CoNVict: An Agentic AI System for Copy Number Variation Prioritization in Rare Disease Diagnosis

The paper introduces CoNVict, a two-stage agentic AI system that leverages large language models to automate the prioritization of copy number variants in rare disease diagnosis by integrating patient phenotypes and performing pairwise comparisons, thereby outperforming existing tools in identifying causal variants while bridging the gap between automated annotation and clinical reasoning.

The Gist

Imagine your body is a massive, intricate library containing billions of books (your genes). Sometimes, a section of this library gets accidentally copied twice (a duplication) or a whole shelf gets torn out (a deletion). In the world of genetics, these are called Copy Number Variants (CNVs).

For patients with rare, unexplained diseases, finding the specific "missing shelf" or "extra copy" that caused their illness is like finding a single needle in a haystack the size of a city.

Here is a simple breakdown of the paper's solution, CoNVict, using everyday analogies.

The Problem: The Overwhelmed Librarian

Currently, when a doctor gets a genetic test result, they are handed a list of thousands of these "shelf changes."

• The Old Way: Existing computer tools act like a basic barcode scanner. They scan every book and say, "This one looks suspicious," or "This one looks safe." But they can't tell the difference between a book that is suspicious for this specific patient and a book that is suspicious but irrelevant to the patient's symptoms.
• The Gap: The computer can flag the "bad" books, but it can't read the patient's medical story (their symptoms) to figure out which bad book actually explains why the patient is sick. This leaves the human doctor to do the heavy lifting of cross-referencing thousands of possibilities, which is slow and prone to error.

The Solution: CoNVict (The AI Detective)

The authors created CoNVict, an AI system that acts like a brilliant, tireless Clinical Detective. Instead of just scanning barcodes, CoNVict reads the patient's story and investigates the suspects one by one.

It works in two main stages, like a two-round tournament:

Round 1: The "Triage" (CNVerdict)

Imagine the detective receives a stack of 500 suspect files (the genetic variants).

• The Action: CoNVict reads the patient's symptoms (e.g., "seizures," "short stature") and compares them against the "criminal record" of every gene involved in the variants.
• The Decision: It sorts the files into three piles:
  1. Relevant: "This gene's history matches the patient's symptoms. Keep investigating."
  2. Abstain: "This gene is weird, but I'm not sure. Let's keep it on the back burner."
  3. Irrelevant: "This gene has nothing to do with these symptoms. Throw it out."
• The Result: The stack of 500 suspects is quickly whittled down to the top 32 most promising ones.

Round 2: The "Tournament" (Pairwise Battles)

Now the detective has a shortlist of 32 suspects. Who is the real culprit?

• The Action: CoNVict pits the suspects against each other in head-to-head battles. It asks the AI: "Between Suspect A and Suspect B, which one explains the patient's symptoms better?"
• The Logic: The AI doesn't just look at the gene; it looks at the context. It considers:
  • Does the gene break in a way that hurts the body?
  • Does the gene control parts of the body that are sick?
  • Is the gene known to be fragile?
• The Winner: The "winner" of the battle moves to the next round. This continues until the AI has ranked the top 4 suspects from most likely to least likely.

Why is this a Big Deal?

The paper tested CoNVict on hundreds of simulated patients and found it was much better than current tools at finding the "needle in the haystack."

1. It Handles the "Noise": In Whole Genome Sequencing (looking at the entire library), there are thousands of background changes that aren't causing the disease. CoNVict is great at ignoring the noise and focusing on the signal.
2. It Solves the "Mystery Cases": Many genetic variants are labeled "Variants of Uncertain Significance" (VUS)—basically, "we don't know if this is bad." CoNVict uses its reasoning skills to figure out that even if a variant is new and unknown, if it breaks a gene that matches the patient's symptoms, it's likely the culprit.
3. It Reads the Fine Print: It can even look at "non-coding" regions (the spaces between the books) which other tools often ignore, realizing that damaging the "margins" can sometimes be just as bad as damaging the "text."

The Bottom Line

CoNVict is like upgrading from a spell-checker to a senior editor. It doesn't just check if words are spelled right; it understands the story, the context, and the plot, helping doctors diagnose rare diseases faster and more accurately. It bridges the gap between raw data and human reasoning, acting as a super-powered assistant to the clinical geneticist.

Technical Summary

1. Problem Statement

Copy Number Variants (CNVs)—deletions or duplications ranging from 50 bp to megabases—are a major cause of rare genetic disorders (15–20% of neurodevelopmental cases). However, their clinical interpretation is significantly more complex than Single Nucleotide Variants (SNVs) due to:

• Multi-gene impact: A single CNV can span dozens of genes, disrupt regulatory elements, or create gene fusions.
• Context dependence: Pathogenicity depends on the dosage sensitivity of overlapping genes, the integrity of regulatory architecture, and the specific patient's phenotype.
• Limitations of current tools: Existing computational methods (e.g., Exomiser, PhenoSV, SvAnna) either rely on static, phenotype-agnostic scoring (failing to distinguish relevant from irrelevant pathogenic variants) or evaluate variants in isolation. They lack the ability to perform comparative evidence synthesis and iterative reasoning required to rank candidates within a specific clinical context. This creates a gap between high-throughput annotation and the nuanced decision-making required for diagnosis, particularly for Variants of Uncertain Significance (VUS) and non-coding variants.

2. Methodology: The CoNVict System

CoNVict is a two-stage Agentic AI system designed to prioritize CNVs by integrating structured genomic data with Large Language Model (LLM) reasoning. It moves beyond static scoring to simulate the iterative evidence evaluation of a clinical geneticist.

A. Pipeline Overview

1. Pre-filtering:

   • Input: VCF files containing CNVs.
   • Annotation: Variants are annotated using AnnotSV.
   • Initial Ranking: A pre-filtering step using CNVoyant scores retains the top 64 candidates for Whole Exome Sequencing (WES) and top 512 for Whole Genome Sequencing (WGS) to reduce computational load.

2. Stage 1: CNVerdict (Triage)

   • Goal: Classify variants into three categories: Relevant, Abstain, or Irrelevant.
   • Mechanism: An LLM (Gemini 2.5 Flash) processes a structured prompt containing:
     • Variant annotations (genomic coordinates, size, overlapping genes).
     • Gene Knowledge Cache: Pre-computed, web-retrieved keywords for each gene (derived from OMIM, ClinVar, HPO, etc.) capturing core clinical associations.
     • Patient Context: The patient's Human Phenotype Ontology (HPO) terms and a synthetic or real clinical epicrisis (summary).
   • Output: Variants are re-ranked based on the verdict, retaining the top 32 (WES) or 64 (WGS) for the next stage.

3. Stage 2: Tournament Ranking (Pairwise Comparison)

   • Goal: Perform fine-grained ranking of the remaining candidates.
   • Mechanism: A tournament-style sorting algorithm where the LLM compares pairs of CNVs.
     • Context: Comparisons are conditioned on the patient's HPO terms, clinical notes, and detailed variant reports (including regulatory targets).
     • Logic: The LLM evaluates which variant is more likely to be the causal driver based on gene-phenotype matching, dosage sensitivity (Haploinsufficiency/Triplosensitivity), and structural impact.
     • Process: Variants are paired randomly; winners advance. To identify the top-K (4 for WES, 8 for WGS), the system runs sub-tournaments on candidates that lost to the current top-ranked variants, ensuring a robust global ranking.

B. Key Technical Innovations

• Agentic Reasoning: Unlike static scoring functions, CoNVict uses LLMs to dynamically weigh heterogeneous evidence (genomic, phenotypic, literature) against the specific patient context.
• Gene Knowledge Cache: A pre-computed database of phenotype-relevant keywords for ~48,000 genes, generated via LLM-grounded web searches, allowing the system to reason about gene function without overflowing context windows.
• Handling Non-Coding Variants: The system explicitly incorporates regulatory annotations and treats low coding overlap as non-disqualifying, focusing instead on regulatory disruption and phenotype fit.

3. Key Contributions

• First Agentic AI for CNV Prioritization: Introduces a framework that uses multi-step, LLM-driven reasoning to bridge the gap between automated annotation and clinical decision-making.
• Superior Performance on VUS and Novel Variants: Demonstrates that agentic reasoning can effectively prioritize variants lacking definitive pathogenicity labels in databases.
• Robustness to Noise: Maintains high performance even in Whole Genome Sequencing (WGS) scenarios where background noise (non-causal CNVs) is ~7.5x higher than in WES.
• Open Source: The system, source code, and data are publicly available.

4. Results

The system was evaluated on 397 simulated diagnostic cases (320 WES, 77 WGS) spanning multiple clinical subspecialties and variant sources (ClinVar, DECIPHER, UDN, and novel computationally derived cases).

• Overall Performance:
  • CoNVict ranked the causal CNV at Position 1 in 74.3% of cases and within the Top 5 in 89.2%.
  • This represents a 10.3 percentage point improvement over the next best method (Exomiser) at Top-1.
• WGS vs. WES:
  • CoNVict achieved 78.1% Top-1 recall in WES and 58.4% in WGS.
  • In WGS, competing tools (PhenoSV, SvAnna) saw significant performance drops due to increased background noise, while CoNVict retained 70.1% Top-5 recall, demonstrating superior noise filtering via contextual reasoning.
• VUS and Novel Variants:
  • On variants classified as VUS by AnnotSV (including novel cases), CoNVict achieved 54.8% Top-1 recall, significantly outperforming Exomiser (38.7%).
  • For novel variants (absent from curated databases), CoNVict led with 49.0% Top-1 recall vs. 25.5% for Exomiser.
• Non-Coding Variants:
  • CoNVict achieved 44.4% Top-1 recall for non-coding variants, compared to 11.1% for Exomiser and 0.0% for SvAnna. It was the only tool to rank any intergenic variant first.
• Ablation Studies:
  • The system's performance relies on both the CNVerdict triage and the Tournament ranking stages; removing either significantly degrades recall.

5. Significance and Future Directions

• Clinical Impact: CoNVict addresses the critical bottleneck in rare disease diagnosis where clinicians must manually synthesize complex variant data with patient phenotypes. By automating this reasoning, it reduces the time to diagnosis and improves the detection of difficult cases (VUS, non-coding, novel).
• Paradigm Shift: The paper validates that agentic AI is well-suited for biomedical tasks requiring iterative evidence evaluation and contextual judgment, outperforming rigid, static scoring functions.
• Limitations: The study relies on simulated cases with synthetic phenotypes; real-world clinical validation is needed. The system currently does not incorporate trio inheritance data (parent-child) and incurs higher computational costs due to LLM usage.
• Future Work: Plans include prospective clinical validation, integration of trio-based inheritance reasoning, and incorporating functional genomic evidence to further improve non-coding variant interpretation.

In summary, CoNVict represents a significant advancement in computational genomics, demonstrating that agentic AI can effectively prioritize CNVs in rare disease diagnosis, particularly in challenging scenarios where traditional tools fail.