EpiExpr: Predicting gene expression using epigenetic data and chromatin interactions

EpiExpr is a flexible deep learning framework that integrates 1D epigenetic tracks and 3D chromatin interactions to accurately predict gene expression and prioritize regulatory elements, offering a computationally efficient alternative to sequence-based models.

The Gist

Imagine your DNA is a massive, 3-billion-letter instruction manual for building a human. But here's the catch: only about 1.5% of that manual is the actual "recipe" for making proteins (the genes). The other 98.5% is a chaotic mix of notes, sticky tabs, and folded pages that tell the cell when and how much of each recipe to use. This is the world of gene regulation.

For a long time, scientists have struggled to read this messy manual. They have maps of the "sticky tabs" (epigenetics) and photos of how the paper is folded (3D chromatin structure), but they haven't had a good way to predict exactly how loud a specific gene will "sing" based on those clues.

Enter EpiExpr, a new AI tool introduced in this paper that acts like a super-smart translator for this biological manual.

The Problem: The "Too Big to Read" Manual

Think of the genome like a giant library.

• Old AI models (like Enformer or EPInformer) tried to read the entire library at once to predict a book's popularity. To do this, they needed massive supercomputers and huge amounts of time. They were like trying to read a whole encyclopedia to find one sentence.
• Older, simpler models were fast but often missed the big picture, like only reading the first page of a book and guessing the ending.

The Solution: EpiExpr

The researchers built EpiExpr, which is like a smart librarian who doesn't need to read every single letter of the DNA. Instead, they look at the clues left on the pages.

1. The "One-Dimensional" Librarian (EpiExpr-1D)

Imagine you are trying to guess how popular a song is just by looking at the volume knobs and light switches on the mixing board, without hearing the music itself.

• The Clues: These are "epigenetic tracks" (like ATAC-seq or ChIP-seq). They show which parts of the DNA are "open" (easy to read) or "marked" with chemical tags.
• The Magic: EpiExpr-1D uses a Residual CNN (a type of AI that learns by looking for patterns in layers). It's like a detective who looks at the volume knobs, realizes "Oh, the bass is turned up high here, so the song must be loud," and makes a prediction.
• The Win: It predicted gene activity just as well as the massive, slow models that read the DNA letters, but it did it much faster and with less computing power. It's like using a flashlight instead of a searchlight.

2. The "3D" Librarian (EpiExpr-3D)

Here is where it gets really cool. DNA isn't just a straight line; it's a tangled ball of yarn. Sometimes, a "volume knob" (an enhancer) is physically far away from the "song" (the gene) on the straight line, but because the DNA is folded, they are actually touching!

• The Analogy: Imagine a long string of beads. Bead #100 is the song, and Bead #500 is the volume knob. On the string, they are far apart. But if you fold the string so they touch, the knob controls the song.
• The Magic: EpiExpr-3D adds a Graph Neural Network (GNN). Think of this as a map of the folded yarn. It connects the distant volume knobs to the songs they actually touch.
• The Win: By adding this "folding map," the AI gets even better at predicting gene activity, especially for genes that are controlled by distant parts of the genome.

Why This Matters (The "So What?")

1. It's Fast and Cheap: You don't need a billion-dollar supercomputer to run this. A standard laptop or a single graphics card can do the job. This means more labs can use it.
2. It's Flexible: The researchers built a "Lego kit" (called a Snakemake pipeline) that lets scientists plug in their own data from any cell type (liver, brain, skin) without rewriting the whole code.
3. It's Accurate: They tested it against real-world experiments (CRISPRi), where they physically turned off enhancers to see if the gene stopped working. EpiExpr correctly identified which "volume knobs" were the real deal, proving it understands the biology, not just the math.

The Bottom Line

EpiExpr is a new, efficient, and flexible tool that helps us understand how the "folding" and "marking" of our DNA control our genes. It proves that you don't need to read every single letter of the genetic code to understand how life works; sometimes, just looking at the notes and the folds is enough to predict the song.

It's a step toward a future where we can easily simulate how changing our environment or our genes might affect our health, all without needing a supercomputer in our basement.

Technical Summary

1. Problem Statement

Decoding gene expression from epigenomic landscapes is a fundamental challenge in genomics. While deep learning models have advanced the field, existing approaches face significant limitations:

• Sequence-based models (e.g., Enformer, Borzoi, AlphaGenome): These rely on DNA sequence embeddings and transformer architectures. They are computationally expensive, often requiring TPUs, and are constrained by input window sizes (typically 200 kb to 1 Mb), limiting their ability to capture distal regulatory effects.
• Existing epigenetic models (e.g., Epi-GraphReg): While these use epigenetic tracks and 3D interactions, they are often rigid, supporting only fixed cell types, fixed numbers of tracks, and fixed resolutions (e.g., 100 bp epigenetic, 5 kb expression).
• Hybrid models (e.g., EPInformer): These integrate sequence, epigenetics, and 3D data but remain computationally intensive due to transformer layers and reliance on pre-computed activity-by-contact (ABC) scores.

There is a need for a scalable, flexible, and computationally efficient framework that predicts gene expression using epigenetic features and 3D chromatin interactions without the heavy computational burden of sequence-based transformers.

---

2. Methodology

The authors introduce EpiExpr, a flexible deep learning framework comprising two models: EpiExpr-1D and EpiExpr-3D.

A. Data Curation & Pipeline

• Open-source Snakemake Pipelines: The authors developed flexible pipelines to construct training datasets for arbitrary cell types, variable numbers of epigenetic tracks, and user-defined resolutions.
• Input Data:
  • 1D Tracks: ChIP-seq, ATAC-seq, DNase-seq, etc.
  • 3D Interactions: Hi-C, HiChIP, or PCHi-C loops (processed via FitHiChIP).
  • Target: CAGE (Cap Analysis of Gene Expression) tracks.
• Resolution: Supports user-defined resolutions (e.g., 100 bp for epigenetics, 5 kb for expression), provided the expression resolution is an integer multiple of the epigenetic resolution.
• Chunking: Genomes are segmented into 6 Mb windows; the central 2 Mb is used for prediction, flanked by 2 Mb background regions.

B. Model Architectures

1. EpiExpr-1D (Residual CNN):

• Backbone: A Residual Convolutional Neural Network (ResNet) inspired by ResNet18.
• Mechanism: Uses iterative, adaptive downsampling to map high-resolution epigenetic inputs ($e$) to lower-resolution expression outputs ($c$).
• Structure:
  • Input tracks are projected to a fixed channel size ($M$, a power of 2).
  • Residual Blocks: The model computes the prime factors of the downsampling ratio ($e/c$) and assigns them to successive residual blocks. If fewer than three factors exist, blocks with a factor of 1 are inserted to ensure a minimum of three blocks.
  • Each block contains four convolutional layers, batch normalization, and activation (GELU/ELU).
  • Final layers compress channels to a single expression prediction.

2. EpiExpr-3D (CNN + Graph Neural Network):

• Hybrid Approach: Integrates 3D chromatin interactions into the EpiExpr-1D backbone.
• Workflow:
  1. The EpiExpr-1D residual blocks process epigenetic data to generate node embeddings at the target expression resolution.
  2. Graph Construction: Nodes represent expression bins; edges represent significant chromatin loops (FDR < 0.1) derived from FitHiChIP.
  3. GNN Layers: Two architectures are tested:
     • Graph Attention Network (GATv2Conv): Uses 8 attention heads and 2 layers.
     • Graph Transformer (TransformerConv): Combines message passing with label propagation.
  4. Edge Normalization: Tested scikit-learn row normalization (E1) and double-stochastic normalization (E2).
  5. Residual Connections: Initial features are added to GNN layers to improve stability and performance.
  6. Output: The GNN output is passed through the final convolutional layers of the CNN to generate the final expression prediction.
• Training Strategy: End-to-end training is used to avoid gradient collapse issues associated with pre-training CNNs separately from GNNs.

---

3. Key Contributions

1. Flexibility: Unlike previous models (e.g., Epi-GraphReg), EpiExpr supports multiple cell types, variable numbers of epigenetic tracks, and arbitrary user-defined resolutions.
2. Computational Efficiency: EpiExpr achieves performance comparable to massive sequence-based transformer models (like EPInformer) but relies solely on epigenetic data and lightweight CNN/GNN architectures, requiring significantly fewer resources (no DNA sequence embeddings, no massive pre-training).
3. Integration of 3D Data: Successfully integrates 3D chromatin loops (HiChIP/Hi-C) via Graph Neural Networks to capture distal regulatory effects without the computational overhead of sequence-based transformers.
4. Open Source: Provides comprehensive Snakemake pipelines for data curation and model training, facilitating reproducibility and adaptation to new datasets.

---

4. Results

The models were benchmarked on GM12878 and K562 cell lines using data from the Epi-GraphReg and Basenji repositories.

• vs. Epi-GraphReg:
  • EpiExpr-1D consistently outperformed the 1D version of Epi-GraphReg in Pearson correlation and Mean Absolute Error (MAE) across both cell types and expression bins.
  • EpiExpr-3D (specifically with Graph Transformer) showed marginal improvements over EpiExpr-1D for non-expressing bins and in K562, demonstrating the value of 3D data for specific contexts. Both EpiExpr variants significantly outperformed Epi-GraphReg-3D.
• vs. Sequence-Based Models (EPInformer):
  • EpiExpr-1D achieved correlation scores comparable to or higher than EPInformer variants (which use DNA sequence + epigenetics + ABC scores) in K562, and similar in GM12878.
  • EpiExpr-3D (Graph Transformer) matched the performance of EPInformer-PE-Activity-HiC (the most complex variant) while using substantially lower computational resources and avoiding the need for ABC score pre-computation.
• Validation with CRISPRi-FlowFISH:
  • Using DeepSHAP for interpretability, EpiExpr models were tested against experimentally validated enhancer-gene pairs.
  • EpiExpr-1D achieved higher mean AUPRC (0.3677) than the ABC model (0.3508) in identifying functional regulatory regions.
  • Specificity: In the KLF1 locus analysis, EpiExpr models correctly identified validated enhancers while avoiding false positives (70–100 kb downstream) that the ABC model predicted, indicating higher specificity.
• Resource Usage:
  • Data curation took <30 mins (~1 GB RAM).
  • Inference took ~40 mins on a single GPU (peak 10 GB memory), demonstrating high efficiency.

---

5. Significance

• Scalability: EpiExpr offers a practical solution for large-scale, multi-cell-type gene expression modeling, overcoming the computational bottlenecks of transformer-based sequence models.
• Interpretability: The framework successfully prioritizes regulatory elements (enhancers) consistent with experimental CRISPRi screens, validating its biological relevance.
• Paradigm Shift: The study demonstrates that epigenetic signals and 3D chromatin architecture alone are sufficient to predict gene expression with high accuracy, challenging the necessity of massive DNA sequence embeddings for this specific task.
• Future Utility: The flexible pipeline enables researchers to dissect the contributions of specific epigenetic modifications and 3D genome organizations across diverse experimental settings, paving the way for broader applications in regulatory genomics.