Structural variants in human congenital heart disease disrupt distal genomic regulatory contacts of developmental genes

This study introduces CardioAkita, a machine-learning model that predicts how structural variants disrupt 3D chromatin interactions, demonstrating that such disruptions lead to aberrant gene expression and contribute to the genetic etiology of congenital heart disease.

The Gist

Imagine your DNA isn't just a long list of instructions (like a recipe book), but also a complex, 3D origami sculpture. To make a heart, the cell needs to fold this paper in very specific ways so that the right instructions can "talk" to each other. If you tear a page or fold it wrong, the instructions get scrambled, and the heart might not form correctly. This is what happens in many cases of Congenital Heart Disease (CHD).

Here is a simple breakdown of what this paper discovered, using everyday analogies:

1. The Problem: The "Missing Link" in Heart Defects

Heart defects are the most common birth defects. Scientists have been trying to find the "smoking gun" (the specific genetic mistake) for years. They found many structural variants (SVs)—big chunks of DNA that are deleted, duplicated, or flipped upside down.

• The Analogy: Imagine a library where some books are missing pages, some have extra pages glued in, and some chapters are in the wrong order.
• The Issue: For a long time, scientists only looked at the words on the pages (the genes). But many of these "structural variants" happen in the empty space between the words. Scientists didn't know if these empty spaces mattered. They were like "undetermined significance" errors—clues that might be important, but we didn't know how to read them.

2. The Solution: A Crystal Ball for DNA Folding (CardioAkita)

The researchers built a super-smart computer program called CardioAkita.

• The Analogy: Think of CardioAkita as a 3D architectural simulator. If you give it a blueprint of a house (the DNA sequence), it can predict exactly how the house will be built and how the rooms will connect, even before a single brick is laid.
• The Innovation: Previous simulators were trained on generic "houses" (stem cells). CardioAkita was specifically trained on heart cells as they grow from babies to adults. It learned the unique way heart cells fold their DNA origami.

3. The Discovery: The "Disruption Score"

The team used CardioAkita to look at the DNA of people with heart defects. They wanted to see if the "structural variants" (the torn or flipped pages) messed up the 3D folding.

• The Analogy: They assigned a "Disruption Score" to every mistake.
  • Low Score: The DNA was folded, but the room layout was fine. The house still works.
  • High Score: The DNA was folded so badly that the kitchen was connected to the roof, or the front door opened into the bathroom. The house is broken.
• The Result: They found that people with severe heart defects had the highest disruption scores. The more the DNA origami was messed up, the worse the heart condition was. This proved that these "empty space" mistakes are actually breaking the heart's blueprint.

4. The Experiment: Building the Broken House in a Lab

To prove the computer wasn't just guessing, they went into the lab. They took human stem cells (which can turn into anything) and used gene-editing tools (like molecular scissors) to create the exact same DNA mistakes found in three real patients.

• The Analogy: They took a pristine model house and physically cut out a wall or glued two rooms together, just like the computer predicted.
• The Outcome:
  1. The Fold: The cells folded their DNA exactly as CardioAkita predicted. The "rooms" connected in the wrong ways.
  2. The Chaos: Because the rooms were connected wrong, the wrong instructions were turned on or off. Genes that should have been quiet were shouting, and genes that should have been loud went silent.
  3. The Result: The cells started acting confused, showing signs of heart development going off the rails.

5. The Big Picture: Why This Matters

This paper changes how we diagnose and understand heart disease.

• Before: If a patient had a DNA mistake in a "junk" area, doctors would say, "We don't know if this causes the heart defect."
• Now: We have a tool (CardioAkita) that can look at that "junk" area, simulate the 3D folding, and say, "Yes, this mistake breaks the folding, and that is why the heart is failing."

In summary:
This study showed that the "furniture arrangement" of our DNA (3D structure) is just as important as the "furniture" itself (genes). By building a computer model that understands how heart cells fold their DNA, scientists can now find the hidden causes of heart defects that were previously invisible, opening the door to better diagnoses and potentially new treatments.

Technical Summary

1. Problem Statement

Congenital Heart Disease (CHD) is the most common major congenital anomaly, affecting ~1% of live births. While structural variants (SVs)—including deletions, duplications, inversions, and insertions—are known contributors to CHD, a significant diagnostic gap remains: 60% of sequenced CHD cases lack a known genetic etiology.

• The Challenge: Many pathogenic SVs are non-coding and located far from the genes they regulate. Traditional methods struggle to predict whether an SV disrupts 3D chromatin architecture (e.g., topologically associating domains or TADs) and alters distal enhancer-promoter interactions.
• The Gap: Existing computational models (like the original Akita) were trained on immortalized cell lines (e.g., H1 hESCs) and lack the cell-type specificity required to model the dynamic chromatin folding during cardiac development.

2. Methodology

The authors developed a multi-stage pipeline combining deep learning, large-scale genomic analysis, and experimental validation.

A. Model Development: CardioAkita

• Architecture: An extension of the Akita convolutional neural network (CNN).
• Training Data: High-resolution Hi-C data from WTC11 induced pluripotent stem cells (iPSCs) differentiating into atrial and ventricular cardiomyocytes across multiple time points (Day 0 to Day 45).
• Input/Output: The model takes ~1 Megabase (Mb) of 1-hot encoded DNA sequence as input and predicts chromatin contact maps (Hi-C) for specific developmental stages (e.g., cardiac mesoderm, precursors, mature cardiomyocytes).
• Validation: Performance was evaluated using Mean Squared Error (MSE) and Pearson correlation (1-CORR) against held-out genomic loci. CardioAkita demonstrated superior performance over the original Akita in predicting cell-type-specific interactions, particularly in loci with significant developmental changes.

B. Computational Screening (SuPreMo Pipeline)

• Cohort Analysis: The authors analyzed 42 high-confidence de novo SVs (dnSVs) from CHD probands (Pediatric Cardiac Genomics Consortium) and compared them against 67 control dnSVs from unaffected siblings (Simons Foundation Autism Research Initiative).
• Scoring: Using the SuPreMo pipeline, they generated reference and alternate DNA sequences for each SV. CardioAkita predicted contact maps for both, and the difference was quantified as a "disruption score" (MSE and 1-CORR).
• Hypothesis Testing: They tested if disruption scores correlated with CHD severity (conotruncal/outflow tract defects vs. non-severe).

C. Discovery & Experimental Validation

• WGS Discovery: Whole-genome sequencing was performed on three new CHD trios (proband + parents) with no known exonic mutations. Six dnSVs were identified.
• Prioritization: CardioAkita was used to score these variants, identifying the most disruptive SVs in each trio.
• Genome Editing: The top disruptive SVs from each trio were engineered into WTC11 iPSCs using CRISPR/Cas9:
  • Proband A: 68-bp non-coding deletion (chr15).
  • Proband B: 84-kb deletion (chr1).
  • Proband C: 58-kb deletion (chr15, overlapping MESP1/MESP2).
• Assays:
  • Capture-C: To validate predicted changes in chromatin interactions at specific developmental stages (D0, D2, D15).
  • RNA-seq: To measure global and local gene expression changes.
  • Heterozygous vs. Homozygous: Both homozygous and heterozygous lines were generated to assess dosage effects.

3. Key Results

A. Model Performance

• CardioAkita accurately predicted cell-type-specific chromatin folding. It outperformed the original Akita (trained on hESCs) in predicting interactions in differentiating cardiomyocytes, correctly identifying gains/losses of contacts at loci like CHRDL2 and KCNE3.

B. Cohort-Wide Association

• There was a positive association between CHD severity and CardioAkita disruption scores.
• Probands with severe CHD had significantly higher median disruption scores than controls.
• Nearly all variants in the top 5% most disruptive were from the severe CHD group, suggesting that SVs causing major 3D genome reorganization are more likely to cause severe phenotypes.

C. Experimental Validation of Specific Variants

The engineered cell lines confirmed CardioAkita's predictions:

1. Proband A (68-bp Deletion):

   • Prediction: A small intergenic deletion would disrupt a CTCF-bound insulator, merging two chromatin domains.
   • Validation: Capture-C confirmed domain merging.
   • Outcome: Significant upregulation (up to 32-fold) of flanking genes (ISLR2, ISLR, STRA6) and widespread dysregulation of genes involved in left-right asymmetry and retinoic acid signaling.

2. Proband B (84-kb Deletion):

   • Prediction: Merging of two insulated domains (COX20 and HNRNPU), leading to ectopic interactions.
   • Validation: Confirmed domain merging and strengthened loops.
   • Outcome: Downregulation of genes involved in left-right axis specification (LEFTY2, TFAP2A) and upregulation of vascular development genes, suggesting impaired early patterning.

3. Proband C (58-kb Deletion):

   • Prediction: Deletion of MESP1/MESP2 (key cardiac progenitor genes) and merging of domains affecting KIF7.
   • Validation: Confirmed loss of MESP1/2 and altered loop anchors.
   • Outcome: Both homozygous and heterozygous lines showed significant transcriptional changes, including upregulation of cardiac transcription factors (TBX1, TBX5) and downregulation of early developmental genes. This confirmed the variant is haploinsufficient.

4. Key Contributions

• CardioAkita: The first deep learning model trained specifically on differentiating cardiomyocytes to predict 3D genome organization from DNA sequence.
• Mechanistic Insight: Provided the first cohort-wide evidence that CHD-associated SVs disrupt 3D genome folding, linking non-coding variants to disease severity via chromatin reorganization.
• Diagnostic Pipeline: Demonstrated a successful workflow (WGS $\rightarrow$ CardioAkita scoring $\rightarrow$ CRISPR validation) to identify pathogenic non-coding SVs that would otherwise be classified as "Variants of Uncertain Significance" (VUS).
• Haploinsufficiency: Showed that even heterozygous deletions of regulatory regions can cause profound 3D structural changes and gene expression dysregulation relevant to CHD.

5. Significance

This study bridges the gap between computational genomics and experimental biology in the context of congenital heart disease.

• Clinical Impact: It offers a scalable method to prioritize non-coding SVs in undiagnosed CHD patients, potentially solving cases where standard exome sequencing fails.
• Biological Understanding: It reinforces the hypothesis that distal regulatory disruption is a primary mechanism for SV-induced developmental disorders, not just gene dosage changes.
• Future Directions: The authors suggest that such models can be expanded to other tissues and developmental stages, improving the interpretation of clinical whole-genome sequencing data and accelerating the discovery of causal variants for complex genetic diseases.

Limitations noted: The model relies on Hi-C data from relatively immature cardiomyocytes (though adult heart failure often reverts to a fetal phenotype), and the study primarily used homozygous edits for validation, though heterozygous effects were also tested for the 58-kb deletion.