Identifying Inheritance Patterns of Allelic Imbalance, using Integrative Modeling and Bayesian Inference

This paper presents a Bayesian integrative modeling approach that leverages joint inference across family trios to simultaneously improve the estimation of allelic imbalance and identify its mode of inheritance, thereby enhancing the detection of causal regulatory variants and their impact on phenotypic traits.

The Gist

The Big Picture: Solving the Genetic "Whodunit"

Imagine you are a detective trying to solve a mystery: Why is a specific person sick or have a certain trait?

In the world of genetics, the "suspects" are tiny changes in our DNA called mutations. Some mutations are obvious (like a typo that breaks a word), but many are sneaky. They are like subtle changes in the instructions for how loud a machine should run, rather than breaking the machine itself. These are called cis-regulatory variants.

The problem is that for rare mutations, a person usually only has one bad copy of a gene (they inherited it from one parent, while the other parent gave them a healthy copy). In a healthy person, both copies of a gene usually work equally well, like two identical speakers playing the same song at the same volume. But if one copy is broken, the "volume" from that side drops. This is called Allelic Imbalance (one side is louder than the other).

The challenge for scientists has been: How do we prove this imbalance is real, and how do we know which parent passed it down?

The Old Way vs. The New Way

The Old Way (The Solo Detective):
Previously, scientists looked at one person at a time. They would check if the volume was uneven. But this is like trying to hear a whisper in a noisy room. It's hard to tell if the whisper is real or just random noise. Also, they couldn't easily tell if the "bad volume" came from Mom or Dad, or if it was a brand new mistake that happened only in the child.

The New Way (The Trio Detective Team):
This paper introduces a new tool called TrioBEASTIE. Instead of looking at one person, it looks at a family trio: Mom, Dad, and the Child.

Think of it like a relay race.

• Mom and Dad are the runners passing the baton (the genes) to the Child.
• The new model looks at the whole race at once. It uses the parents' data to help "fill in the blanks" for the child.

If Mom has a "quiet" gene copy, and the Child also has a "quiet" copy, the model can say with high confidence: "Aha! The child inherited the quiet gene from Mom!"

Even if the Child doesn't have enough data to prove it on their own (like a runner who dropped the baton), the model can look at Mom and Dad and say, "Based on what we see in the parents, there's a 90% chance the child inherited the quiet gene."

How the Model Works (The "Magic" Ingredients)

The authors built a Bayesian Model. In plain English, this is a mathematical system that acts like a super-smart weather forecaster.

1. Gathering Clues: It looks at the "read counts" (the volume levels) from RNA sequencing.
2. The 11 Scenarios: The model considers 11 different ways a family could be affected.
   • Scenario A: No one is sick.
   • Scenario B: Only Mom is sick, and she passed it to the child.
   • Scenario C: Only Mom is sick, but she didn't pass it to the child.
   • Scenario D: A brand new mistake happened in the child (a "de novo" mutation).
   • Scenario E: A "recombination" happened (like a mix-up in the middle of the race where the child got the healthy baton but the sick instructions).
3. The Verdict: Instead of just saying "Yes" or "No," it gives a probability score. It says, "There is a 95% chance this is Scenario B, and a 5% chance it's Scenario C."

Why This is a Game-Changer

1. It borrows strength:
Imagine you are trying to guess the weight of a watermelon. If you lift it alone, it's hard to tell. But if you lift it with your mom and dad, and you know how much they weigh, you can guess the watermelon's weight much more accurately. TrioBEASTIE does this with genes. It combines data from the whole family to make a stronger, more accurate guess.

2. It finds the "Smoking Gun":
The researchers tested this on real families (the famous CEPH trios). They found genes where the volume was uneven. But they didn't stop there. They also looked at ATAC-seq data, which is like checking the openness of the door to the gene.

• If the "door" (chromatin) is closed on one side, the gene can't be read, so the volume drops.
• By linking the "closed door" (ASA) to the "low volume" (ASE), they could pinpoint the exact genetic typo causing the problem.

3. It's better than guessing:
When they tested their model against old methods using computer simulations, TrioBEASTIE was much better at finding the right answer, especially when the data was messy or the "volume difference" was small.

The Real-World Impact

Why should you care?

• Rare Diseases: Many rare diseases are caused by these subtle "volume" changes that old tools miss. This tool helps doctors find the cause faster.
• Personalized Medicine: As DNA sequencing gets cheaper, we will have more family data. This tool helps us make sense of that data, turning a pile of numbers into a clear story about why a person is sick.
• Future Proofing: The authors show that this method works not just for RNA (the message), but also for chromatin (the packaging). This means it can be used to study many different types of genetic regulation.

Summary Analogy

Imagine a choir where one singer is off-key.

• Old Method: You listen to one singer and try to guess if they are off-key, but the room is noisy. You might be wrong.
• TrioBEASTIE: You listen to the Mom, Dad, and Child together. You hear that Mom is off-key, and the Child is off-key in the exact same way. You instantly know the Child inherited the off-key voice. Even if the Child is singing quietly, you know the problem is inherited, not a random mistake.

This paper gives scientists a powerful new pair of ears to hear the subtle genetic whispers that cause disease.

Technical Summary

1. Problem Statement

Interpreting the phenotypic impact of novel mutations, particularly cis-regulatory variants, remains a significant challenge in genomics. For rare variants, individuals often possess only one affected copy of a causal allele, leading to Allelic Imbalance (AI) or Allele-Specific Expression (ASE).

• Limitations of Current Methods: Existing ASE detection methods primarily focus on single individuals using Null Hypothesis Significance Testing (NHST) (e.g., binomial tests). These approaches treat individuals as independent, failing to leverage shared genetic information within families.
• The Gap: While co-segregation analysis in families (trios) is powerful for identifying causal variants, there is a lack of formal statistical frameworks that jointly model ASE across related individuals to infer inheritance patterns, estimate effect sizes with uncertainty, and distinguish between modes of inheritance (e.g., inherited vs. de novo, recombination events).

2. Methodology: TrioBEASTIE

The authors propose TrioBEASTIE (Trio-aware Bayesian Estimation of Allele-Specific Transcription by Integrating Evidence), a novel probabilistic graphical model designed to analyze RNA-seq (and ATAC-seq) data across a mother-father-offspring trio.

Core Architecture

• Bayesian Framework: Unlike NHST methods, TrioBEASTIE uses Bayesian inference to assess evidence for both the null hypothesis (no ASE) and alternative hypotheses (ASE present) simultaneously. It incorporates prior probabilities to stabilize effect size estimates and mitigate false positives.
• Joint Modeling: The model jointly estimates the affected status (presence of ASE on a specific haplotype) and the effect size ($\theta$) across all three individuals in a trio.
• Modes of Inheritance: The model evaluates 11 distinct modes of inheritance (Fig. 1), including:
  • Null: All individuals unaffected.
  • Simple Inheritance: One parent is affected; the child either inherits or does not inherit the affected allele.
  • De Novo: Only the child is affected (new mutation).
  • Recombination: A recombination event occurs in an affected parent between the unobserved causal element and the gene body, causing the child to inherit the gene allele associated with full expression but the causal variant associated with reduced expression (or vice versa).
• Latent Variables: The model uses latent variables ($V_{i,k}$) to represent the affected status of individual $i$ and gene copy $k$. It assumes at most one affected copy per individual and at most one affected parent.
• Implementation: Implemented in Stan using Hamiltonian Monte Carlo (HMC) sampling via RStan.

Data Inputs

• Read Counts: Allele-specific read counts from heterozygous sites.
• Phasing: Utilizes trio phasing to determine which allele was transmitted from parent to child.
• Effect Size ($\theta$): Represents the odds ratio of read counts between the two alleles. The model assumes a shared $\theta$ across affected individuals within a specific mode.

3. Key Contributions

1. Joint Inference across Trios: The first method to explicitly model Mendelian transmission probabilities, de novo mutation rates, and recombination events to infer ASE inheritance patterns across a family unit.
2. Borrowing Strength: By aggregating information across related individuals, the model improves statistical power, allowing for ASE detection in individuals with low read depth or no heterozygous sites (by inferring status from parents).
3. Uncertainty Quantification: Provides full posterior probability distributions for effect sizes and posterior probabilities for each mode of inheritance, enabling users to rank inheritance modes by confidence.
4. Multi-Omics Applicability: Demonstrates that the framework is not limited to RNA-seq but is equally applicable to ATAC-seq for detecting Allele-Specific Accessibility (ASA), linking chromatin state imbalance to gene expression imbalance.

4. Results

Simulation Studies

• Accuracy: TrioBEASTIE outperformed a baseline "independence model" (which analyzes individuals separately) and a standard binomial test.
  • It achieved higher Area Under the Curve (AUC) scores, particularly when effect sizes ($\theta$) were extreme or read depths were low.
  • It successfully distinguished between the 11 modes of inheritance, with confusion primarily occurring only when data was insufficient (leading to conservative "null" calls).
• Effect Size Estimation: The model provided more accurate estimates of $\theta$ (lower Root Mean Squared Error) compared to the independence model, especially when leveraging data from multiple affected individuals.
• Robustness: The model correctly identified de novo events and recombination scenarios when sufficient data was available.

Application to Real Data (CEPH 1463 Trios)

The model was applied to RNA-seq and ATAC-seq data from two genetically independent trios (NA12878 and NA12877).

• ASE Detection:
  • Identified 132 ASE genes in the NA12878 trio and 200 in the NA12877 trio.
  • Found 5 genes (e.g., RPL29P4, CBX3, SLFN5) showing ASE in both trios with 100% marginal probability, suggesting shared biological mechanisms or variants.
  • Top genes showed significant log2 fold changes (e.g., IGKV3 with 5.17).
• Linking ASE to ASA (Causal Variant Discovery):
  • Applied the model to ATAC-seq data to detect Allele-Specific Accessibility (ASA).
  • Identified 6 peak-gene links where both the regulatory element (peak) and the gene showed concordant ASE/ASA.
  • Case Study (TBC1D4): An intronic ASA peak overlapping an ENCODE cCRE was linked to ASE in the TBC1D4 gene. The peak contained a known eQTL (rs1560540) and showed concordant inheritance with the gene's expression imbalance, strongly suggesting a causal regulatory variant.
• Validation: Enrichment analysis showed a trend (though not statistically significant in this small sample) toward higher prevalence of stop-gain variants in ASE genes, validating the biological plausibility of the findings.

5. Significance and Future Directions

• Clinical Utility: As sequencing costs decrease, TrioBEASTIE offers a powerful tool for routine analysis of family trios to identify causal non-coding variants and interpret rare genetic diseases.
• Mechanistic Insight: By integrating RNA-seq and ATAC-seq, the method moves beyond detecting that imbalance exists to suggesting why (e.g., via a specific cis-regulatory variant affecting chromatin accessibility).
• Scalability: While currently designed for trios, the Bayesian framework is flexible enough to be extended to larger pedigrees, which would further improve phasing and inference accuracy.
• Limitations & Future Work: The current model assumes uniform effect sizes across affected individuals and at most one affected parent. Future iterations could incorporate hierarchical modeling to allow variable effect sizes and negative binomial likelihoods to handle over-dispersion and include unrelated controls.

Conclusion: TrioBEASTIE represents a significant advancement in the statistical genetics of gene regulation. By shifting from single-individual analysis to joint Bayesian inference across families, it provides a more robust, accurate, and interpretable framework for uncovering the genetic architecture of allelic imbalance and its contribution to phenotypic traits.