Quantification of the effects of single nucleotide variants in NKX2.1 transcription factor binding sites

This study employs EMSA-seq and neural network modeling to quantify how single nucleotide variants affect NKX2.1 transcription factor binding, providing a framework to identify regulatory mutations that may cause CAHTP in patients lacking coding region mutations.

The Gist

The Big Picture: The "Lock and Key" Problem

Imagine your body is a massive, bustling city. The DNA in your cells is the City Blueprint, containing all the instructions for how to build and run everything.

But a blueprint is useless without a foreman to read it. In biology, these foremen are called Transcription Factors (TFs). One very important foreman is named NKX2.1. NKX2.1 is like a specialized construction manager who only shows up to build specific neighborhoods: the Thyroid (which controls metabolism), the Lungs, and parts of the Brain.

NKX2.1 works by finding specific "addresses" on the DNA blueprint (called binding sites) and locking onto them to say, "Start building here!"

The Problem:
Sometimes, people get sick with a condition called CAHTP (which causes thyroid issues, lung problems, and movement disorders). Usually, doctors find the culprit by looking at the "coding" part of the blueprint—the part that builds the foreman (NKX2.1) itself. If the foreman is broken, the city stops working.

However, in about 73% of patients, the foreman (NKX2.1) looks perfect. The problem isn't the foreman; it's the address on the blueprint. A tiny typo (a single letter change) in the address where the foreman is supposed to lock on might be preventing him from finding the job site.

Until now, we didn't have a good way to spot these tiny typos in the addresses. This paper is about building a super-smart detector to find them.

---

The Experiment: The "Speed Dating" for DNA

To understand how NKX2.1 reads these addresses, the scientists needed to test millions of variations. They couldn't do this one by one; it would take forever. So, they used a clever trick called EMSA-seq.

The Analogy: The Speed Dating Event
Imagine a massive speed dating event.

• The Foreman (NKX2.1): He is the guest of honor.
• The Dates (DNA Sequences): Instead of one person, they invited millions of different DNA sequences to the party. Some have the perfect address, some have a typo, and some are completely wrong.
• The Match: The foreman walks around and shakes hands (binds) with the DNA sequences he likes.
• The Result: The scientists take a photo of who he shook hands with. They then use a high-tech scanner (sequencing) to count exactly how many times he shook hands with each specific DNA address.

This allowed them to see, in one go, which typos made the foreman say, "No thanks," and which ones he still liked.

The Brain: Training a "Digital Foreman"

Once they had the data from the speed dating, they needed a way to predict what would happen with new addresses they hadn't tested yet. They built an Artificial Intelligence (AI) model—a digital brain.

The Analogy: Learning to Read a Language
Think of the DNA sequence as a language. The scientists taught the AI to read this language by showing it the results of the speed dating.

• They showed the AI: "Here is a perfect address. The foreman loved it."
• They showed the AI: "Here is an address with a 'G' instead of an 'A'. The foreman hated it."
• They showed the AI: "Here is an address where two letters changed. The foreman was confused."

The AI (a Neural Network) learned the complex grammar of this language. It figured out that it's not just about one letter; sometimes, the combination of letters matters. It learned that the "context" (the letters surrounding the main address) changes how the foreman feels.

The Surprise:
The AI was so smart that even when they only showed it a small part of the address (the "core"), it could guess the importance of the surrounding letters because it learned the "vibe" of the whole neighborhood.

The Reality Check: Does the AI Work in the Real World?

The scientists didn't just trust the AI. They tested it three different ways to make sure it wasn't just guessing.

1. The "One-on-One" Test (MST):
   They took the AI's predictions and compared them to a very precise lab test where they measured how tightly the foreman held onto a single DNA strand.

   • The Twist: The AI and the precise test didn't always agree perfectly. Why? Because the "Speed Dating" (EMSA-seq) was a competitive environment. The foreman had to choose between millions of options at once. In the real body, the foreman is also competing against millions of other DNA strands. The AI learned this "competition" better than the isolated lab test did.

2. The "X-Ray Vision" Test (Crystallography):
   They took a snapshot of the foreman actually holding the DNA using X-ray crystallography.

   • The Result: The X-ray pictures showed exactly how the foreman's hands touched the DNA. When they looked at the AI's "brain map" (what it thought was important), it matched the X-ray pictures perfectly! The AI knew exactly which letters the foreman was touching, even though it had never seen an X-ray before.

3. The "City Map" Test (ChIP-seq):
   Finally, they asked the AI to look at real maps of the human body (genomic data from living cells) to find where the foreman actually lives.

   • The Result: The AI was excellent at finding the foreman's real addresses in the messy, complex city of the human genome. It was better than the old, simple methods (like looking for a single keyword) because it understood the whole sentence, not just the word.

Why Does This Matter?

The "Missing Puzzle Piece"
For years, doctors have been looking at patients with CAHTP, finding that their "foreman" (NKX2.1) is perfect, but they still can't explain why the patient is sick. They were missing the puzzle piece.

This paper provides a magnifying glass for that missing piece.

• If a patient has a genetic typo in the "address" where NKX2.1 is supposed to bind, this new AI tool can tell the doctor: "This typo is the problem. It's breaking the lock."
• This means we can finally diagnose patients who were previously "unsolved" cases.

Summary in a Nutshell

1. The Issue: Some diseases are caused by typos in the "addresses" on our DNA, not the "workers" themselves.
2. The Method: The scientists ran a massive "speed dating" event to see which DNA addresses a specific worker (NKX2.1) likes.
3. The Tool: They trained an AI to learn the rules of these addresses.
4. The Proof: They proved the AI works by comparing it to X-ray photos and real-world data.
5. The Future: Doctors can now use this AI to find the hidden typos causing diseases in patients who previously had no answers.

It's like upgrading from a simple spell-checker to a genius editor that understands the meaning of the sentence, helping us fix the typos that cause our bodies to malfunction.

Technical Summary

1. Problem Statement

• Clinical Context: Mutations in the coding region of the NKX2-1 gene cause CAHTP (Choreoathetosis, Congenital Hypothyroidism, with or without Pulmonary Dysfunction). However, in a cohort of 101 patients with the characteristic phenotype, 73% remained genetically unsolved, lacking coding mutations.
• Hypothesis: The authors hypothesize that pathogenic variants reside in regulatory elements (promoters/enhancers) rather than coding regions. Specifically, Single Nucleotide Variants (SNVs) within Transcription Factor Binding Sites (TFBSs) may disrupt NKX2.1 binding, altering gene regulation and causing disease.
• Technical Gap: Existing models for predicting TF binding, such as Position Weight Matrices (PWMs), assume nucleotide independence and fail to capture complex interdependencies (epistasis) or dinucleotide effects. Furthermore, there is a lack of high-throughput in vitro binding data for human NKX2.1 to train more sophisticated models.

2. Methodology

The study employed a multi-modal approach combining high-throughput experimental assays, structural biology, and deep learning.

A. High-Throughput Binding Assay: EMSA-seq

• Technique: The authors adapted Electromobility Shift Assay sequencing (EMSA-seq). Unlike HT-SELEX, EMSA-seq uses a single round of binding, allowing for the detection of low-affinity sites.
• Libraries: Three mutant libraries were designed based on the rat thyroglobulin promoter NKX2.1 binding site (24 bp total):
  1. CORE: Randomization of the 4 bp core motif (CAAG).
  2. FLANK: Randomization of the 10 bp flanking regions (keeping the core constant).
  3. ALL: Randomization of the entire 14 bp region.
• Process: Libraries were incubated with recombinant NKX2.1 DNA-binding domain (DBD)-GFP. Bound DNA was separated from unbound DNA via gel electrophoresis, excised, and sequenced.
• Analysis: DESeq2 was used to calculate the Log2 Fold Change (LFC) of sequence enrichment in the bound fraction versus the unbound reference.

B. Deep Learning Model Training

• Architecture: A VCNNBPNet (Variable Convolutional Neural Network based on BPNet) was developed.
  • Input: One-hot encoded DNA sequences (24 bp).
  • Structure: Stacked dilated convolutions to capture multi-scale motifs and long-range dependencies, followed by adaptive max pooling and dense layers.
  • Training: Models were trained to predict the LFC derived from EMSA-seq data.
  • Interpretability: DeepSHAP (Shapley values) and in silico saturation mutagenesis were used to attribute importance to specific nucleotides and identify epistatic interactions.

C. Validation Methods

1. Microscale Thermophoresis (MST): Measured absolute dissociation constants ($K_d$) for 29 specific variants to validate binding affinity.
2. X-ray Crystallography: Solved high-resolution structures (PDB: 9U18, 9U19) of NKX2.1 DBD bound to wildtype and variant (CACG) DNA to visualize molecular interactions.
3. AlphaFold: Used to predict structures for non-crystallized variants and validate the experimental findings.
4. ChIP-seq Validation: Tested model performance on in vivo data by classifying NKX2.1 ChIP-seq peaks against negative control TFs (GATA1, MYOD1, etc.).

3. Key Results

Experimental Binding Data

• EMSA-seq Sensitivity: Successfully quantified binding enrichment for millions of sequences. The wildtype CAAG core was the most enriched.
• Discrepancy with MST: There was no significant correlation between EMSA-seq LFC predictions and MST $K_d$ values ($r \approx 0$).
  • Explanation: EMSA-seq is a competitive assay (many sequences competing for limited protein), whereas MST measures isolated binary interactions. The competitive nature of EMSA-seq likely detects subtle affinity differences and "relative" binding strengths that MST misses.
  • Support: The EMSA-seq models correlated much better with Selective Chromatography (SC) data (another competitive method, $r \approx 0.76$) than with MST.

Deep Learning Performance

• Model Accuracy: The FLANK model (10 bp variation) achieved the best balance of complexity and data density, outperforming the CORE and ALL models.
• Epistasis: The models successfully learned non-additive interactions (epistasis) between nucleotides, capturing dependencies that PWMs and TFFMs (which only consider dinucleotides) cannot.
• Attribution: DeepSHAP analysis revealed that the models correctly identified the importance of the core motif and specific flanking nucleotides, even in regions held constant during training, indicating the models learned the full sequence context.

Structural Insights

• Crystallography: The wildtype structure showed the N-loop anchoring in the minor groove and the recognition helix (H3) in the major groove.
• Mechanism of Mutation: The CAAG $\to$ CACG mutation caused a subtle shift ($\le$ 0.5Å) in the R165 side chain, displacing the N-loop and disrupting the hydrogen-bonding network. This structural plasticity explained the altered binding observed in functional assays.
• AlphaFold Validation: AlphaFold predictions for non-crystallized variants (e.g., T $\to$ G substitution) correctly predicted conformational shifts that aligned with the deep learning model's prediction of reduced binding, despite conflicting with some MST data.

In Vivo Validation (ChIP-seq)

• Classification: The FLANK model outperformed both the CORE/ALL models and the traditional PWM (FIMO) in distinguishing NKX2.1 ChIP-seq peaks from negative controls.
• Context Matters: Models performed better when analyzing the entire 500 bp peak rather than a sliding window, suggesting that long-range sequence context is crucial for accurate in vivo binding prediction.
• Peak Centering: The FLANK model correctly localized the highest predicted binding scores to the center of ChIP-seq peaks, a hallmark of true TF binding sites.

4. Key Contributions

1. First Public In Vitro Dataset: Generated the first publicly available, high-throughput in vitro binding data for human NKX2.1 covering millions of DNA sequences.
2. Novel Deep Learning Framework: Developed and validated a VCNNBPNet architecture capable of modeling complex, non-linear nucleotide dependencies and epistasis in TF binding.
3. Methodological Insight: Demonstrated that competitive binding assays (EMSA-seq) provide a more biologically relevant metric for predicting in vivo binding than isolated affinity measurements (MST), likely due to the competitive environment mimicking the cellular nucleus.
4. Clinical Utility: Provided a tool to prioritize non-coding variants in patients with CAHTP who lack coding mutations, potentially solving a significant portion of "unsolved" genetic cases.

5. Significance

This study bridges the gap between molecular biophysics, structural biology, and clinical genomics. By proving that deep learning models trained on competitive in vitro data can accurately predict in vivo binding sites, the authors provide a robust pipeline for interpreting Whole Genome Sequencing (WGS) data in rare diseases.

The work highlights that regulatory variants are a major source of missing heritability in Mendelian disorders. The developed models allow clinicians and researchers to move beyond coding regions and assess the pathogenicity of SNVs in promoters and enhancers, specifically for NKX2.1-related disorders, with high precision. Furthermore, the findings suggest that future TF binding studies should prioritize competitive, high-throughput methods over isolated affinity measurements to better reflect biological reality.