Neretva: Neural Variational Inference for Allele-level Genotyping of Highly Polymorphic Genes

Neretva is an open-source framework that utilizes auto-encoding variational Bayes to efficiently and accurately perform allele-level genotyping and phasing of highly polymorphic gene families like CYP and KIR, outperforming current state-of-the-art methods in scalability and accuracy.

The Gist

The Big Picture: The "Genetic Jigsaw Puzzle" Problem

Imagine your DNA is a massive library of instruction manuals. Most of these manuals are standard and easy to read. However, some sections of the library are incredibly messy. They contain thousands of nearly identical copies of the same book, with tiny differences in punctuation (like a comma instead of a period) that completely change the meaning of the sentence.

In the world of medicine, two specific families of these "books" are critical:

1. CYP (Cytochrome P450): These are the body's "drug metabolism" manuals. They tell your liver how to break down medicines. If you get the recipe wrong, a drug might not work, or it might poison you.
2. KIR (Killer-cell Immunoglobulin-like Receptors): These are the "security guard" manuals for your immune system. They decide if your body attacks a virus, a tumor, or a new organ transplant.

The Problem:
When scientists try to read these specific books using modern sequencing machines (High-Throughput Sequencing), it's like trying to solve a jigsaw puzzle where:

• Half the pieces look exactly the same.
• Some pages are missing (deletions).
• Some pages are duplicated (copy number variations).
• Some pieces from Book A accidentally got glued into Book B.

Current computer programs try to solve this by brute-forcing every possible combination. It's like trying to solve a 10,000-piece puzzle by trying every single piece in every single spot. It works for small puzzles, but for these messy gene families, it's too slow, too rigid, and often gets stuck.

The Solution: Enter "Neretva"

The authors created a new tool called Neretva. Instead of brute-forcing the puzzle, Neretva acts like a super-smart, intuitive detective who uses probability and "gut feeling" (mathematically speaking) to solve the mess.

Here is how Neretva works, broken down into simple steps:

1. The "Shadow" Detective (Handling Ambiguity)

In the messy gene families (especially KIR), a piece of DNA might look like it belongs to Gene A, but it could actually be a "shadow" cast by Gene B.

• Old Way: The computer gets confused and says, "I don't know, maybe it's A, maybe it's B."
• Neretva's Way: It says, "Okay, this piece looks like it belongs to A, but it's casting a shadow on B. Let's assume it could be either, and we'll figure out the most likely story as we go." It doesn't throw away confusing data; it uses it to learn.

2. The "Recipe" Generator (Variational Inference)

Instead of checking every single possible recipe (genotype) one by one, Neretva builds a probabilistic model.

• Imagine you are trying to guess a secret recipe based on the smell of the kitchen. You don't taste every possible combination of ingredients. Instead, you use a mental model to say, "It smells like garlic and onions, so there's a 90% chance it's a garlic-onion soup, and a 10% chance it's a garlic-onion stew."
• Neretva does this with DNA. It creates a "soup" of possibilities and slowly refines the recipe until it finds the most accurate one. It uses Neural Networks (AI) to learn the patterns of how these genes usually look, making it incredibly fast.

3. Counting the Copies (Copy Number Estimation)

Sometimes, a person has two copies of a gene, and sometimes they have five.

• Neretva looks at how much "coverage" (how many times a specific part of the gene was read by the machine) it sees.
• If it sees double the usual amount of "garlic smell," it knows there are likely two copies of the garlic gene. It uses a mathematical "regression" (a fancy way of fitting a line to data) to count exactly how many copies exist, even if the data is noisy.

4. The "Focus" Filter (Core vs. Minor Variants)

Not all differences in the DNA matter. Some are just typos that don't change the meaning (silent variants).

• Neretva has a special filter that says, "Ignore the tiny typos; focus on the words that actually change the meaning of the sentence."
• It uses a mathematical penalty system to ensure the final answer makes biological sense, prioritizing the "important" changes over the noise.

Why is this a Big Deal?

The paper tested Neretva against the current best tools (like Aldy, Geny, and Cyrius) on real human data.

• Speed: While other tools might take an hour to solve a complex KIR puzzle, Neretva does it in minutes. It's like switching from a hand-cranked calculator to a supercomputer.
• Accuracy: On the messy KIR genes, Neretva was the most accurate tool, correctly identifying the genetic makeup in 91% of cases, beating the previous leaders.
• Flexibility: Because it's built on AI (Neural Networks) rather than rigid rules, it can adapt to new types of data or new gene families without needing to be completely rewritten.

The Bottom Line

Neretva is a new, open-source software that helps doctors and researchers read the most confusing parts of our DNA. By using AI to guess the most likely genetic story rather than trying to force a perfect fit, it solves the "messy gene" problem faster and more accurately than ever before.

This matters because getting these genes right is the difference between a patient getting the right dose of medication, a successful organ transplant, or a correct diagnosis for an autoimmune disease. Neretva makes precision medicine more reliable for everyone.

Technical Summary

1. Problem Statement

Accurate genotyping and phasing of highly polymorphic gene families (e.g., CYP pharmacogenes and KIR immune receptors) are critical for precision medicine. However, current methods face significant computational and biological challenges:

• Sequence Similarity: Extreme similarity between related genes (paralogs) and high intra-gene allelic identity (>98%) create ambiguity in assigning sequencing reads to their true origin.
• Structural Complexity: These families exhibit extensive Copy Number Variation (CNV), gene deletions, duplications, and complex structural rearrangements (e.g., fusions).
• Limitations of Current Methods:
  • Combinatorial Optimization (ILP): Tools like Aldy and StellarPGx rely on Integer Linear Programming. While theoretically optimal, ILP is NP-hard, scales poorly to complex families (like KIR), often requires proprietary solvers (limiting privacy-preserving applications), and lacks flexibility in handling uncertainty.
  • Likelihood-Based Approaches: These often suffer from convergence to local optima and lack scalability.
  • Scalability: Existing tools struggle with the combinatorial explosion required for gene families with high diversity (e.g., the 17-gene KIR cluster).

2. Methodology: The Neretva Framework

Neretva reframes the genotyping problem from combinatorial optimization to a probabilistic latent variable model solved via Auto-Encoding Variational Bayes (AEVB). This allows for scalable, gradient-based optimization.

A. Data Preprocessing & Database Construction

• Input: High-throughput sequencing (HTS) reads (BAM/CRAM/FASTQ) and allele databases (PharmVar for CYP, IPD-KIR for KIR).
• Alignment: Reads are aligned to known allele sequences. For complex clusters (like KIR), an "all-to-all" alignment is used.
• Shadow Variant Handling: To address cross-gene mapping artifacts, Neretva constructs an extended variant set. If a read maps to a gene $G'$ but supports a variant from gene $G$, that variant is added to $G'$'s potential set as a "shadow variant." This prevents false negative allele calls due to mapping ambiguity.
• Bias Correction: A Bias Factor Matrix ($H$) is introduced to model sequencing biases and mapping artifacts, adjusting expected coverage for specific variants.

B. Generative Model

The model assumes the observed base count matrix $Y$ is generated by:

1. Allele Proportions ($\theta$): A latent variable representing the relative abundance of $K$ candidate alleles, modeled via a Logistic Normal distribution.
2. Base Emissions ($\Psi$): Latent variables representing the probability of observing specific bases at specific positions for each allele.
3. Observation: The observed counts follow a Multinomial distribution based on the product of allele proportions, base emissions, and bias factors.

C. Inference & Optimization

• Variational Inference: Since the marginal likelihood is intractable, Neretva uses a mean-field variational approximation to maximize the Evidence Lower Bound (ELBO).
• Loss Function Augmentation: To ensure biological relevance, the standard ELBO is augmented with two specific terms:
  1. Jensen-Shannon Divergence ($D_{JS}$): Penalizes discrepancies between the empirical distribution of core (functional) variants and the model's reconstructed distribution. This ensures the model prioritizes variants that define the phenotype.
  2. Entropy Penalty: Encourages the base emission distributions to be sharp (low entropy), reflecting the biological reality that a specific genomic position usually has a single dominant base, minimizing the impact of sequencing noise.
• Optimization: The model is trained using the Adam optimizer and the reparameterization trick to compute gradients efficiently on GPUs.

D. Copy Number Estimation

Before final allele calling, Neretva estimates gene copy numbers by formulating the problem as a coverage reconstruction task. It minimizes the Huber loss between observed read depths and expected depths derived from a linear combination of unknown gene copy numbers.

3. Key Contributions

• Novel Framework: Introduction of Neretva, the first unified framework using AEVB for genotyping both pharmacogenes (CYP) and immune receptors (KIR).
• Scalability: By replacing NP-hard ILP with gradient-based neural inference, Neretva achieves orders-of-magnitude speedups, making it feasible to genotype complex families like KIR in minutes rather than hours.
• Handling Ambiguity: The "shadow variant" extension and bias matrix allow the model to robustly handle cross-gene mapping errors common in highly similar gene clusters.
• Privacy & Flexibility: The approach avoids proprietary solvers, making it compatible with privacy-preserving technologies (e.g., homomorphic encryption) and easily extensible to new gene families.

4. Results

Neretva was benchmarked against state-of-the-art tools (e.g., Aldy, Cyrius, StellarPGx, Geny, T1K) on two datasets:

1. Pharmacogenes (CYP): 70 GeT-RM Illumina WGS samples.
   • CYP2C19/8/9: Achieved 100% accuracy (F1 = 1.0), matching the best existing tools.
   • CYP2D6: Achieved an F1 score of 0.993 (sample-level) and 0.994 (allele-level), performing on par with Aldy and slightly below Cyrius (which had perfect accuracy but is less flexible).
2. KIR Genes: 40 Human Pangenome Reference Consortium (HPRC) samples.
   • Sample-Level: Neretva achieved the highest accuracy (0.842) and F1 score (0.908), significantly outperforming Geny (0.796 accuracy) and T1K (0.520 accuracy).
   • Allele-Level: Achieved an F1 score of 0.912, correctly identifying 682/749 alleles, surpassing Geny (0.884) and T1K (0.732).
3. Computational Efficiency:
   • Neretva: < 20 minutes for KIR (inference < 8 mins).
   • ILP Comparison: A comparable ILP model failed to converge to an optimal solution within 30 minutes (4x the Neretva runtime) and produced strictly worse results.

5. Significance

• Clinical Impact: Neretva provides a fast, accurate, and interpretable solution for genotyping complex gene families essential for drug metabolism (CYP) and transplant compatibility/immune response (KIR).
• Paradigm Shift: It demonstrates that deep learning-based variational inference can effectively replace rigid combinatorial optimization in genomics, offering a better balance between scalability, uncertainty modeling, and biological interpretability.
• Future Directions: The framework is designed to be extensible to long-read sequencing (Nanopore), other gene families (HLA, TCR), and pangenome graph references, potentially revolutionizing how complex genomic loci are analyzed in precision medicine.