Cancer genomic profiling predicts pathogenicity of BRCA1 and BRCA2 variants

This study develops machine learning models that leverage large-scale real-world cancer genomic profiles to accurately predict the pathogenicity of BRCA1 and BRCA2 variants, successfully reclassifying nearly half of the previously uncertain variants by integrating tumor-derived features such as homologous recombination deficiency signatures and co-mutated genes.

The Gist

Imagine your DNA is a massive, intricate instruction manual for building and maintaining a human body. Sometimes, there are typos in this manual called genetic variants. Most typos are harmless (like a misspelled word that doesn't change the meaning of a sentence), but some are dangerous errors that can cause the body's repair crews to fail, leading to cancer.

The genes BRCA1 and BRCA2 are like the "Chief Safety Officers" of the body's DNA repair team. When they work, they fix broken DNA. When they are broken (due to a dangerous variant), the body can't repair itself, and cancer risks skyrocket.

The Problem: The "Uncertain" Typos

Doctors have found thousands of these typos in the BRCA1 and BRCA2 genes. However, for many of them, they don't know if they are harmless typos or dangerous errors. In the medical world, these are called Variants of Uncertain Significance (VUS).

Think of a VUS like a blurry photo of a suspect. You know someone is there, but you can't tell if they are a hero or a villain. This is a huge problem because:

1. Risk: If it's a villain, the patient needs extra screening or surgery.
2. Treatment: If it's a villain, the patient might be eligible for special drugs (PARP inhibitors) that only work if the safety officer is broken.
3. Confusion: If it's a hero, the patient might undergo unnecessary stress and procedures.

Currently, about one-third of these variants are stuck in this "blurry photo" limbo.

The Solution: A Detective's AI

The researchers in this paper decided to solve the "blurry photo" problem using a giant detective AI. Instead of just looking at the typo itself, they looked at the crime scene (the tumor).

They gathered data from over 800,000 cancer patients and looked at 120,000 specific tumor samples where these BRCA typos were found. They trained a computer model (Machine Learning) to spot patterns, much like a detective learning to tell a real criminal from an innocent bystander by looking at their surroundings.

The Clues the AI Used:
The AI didn't just look at the typo; it looked at the "footprints" left behind in the tumor:

• The "Scars" (HRD Signature): When the safety officer is broken, the DNA gets messy and leaves specific "scars." The AI learned that if these scars are present, the typo is likely a villain.
• The Neighborhood: Certain types of cancer (like ovarian or breast) are more likely to have these villains.
• The Accomplices: The AI checked if other bad guys (mutations in genes like TP53) were hanging out with the BRCA typo.
• The Copy Number: Did the tumor have extra copies of the bad gene?

How It Works: The "Tumor Report Card"

The researchers built two AI models (one for BRCA1, one for BRCA2). They fed the AI thousands of cases where the answer was already known (the "Truth Set"). The AI learned: "Ah, when I see a typo in a breast tumor with these specific scars and these accomplices, that typo is almost certainly a villain."

Then, they applied this AI to the blurry photos (the VUS).

• High Confidence (High Score): The AI says, "This typo is definitely a villain." (Pathogenic)
• Low Confidence (Low Score): The AI says, "This typo is definitely harmless." (Benign)
• Indeterminate: The AI says, "I'm not sure yet."

The Results: Clearing the Fog

The results were like turning on a bright light in a dark room:

• Accuracy: The AI was incredibly accurate, almost perfect at distinguishing known villains from known heroes.
• Solving the Mystery: When they tested the AI on the 2,700+ "blurry photos" (VUS), it successfully cleared up the uncertainty for about 40% of BRCA1 variants and 50% of BRCA2 variants.
• Real-World Proof: They checked the patients who received special cancer drugs (PARP inhibitors). The patients whose VUS were labeled "Villain" by the AI responded to the drug just like patients with known villains. The patients labeled "Hero" did not. This proved the AI was right.

Why This Matters

Before this study, a patient with a "blurry photo" VUS might have been told, "We don't know, so we can't help you decide."

Now, thanks to this research:

1. More Answers: Doctors can tell many more patients whether they are at high risk or not.
2. Better Treatment: Patients can get the right drugs (PARP inhibitors) sooner if their tumor shows the "villain" signature.
3. Less Stress: Patients with "harmless" variants can stop worrying about unnecessary surgeries.

The Bottom Line

This paper is like giving doctors a super-powered magnifying glass. By looking at the whole picture of the tumor (the scars, the neighbors, the history) rather than just the single typo, they can finally solve the mystery of thousands of genetic variants. This turns "unknowns" into "knowns," saving lives and reducing anxiety for cancer patients everywhere.

Note: This is a preprint (a draft of a study), meaning it hasn't been fully peer-reviewed by other scientists yet, but the results are very promising and ready for the medical community to start using.

Technical Summary

1. Problem Statement

Accurate classification of BRCA1 and BRCA2 genetic variants is critical for cancer risk assessment, screening, and the selection of targeted therapies (specifically PARP inhibitors). However, a significant challenge remains: over one-third of reported BRCA1/2 variants are classified as Variants of Uncertain Significance (VUS) or have conflicting pathogenicity assertions in public databases like ClinVar.

• Clinical Impact: The inability to classify these variants hinders precise inherited cancer risk stratification and prevents patients from accessing life-saving targeted therapies.
• Data Gap: While routine tumor genomic profiling generates vast amounts of real-world data, this data is currently underutilized for variant interpretation. Pathogenic BRCA1/2 variants induce specific tumor genomic phenotypes (e.g., Homologous Recombination Deficiency or HRD), but these features have not been systematically leveraged at scale to resolve VUS.

2. Methodology

The authors developed a machine learning framework to predict variant pathogenicity using real-world tumor genomic data.

• Data Source & Cohort:

  • Utilized >800,000 cancer samples from the Foundation Medicine dataset.
  • Focused on 120,660 samples containing BRCA1 or BRCA2 variants.
  • Exclusion Criteria: Samples with high Tumor Mutation Burden (TMB >15 Mut/Mb) were excluded to avoid confounding by hypermutator phenotypes (passenger mutations). Samples with multiple BRCA1/2 variants (where one was already known pathogenic) were also excluded.
  • Truth Set: Variants classified as Benign/Likely Benign (LB/B) or Pathogenic/Likely Pathogenic (LP/P) in ClinVar (non-conflicting) served as the training and validation ground truth.
  • Target Set: Variants listed as VUS, conflicting, or absent in ClinVar were reserved for prediction.

• Feature Engineering (389 Features):
  The models integrated four categories of features:

  1. Patient/Sample: Cancer type (37 labels), sex, age, and diagnostic test type.
  2. Tumor Genomic: Tumor Mutation Burden (TMB), Percent Genome Loss of Heterozygosity (LOH), Microsatellite Instability (MSI) status, and a Pan-cancer HRD signature (HRDsig).
  3. Variant-Specific: Germline vs. somatic status, tumor zygosity (homozygous/heterozygous), Mutant Allele Fraction (MAF), and gene copy number.
  4. Co-observed Genomic Events: Binary presence/absence of mutations in 377 other genes (e.g., TP53, KRAS, TERT).

• Machine Learning Model:

  • Algorithm: Random Forest Classifier (implemented via tidymodels and ranger in R).
  • Training: Models were trained separately for BRCA1 and BRCA2 using a 60/20/20 split (Training/Test/Validation) of the truth set.
  • Hyperparameters: Optimized via 10-fold cross-validation (e.g., BRCA1: 60 mtry, 40 min_n; BRCA2: 30 mtry, 40 min_n).
  • Prediction Logic: Individual variant observations were predicted, and the mean score was calculated for variants with multiple observations.
  • Thresholding: A Two-Graph ROC (TG-ROC) approach was used to establish thresholds ensuring minimum sensitivity and specificity of 0.9. Predictions were stratified into Low (Benign), Indeterminate, and High (Pathogenic).

• Validation & Integration:

  • External Validation: Compared against functional assays, AlphaMissense (missense), and SAI-10k-calc (splicing).
  • Clinical Correlation: Analyzed Real-World Progression-Free Survival (rwPFS) in ovarian cancer patients receiving PARP inhibitors.
  • VCEP Integration: Applied ACMG/AMP guidelines to integrate model predictions as "Strong Benign" or "Strong Pathogenic" evidence to reclassify VUS.

3. Key Results

• Model Performance:

  • The models achieved near-perfect performance on the validation set for variants with ≥5 observations:
    • BRCA1: ROC-AUC = 1.000
    • BRCA2: ROC-AUC = 0.989
  • Feature Importance: The Pan-cancer HRD signature was the single most influential feature for both models, followed by genome-wide LOH, TMB, and variant zygosity.
  • Classification Accuracy:
    • Low predictions correctly identified 97.7% (BRCA1) and 94.0% (BRCA2) of known LB/B variants.
    • High predictions correctly identified 97.5% (BRCA1) and 96.2% (BRCA2) of known LP/P variants.

• VUS Re-classification:

  • Applied to 1,073 BRCA1 and 1,639 BRCA2 VUS (with ≥5 observations).
  • Re-classification Success:
    • 39.48% of BRCA1 and 50.52% of BRCA2 assessable variants were strengthened or enabled for classification.
    • Specifically, 56.9% of BRCA1 and 67.6% of BRCA2 VUS were downgraded to Likely Benign/Benign.
    • 8.7% of BRCA1 and 4.4% of BRCA2 VUS were upgraded to Likely Pathogenic/Pathogenic.

• Clinical Concordance:

  • Functional Assays: High concordance between "High" model predictions and reported functional impacts (approx. 50% of "High" variants showed functional impact vs. <1% of "Low").
  • PARP Inhibitor Response: Patients with BRCA1/2 VUS predicted as "High" showed progression-free survival trends similar to those with known pathogenic variants when treated with PARP inhibitors, whereas "Low" predictions aligned with wild-type outcomes.

• Manual Review of Re-classified Pathogenic Variants:

  • Of the 26 variants re-classified as Pathogenic, manual review confirmed strong supporting evidence (e.g., location in functional domains like RING/BRCT, disruption of splicing, and co-occurrence with HRD signatures).

4. Significance and Contributions

• Scalable Framework: This study establishes a scalable method to transform underutilized, routine tumor profiling data into high-confidence evidence for germline variant classification.
• Resolution of VUS: It directly addresses the "VUS bottleneck," providing a mechanism to reclassify thousands of uncertain variants, thereby reducing clinical ambiguity.
• Evidence Strength: The models provide "Strong" evidence (per ACMG/AMP guidelines) for both pathogenicity and benignity, a level of evidence previously difficult to achieve for rare variants without extensive family studies or functional assays.
• Generalizability: While focused on BRCA1/2, the framework is applicable to other cancer genes associated with defined tumor genomic phenotypes (e.g., other Homologous Recombination Repair genes).
• Clinical Utility: By linking tumor genomic features (like HRD) to variant pathogenicity, the approach supports more precise therapy selection (PARP inhibitors) and risk stratification for patients.

5. Limitations

• Observation Threshold: Reliable predictions require a minimum of 5 independent observations per variant, limiting immediate application to extremely rare variants.
• Confounding Factors: While the study excluded samples with co-occurring BRCA1/2 pathogenic variants, HRD phenotypes could theoretically be driven by other HRR gene mutations or promoter methylation, potentially leading to false positives in "High" predictions for rare cases.
• Data Privacy: Individual patient genomic data cannot be shared due to HIPAA, though aggregated data and model predictions are made available.

In conclusion, this paper presents a robust, data-driven solution to a critical problem in precision oncology, demonstrating that tumor-derived genomic features can serve as powerful predictors of germline variant pathogenicity.