TL;DR
ChromBPNet is a novel deep learning model that accurately predicts chromatin accessibility at base resolution, effectively disentangling the influence of transcription factor binding from enzymatic biases inherent in experimental assays. This breakthrough enables more precise identification of regulatory elements, transcription factor binding sites, and the impact of genetic variants on gene regulation, with broad implications for understanding complex traits and diseases.
What Problem Does This Paper Address?
Understanding how DNA sequence dictates gene regulation is fundamental to biology. While chromatin accessibility assays like ATAC-seq and DNase-seq have mapped millions of potential regulatory elements (cREs) across various cellular contexts, deciphering the precise sequence "syntax" that governs transcription factor (TF) binding and accessibility at these sites has remained challenging. Furthermore, genetic variants within these cREs can subtly alter gene regulation, contributing to disease and complex traits, but pinpointing their exact functional impact has been difficult. Existing computational methods often struggle to disentangle the true regulatory signal from technical biases introduced by the experimental assays themselves, particularly the sequence preferences of the enzymes used (like Tn5 transposase in ATAC-seq).
Methods
The core innovation is ChromBPNet, a deep learning model that addresses the problem of assay bias. It employs a "bias-factorized" architecture, meaning it explicitly models and separates the enzyme's sequence preference from the regulatory sequence determinants.
Key methodological aspects include:
- Bias Modeling: A pre-trained "bias model" is used to capture the sequence preferences of the enzymes (Tn5 for ATAC-seq, DNase-I for DNase-seq). This model's weights are frozen during the main training.
- Bias-Factorized Architecture: ChromBPNet integrates this bias model with a "TF Model" (a standard sequence-to-profile model). The training process is split into two stages: first, fitting the bias model's predictions to the measured data, and second, training the TF model on the residual variance. This allows for "bias-corrected" predictions.
- Sequence-to-Profile Prediction: The model takes a 2114 bp DNA sequence and predicts a 1000 bp base-resolution chromatin accessibility profile, outputting both total read counts and a probability distribution across the sequence.
- Model Interpretation: Techniques like DeepLIFT and TF-MODISCO are used to interpret the model's predictions, identifying key TF motifs and understanding how variants affect regulatory syntax.
- Marginal Footprinting: A novel method to generate footprints for individual motifs by marginalizing over the influence of other co-occurring motifs.
Datasets
The study utilizes a variety of datasets for training and evaluation:
- ENCODE datasets: Bulk ATAC-seq and DNase-seq data from multiple cell lines (GM12878, HEPG2, IMR90, H1ESC, K562) were used for training and benchmarking.
- Single-cell ATAC-seq (scATAC-seq) datasets: Data from human microglia and coronary smooth muscle cells were used to train models for specific cell types.
- African ancestry bulk ATAC-seq datasets: Data from six African populations were used to assess cross-ancestry generalization.
- Naked DNA transposition libraries: Used to train bias models to capture enzyme sequence preferences.
- ChIP-seq datasets: Used for comparative analysis and validation of predicted TF binding.
- Variant effect prediction datasets:
- DNase QTLs (dsQTLs) in Yoruba LCLs: Used to benchmark variant effect prediction.
- Chromatin QTLs (caQTLs) in European and African LCLs: Used for benchmarking.
- Massively Parallel Reporter Assays (MPRAs): Used to assess the impact of variants on regulatory element activity.
- Fine-mapped GWAS variants for blood traits: Used to prioritize variants associated with complex traits.
- SPI1 binding QTLs (bQTLs): Used to evaluate predictions of pioneer TF occupancy.
The paper also introduces novel metrics for variant effect prediction (logFC, JSD, AAQ, IES, IPS) and contributes curated benchmarking datasets.
Key Insights and Ideas
- Assay bias is a major confounder: The study highlights that standard deep learning models trained on ATAC-seq and DNase-seq data often learn the sequence preferences of the enzymes (Tn5 and DNase-I) rather than purely the regulatory sequence syntax. This bias significantly distorts profile shape predictions and motif discovery.
- Bias correction is crucial for accurate interpretation: By explicitly modeling and correcting for enzyme bias, ChromBPNet achieves significantly improved concordance between ATAC-seq and DNase-seq predictions and reveals more accurate TF footprints and motif lexicons.
- Compact motif lexicons drive accessibility: Even across diverse cell types, a relatively small set of TF motifs, including composite elements, is sufficient to accurately predict chromatin accessibility, suggesting a conserved underlying regulatory logic.
- Chromatin accessibility predictions can inform TF occupancy and reporter activity: The model's predictions of variant effects on chromatin accessibility correlate well with observed effects on TF binding and reporter gene activity, demonstrating its utility in predicting downstream regulatory consequences.
- Generalizability across assays, ancestries, and cell types: ChromBPNet shows robust performance and generalization capabilities, performing well across different assays, diverse ancestry groups, and even in primary cell types and low-coverage single-cell data.
Key Results
- High prediction accuracy: ChromBPNet achieves high concordance between predicted and observed chromatin accessibility profiles, with Jensen-Shannon Distances (JSD) as low as 0.26 after bias correction.
- Improved motif discovery: Bias correction leads to the discovery of more accurate and biologically relevant TF motifs, eliminating spurious motifs caused by enzyme bias.
- Competitive performance with larger models: ChromBPNet, despite its lightweight design, performs competitively or better than much larger, state-of-the-art models (like Enformer) in predicting variant effects on chromatin accessibility, TF binding, and reporter activity.
- Accurate variant effect prediction: The model accurately predicts the impact of genetic variants on chromatin accessibility, achieving high correlation with observed effect sizes in various QTL studies. For example, it achieves an average precision (AP) of 0.54 for ATAC-seq and 0.43 for DNase-seq in discriminating significant dsQTLs.
- Robustness to low sequencing depth: ChromBPNet can effectively impute profiles and sequence annotations even from moderate coverage (25M reads) ATAC-seq datasets, making it applicable to challenging datasets like scATAC-seq.
Limitations and Future Work
- Distal context: While ChromBPNet excels at modeling local sequence context for profile shape, predicting total accessibility still lags behind pseudo-replicate concordance, suggesting that distal genomic context might be critical. The authors suggest future work on progressively building long-context models from pre-trained local models.
- Other regulatory layers: The current model focuses on chromatin accessibility. Future extensions could integrate other regulatory layers (e.g., TF binding, histone modifications) and multiple axes of regulatory variation.
- "Setter" TFs: The model primarily identifies "pioneer" TFs that influence accessibility. It may not fully capture "setter" TFs that bind open chromatin but don't actively shape it.
- Model architectures: While CNNs are shown to be competitive, exploring alternative architectures could be a future direction.
- Population-specific linkage disequilibrium: While the model learns causal regulatory mechanisms, it's important to acknowledge that population-specific linkage disequilibrium can still influence variant interpretation.
- Distal chromatin QTLs and eQTLs: Current benchmarks are restricted to local QTLs; future evaluations should extend to distal chromatin QTLs and eQTLs.
- Variant types: The current focus is on SNVs; future work should include indels and STRs.
- Integration with other data: Integrating predictions with measured effects on diverse molecular and cellular phenotypes will be crucial for precise mechanism mapping.
Summary updated