data becomes more imbalanced. Interestingly, higher-order dependencies alone outperformed low-order dependencies in 11 out of 13 TFs, suggesting that the broader genomic context is often more informative than the local motif sequence.
Discussion
CNN_TF effectively demonstrates that incorporating higher-order dependencies via histone modifications significantly enhances TFBS prediction. However, the model's reliance on histone data limits its application to cell types where such experimental data is available. Additionally, the use of fixed 100-bp windows for histone features may be a subject for future optimization. Future research could explore cross-cell-type prediction to mitigate data dependency and investigate Transformer-based architectures to capture even longer-range genomic dependencies.
Summary updated