Cross-Cell-Type Prediction of TF-Binding Site by Integrating Convolutional Neural Network and Adversarial Network

Lan Gongqiang, Jiyun Zhou, Ruifeng Xu, Qin Lu, Hongpeng WangInternational Journal of Molecular Sciences 2019 · DOI

TL;DR

This paper introduces DANN_TF, a novel deep learning method that uses a Convolutional Neural Network (CNN) combined with an Adversarial Network to predict transcription factor binding sites (TFBSs). DANN_TF excels at predicting TFBSs in cell types with limited or no labeled data by learning common features across different cell types, significantly outperforming existing methods.

What Problem Does This Paper Address?

Gene expression is tightly regulated by transcription factors (TFs) binding to specific DNA sequences called Transcription Factor Binding Sites (TFBSs). Accurately predicting these sites is crucial for understanding gene regulation. While experimental methods like ChIP-seq can identify TFBSs, they are costly and time-consuming. Computational methods, particularly deep learning approaches, have shown promise, but they often require substantial amounts of labeled data for each specific TF and cell type. The challenge arises because many TFs lack sufficient labeled data in certain cell types, making accurate prediction difficult. This paper tackles the problem of predicting TFBSs in these data-scarce scenarios by leveraging knowledge from cell types where more data is available.

Methods

The core innovation of DANN_TF lies in its integration of a Convolutional Neural Network (CNN) with an Adversarial Network.

  • Feature Extractor (CNN): This component, built with convolutional and pooling layers, learns to extract relevant sequence features from DNA segments.
  • Label Predictor: This part of the network takes the extracted features and predicts whether a given DNA sequence is a TFBS or not.
  • Domain Classifier (Adversarial Network): This is the key to DANN_TF's cross-cell-type prediction capability. It's trained to distinguish between DNA sequences from the "target" cell type (where prediction is needed) and "source" cell types (where labeled data is available). Crucially, the feature extractor is trained adversarially against this classifier. This means the feature extractor tries to learn features that are indistinguishable between cell types, forcing it to capture general TFBS characteristics rather than cell-type-specific noise. This adversarial training ensures that the learned features are common across different cell types.

The overall framework aims to learn a shared representation of TFBSs that generalizes well, even when the target cell type has limited or no training data.

Datasets

The study utilized two main datasets:

  1. Five TFs in Five Cell Types: This dataset comprised 25 cell-type TF pairs. The cell types included GM12878, H1-hESC, HeLa-S3, HepG2, and K562. The TFs were CTCF, JunD, GABPA, REST, and USF2. TFBS data was obtained from ENCODE.
  2. Additional 13 TFs in Five Cell Types: This larger dataset was used to evaluate the generalization ability of DANN_TF, resulting in 65 cell-type TF pairs.

For training and evaluation, data was split into training, validation, and test sets. The paper also explored scenarios with varying percentages of labeled data in the target cell type (50%, 20%, 10%) for semi-supervised prediction.

Key Insights and Ideas

  • Adversarial Learning for Generalization: The central idea is that by forcing the feature extractor to produce features that fool a cell-type classifier, it learns representations that are robust and generalizable across different cellular contexts. This is a powerful way to overcome data scarcity.
  • Leveraging "Unlabeled" Target Data: DANN_TF can effectively use unlabeled data from the target cell type alongside labeled data from source cell types. This is a significant advantage over methods that strictly require labeled target data.
  • Common TFBS Features Exist: The success of DANN_TF strongly suggests that TFs share fundamental binding characteristics that are conserved across different cell types, even if the specific regulatory context varies.
  • Data Augmentation vs. Cross-Cell-Type Prediction: The paper demonstrates that DANN_TF's cross-cell-type prediction approach, which leverages source cell data, can outperform simple data augmentation strategies when target data is scarce.

Key Results

  • Superior Performance: DANN_TF consistently outperformed a baseline method (which lacked the adversarial component) across various scenarios.
  • For data augmentation (using labeled target data), DANN_TF achieved higher AUC and F1 scores for 21 out of 25 cell-type TF pairs.
  • In semi-supervised prediction (with limited labeled target data), DANN_TF showed significant improvements, especially as the amount of labeled target data decreased.
  • For cross-cell-type prediction (using only unlabeled target data and labeled source data), DANN_TF significantly outperformed the baseline on all 25 cell-type TF pairs.
  • Significant AUC Gains: On a larger set of 65 cell-type TF pairs, DANN_TF achieved significantly higher AUC than the baseline method on 96.9% of pairs.
  • State-of-the-Art Comparison: DANN_TF outperformed established deep learning methods like DeepSEA and DanQ (and their variants) on 24 out of 25 cell-type TF pairs, demonstrating its effectiveness against current benchmarks.
  • Efficiency: DANN_TF had a competitive training time, generally taking less time than methods like DanQ and DanQ-JASPAR, and only slightly more than the baseline method.

Limitations and Future Work

  • Equal Treatment of Source Cell Types: The current DANN_TF treats all source cell types equally. The authors acknowledge that different source cell types might have varying degrees of functional relevance to the target cell type. Future work could explore methods to weigh or select source cell types based on their predicted relevance.
  • Sequence-Specific vs. Non-Specific TFs: While DANN_TF showed improvements for TFs without specific binding motifs, the gains were slightly less pronounced compared to sequence-specific TFs. Further investigation into how to better capture binding features for non-specific TFs could be beneficial.
  • Exploration of Other Architectures: While CNNs are effective, exploring other deep learning architectures or combinations might yield further improvements.

Summary updated

Abstract

Transcription factor binding sites (TFBSs) play an important role in gene expression regulation. Many computational methods for TFBS prediction need sufficient labeled data. However, many transcription factors (TFs) lack labeled data in cell types. We propose a novel method, referred to as DANN_TF, for TFBS prediction. DANN_TF consists of a feature extractor, a label predictor, and a domain classifier. The feature extractor and the domain classifier constitute an Adversarial Network, which ensures that learned features are common features across different cell types. DANN_TF is evaluated on five TFs in five cell types with a total of 25 cell-type TF pairs and compared to a baseline method which does not use Adversarial Network. For both data augmentation and cross-cell-type prediction, DANN_TF performs better than the baseline method on most cell-type TF pairs. DANN_TF is further evaluated by an additional 13 TFs in the five cell types with a total of 65 cell-type TF pairs. Results show that DANN_TF achieves significantly higher AUC than the baseline method on 96.9% pairs of the 65 cell-type TF pairs. This is a strong indication that DANN_TF can indeed learn common features for cross-cell-type TFBS prediction.
BibTeX
@article{gongqiang2019cross,
  author = {Lan Gongqiang and Jiyun Zhou and Ruifeng Xu and Qin Lu and Hongpeng Wang},
  title = {{Cross-Cell-Type Prediction of TF-Binding Site by Integrating Convolutional Neural Network and Adversarial Network}},
  journal = {International Journal of Molecular Sciences},
  year = {2019},
  doi = {10.3390/ijms20143425},
}
Figure