TL;DR
This paper introduces TAMC, a novel deep learning framework that accurately predicts where transcription factors (TFs) bind to DNA using ATAC-seq data. TAMC excels by not requiring complex bias correction and by effectively integrating both footprint and non-footprint features, outperforming existing methods, especially with limited sequencing data.
What Problem Does This Paper Address?
Understanding where transcription factors (TFs) bind to DNA is fundamental to deciphering how genes are regulated. Traditional methods for identifying these binding sites (TFBSs) often rely on experimental assays like ChIP-seq, which are resource-intensive and can only profile one TF at a time. Computational approaches have emerged as alternatives, but many struggle with accuracy, especially when dealing with the nuances of ATAC-seq data. ATAC-seq measures chromatin accessibility, which indirectly reflects TF binding. A key challenge is that TF binding events leave "footprints" – regions of reduced enzymatic cleavage – in ATAC-seq profiles. However, existing footprinting tools often require complex bias correction steps to account for sequence biases in the enzymatic cleavage process, which can limit their effectiveness and introduce their own artifacts. Furthermore, the assumption that all TFs leave similar footprint patterns is being challenged, suggesting a need for more TF-specific approaches.
Methods
The core of TAMC is a 1D convolutional neural network (1D-CNN) architecture. Unlike previous methods, TAMC bypasses the need for explicit bias correction during signal processing. Instead, it takes a multi-faceted input signal derived from ATAC-seq data. This input signal is constructed by combining:
- Footprint profiles: Calculated using the TOBIAS footprint score metric.
- Genomic cleavage signals and slopes: Processed using modified HINT-ATAC scripts.
These signals are further enriched by separating them into channels based on read strand and fragment size, creating a 9-channel input vector for each potential binding site. The 1D-CNN then learns to extract features from this combined signal. The network architecture involves multiple convolutional layers with varying kernel sizes (3, 5, and 7) to capture features at different scales, followed by max-pooling, concatenation, and fully connected layers with a sigmoid activation for predicting binding probability. TAMC can also generate TF-specific models, meaning a separate model is trained for each TF, allowing it to capture TF-specific binding characteristics.
Datasets
The study utilized publicly available ATAC-seq and ChIP-seq datasets from three human cell types: GM12878, HepG2, and K562. These datasets were obtained from the UCSC ENCODE portal. ChIP-seq data was used to label motif-predicted binding sites (MPBSs) as either bound or unbound. For training TAMC models, ATAC-seq data from GM12878 and HepG2 were downsized to 150 million high-quality aligned reads. For testing, K562 ATAC-seq data was used across a range of sequencing depths (from 5 million to 300 million reads) to evaluate performance under varying data availability. The study also applied TAMC to predict CTCF binding sites in human embryonic cells using ATAC-seq data from 2-cell and 8-cell embryos.
Key Insights and Ideas
- Bias Correction is Not Always Necessary: TAMC demonstrates that a well-designed deep learning model can implicitly learn to correct for biases in ATAC-seq data, eliminating the need for manual bias correction steps that can be error-prone and computationally intensive.
- Holistic Feature Integration: By combining footprint information with cleavage signals, strand, and fragment size, TAMC captures a richer set of features than methods relying solely on footprints or cleavage patterns.
- TF-Specific Models Matter: The performance gains observed when using TF-specific models highlight that different TFs have distinct binding preferences and chromatin interaction patterns, which a single generic model might miss.
- Deep Learning's Power for Low-Depth Data: TAMC's superior performance with limited sequencing depth is a significant advantage, making it more applicable to challenging datasets like single-cell ATAC-seq.
Key Results
- TAMC consistently outperformed existing footprinting tools (TOBIAS and HINT-ATAC) in TFBS prediction across multiple TFs and cell types, particularly in cross-data prediction scenarios.
- When trained on ATAC-seq data with higher sequencing depth, TAMC models showed significantly better cross-data prediction performance.
- TAMC models trained on multiple cell types (GM12878 and HepG2) generally performed better in cross-data predictions than models trained on a single cell type.
- Removing footprint scores from the input signal drastically compromised TAMC's performance for all TFs, indicating the importance of footprint information. However, cleavage profiles, strand, and fragment size also contributed significantly for specific TFs.
- TAMC successfully predicted changes in constitutive CTCF binding sites during early human embryonic development, showing a significant increase in constitutive binding sites at the 8-cell stage compared to the 2-cell stage.
Limitations and Future Work
- Data Scarcity for Some TFs: For TFs with very few labeled bound sites, TAMC's performance was comparable to existing methods, suggesting that insufficient training data remains a challenge. The authors suggest combining data from multiple cell types to mitigate this.
- ChIP-seq Noise: TFs with high background noise in their ChIP-seq data (indicating poor antibody specificity) showed lower performance with TF-specific models. The authors propose a two-step training approach: first, train on common binding features across all TFs, and then fine-tune with TF-specific data.
- Optimizing TF-Specific Features: While TAMC can capture TF-specific features, further optimization of input parameters (like input region length, strand, and fragment size separation) for specific TFs could potentially improve performance even further.
- Single-Cell ATAC-seq Application: The authors express optimism that TAMC's ability to handle low-depth data makes it a promising tool for analyzing TF binding dynamics in single-cell ATAC-seq experiments.
Summary updated