BICEP: an extension to indels and copy number variants for rare variant prioritisation in pedigree analysis

This paper presents an extension to the BICEP Bayesian inference model that enables the prioritization of rare indels and copy number variants in pedigree-based analyses, demonstrating performance comparable to its original single nucleotide variant model.

The Gist

Imagine your family tree is a giant, ancient library filled with books of DNA. Sometimes, a typo in one of these books causes a family to suffer from a specific health condition. For a long time, scientists had a very smart librarian named BICEP who could help find these typos.

However, until now, BICEP was only good at finding single-letter typos (like changing an 'A' to a 'G'). But DNA isn't just about single letters; sometimes whole words are missing, extra words are added, or entire paragraphs are duplicated. These are called Indels (insertions/deletions) and Copy Number Variants (CNVs).

This paper is about giving BICEP a massive upgrade so it can read all types of typos, not just single letters.

Here is the breakdown of how they did it, using some everyday analogies:

1. The Problem: The "Single-Letter" Librarian

Previously, if you asked BICEP to find a missing paragraph in a book, it would just shrug and say, "I only know how to spot single-letter mistakes." This meant scientists were ignoring a huge chunk of potential causes for genetic diseases because they didn't have the right tools to prioritize them.

2. The Solution: Teaching BICEP New Tricks

The researchers taught BICEP two new skills:

• Spotting "Word" Typos (Indels): These are small chunks of DNA that are missing or added.
• Spotting "Paragraph" Typos (CNVs): These are large sections of DNA that are either deleted or duplicated.

To teach BICEP, they didn't just guess. They used a massive training dataset (like a giant stack of "Answer Keys" from the medical database ClinVar) to show BICEP examples of what a "bad" (disease-causing) typo looks like versus a "good" (harmless) one.

3. The Detective Tools: How BICEP Decides

When BICEP looks at a new typo, it acts like a detective weighing evidence. To decide if a typo is dangerous, it checks specific clues:

• For Indels (The Word Typos): BICEP checks how rare the typo is in the general population (if it's common, it's probably harmless) and how badly it breaks the "grammar" of the gene.
• For CNVs (The Paragraph Typos): This is trickier. The researchers tested four different "detective kits" to see which worked best:
  1. The "Nothing" Kit: Just guessing based on size.
  2. The "CADD-SV" Kit: A computer score that predicts how damaging a structural change is.
  3. The "Loeuf" Kit: A score that measures how much a gene hates losing its function (some genes are very sensitive; others are tough).
  4. The "Super Kit" (The Winner): Combining the CADD-SV score and the Loeuf score.

The Result: The "Super Kit" was the best detective. It gave the most accurate answers for both missing paragraphs (deletions) and extra paragraphs (duplications).

4. The Performance Check: Did it Work?

The researchers put the new BICEP through a final exam. They hid the answers and asked BICEP to guess which typos were dangerous.

• The Good News: BICEP performed just as well on these new, complex typos as it did on the old single-letter ones.
• The "High Precision" Rule: The most important thing for BICEP is Precision. If BICEP says, "This typo is dangerous," it wants to be right 99% of the time. The new models achieved this high level of confidence.
• The Catch: BICEP is still a little bit "cautious." Sometimes, it misses a truly dangerous typo and says, "I'm not sure." This is better than crying wolf, but the researchers hope to fix this as they get more data in the future.

5. Why This Matters

Think of genetic analysis like searching for a needle in a haystack.

• Before: BICEP could only find the needles made of steel (Single Letter variants).
• Now: BICEP can find needles made of steel, wood, and plastic (Indels and CNVs).

This means doctors and researchers can now look at a much wider range of genetic errors when trying to solve the mystery of why a family has a rare disease. The tool is now available for everyone to use, making the search for genetic answers faster and more complete.

In short: The BICEP tool got a software update that lets it understand the full complexity of our DNA, not just the simple parts, helping us find the root causes of genetic diseases more effectively.

Technical Summary

1. Problem Statement

Co-segregation analysis is a standard method for identifying genetic variants responsible for traits within pedigrees. However, existing methodologies often lack the ability to integrate prior evidence of causality (e.g., functional impact, conservation, deleteriousness scores) directly into the statistical model. While the authors previously developed BICEP (Bayesian Inference for Co-segregation and Prioritisation) to address this for Single Nucleotide Variants (SNVs), the model was limited to SNVs only.

This limitation excludes other critical variant types—specifically indels (insertions/deletions) and Copy Number Variants (CNVs)—which play significant roles in complex trait aetiology but are often ignored due to detection and interpretation challenges. The paper addresses the need to extend BICEP's Bayesian framework to include these variant types to enable comprehensive genomic analysis.

2. Methodology

The authors extended the BICEP framework by developing new prior models for indels and CNVs using logistic regression trained on large-scale genomic databases.

• Data Sources & Preprocessing:

  • SNVs: Updated ClinVar data (downloaded 08/12/2025) for both GRCh37 and GRCh38 builds.
  • Indels: ClinVar data including "single submitter" and "multiple submitter" records to increase sample size.
  • CNVs: Structural variants from dbVar (study ID nstd102). Data was filtered to include deletions and duplications between 50bp and 30Mbp overlapping MANE v1.4 protein-coding transcripts. Variants were classified as Pathogenic/Likely Pathogenic or Benign/Likely Benign.
  • Frequency Filtering: Variants with allele frequencies >1% were removed from the pathogenic set to avoid benign misclassifications. gnomAD v4.1 (GRCh38) and v2.1.1 (GRCh37) were used for allele frequency annotation.

• Feature Engineering & Predictors:

  • Indel Model: Used Allele Frequency and IMPACT annotation (HIGH, MEDIUM, LOW, MODIFIER) as predictors. The authors avoided specific deleteriousness scores (e.g., CAPICE) to prevent circularity bias, as many are trained on ClinVar data.
  • CNV Model: Evaluated four predictor configurations for deletions and duplications separately:
    1. NULL: No predictors.
    2. CADD-SV: Deleteriousness scores (converted to GRCh37 via liftOver).
    3. pLoEUF: Sum of reciprocal loss-of-function intolerance scores (gnomAD v4.1) for genes fully contained within the CNV.
    4. BOTH: Combination of CADD-SV and pLoEUF.

• Model Training & Benchmarking:

  • Data was split into 80% training and 20% testing sets.
  • The output is a logPriorOC (log10 prior odds of causality). A value >0 indicates a prediction of "pathogenic."
  • Metrics: Sensitivity, Specificity, PPV (Positive Predictive Value), NPV (Negative Predictive Value), and MCC. PPV was prioritized as the primary metric, reflecting the proportion of high-priority variants that are truly pathogenic.

3. Key Contributions

• Extension of BICEP: Successfully integrated prior models for indels, deletions, and duplications into the existing BICEP software pipeline.
• Predictor Optimization: Systematically evaluated different input predictors for CNVs, identifying that combining CADD-SV and pLoEUF scores yields the best performance for both deletions and duplications.
• Implementation: The updated models are now available as the default in the BICEP GitHub repository, supporting both GRCh37 and GRCh38 genome builds.
• Bias Mitigation: Explicitly avoided circularity bias in the indel model by excluding ClinVar-trained deleteriousness scores.

4. Results

• Indels:
  • Achieved an exceptionally high PPV of 99.8%, indicating that variants with a positive logPriorOC are almost certainly pathogenic.
  • NPV was modest (62.3%), suggesting the model misses some truly pathogenic indels (false negatives), likely due to data imbalance and smaller sample sizes compared to SNVs.
• CNVs (Deletions & Duplications):
  • Deletions: The BOTH model (CADD-SV + pLoEUF) achieved the highest PPV (92.0%). While differences between models were small, the combined model is recommended.
  • Duplications: The BOTH model significantly outperformed others with a PPV of 84.6%.
  • General Performance: While CNV PPVs were slightly lower than short variants (SNVs/Indels), they remain robust. The authors attribute the gap to the scarcity of high-quality deleteriousness predictors for CNVs compared to SNVs.
• Overfitting: Confidence intervals between training and testing sets overlapped well, indicating the models are not overfitted.

5. Significance

• Comprehensive Analysis: This extension allows researchers to prioritize rare variants across the entire spectrum of genomic variation (SNVs, indels, CNVs) within a single Bayesian framework, rather than filtering them independently.
• Pedigree Utility: By integrating prior evidence with co-segregation data, BICEP improves the accuracy of identifying causal variants in family-based studies, which is crucial for diagnosing rare genetic disorders.
• Future-Proofing: The authors acknowledge that performance will improve as more training data becomes available and better predictors for CNVs are developed. Providing these models now ensures that the community can utilize modern sequencing technologies effectively while the field matures.
• Open Access: The tools are freely available, promoting reproducibility and adoption in neuropsychiatric and general genetic research.