New system helps identify real cancer mutations

A new framework called AVA automates the review of DNA sequencing data to distinguish true mutations from technical errors.

Researchers created AVA, which stands for Agentic Variant Adjudication. This system helps identify real mutations in tumor DNA. It works by reviewing sequencing data to separate true mutations from technical artifacts.

Finding mutations is vital for cancer diagnosis and treatment. Doctors use genomic sequencing to look for these changes. However, distinguishing real mutations from sequencing artifacts is often difficult and error-prone.

Current methods use fixed cutoffs for depth and variant-allele-frequency (VAF). These cutoffs cannot separate artifacts from real mutations in the borderline low-VAF zone. In this zone, the two types of data overlap.

Because of this overlap, laboratory specialists must manually inspect the data. This manual review is slow and can vary between different people. This process can also cause delays in making a diagnosis.

AVA uses a framework of subagents to automate this inspection. These subagents reason over evidence and generate investigative code. They provide a verdict and an auditable reasoning trace without needing prior training on existing data.

The researchers tested AVA on the SEQC2 DNA-seq benchmark and patient RNA-seq data. They focused on borderline variants where standard thresholds fail. The tests used a VAF interval of $[0.05, 0.20)$.

Both closed-source models like Claude Sonnet 5 and open-weight models like GLM 5.3 Flash and Gemma 4 31b performed well. These agents outperformed the best possible VAF and depth thresholds. They also beat the FilterMutectCalls method.

A variant is a location where DNA reads disagree with a reference genome. Depth is the number of reads covering a specific location. VAF is the fraction of reads that support a specific variant.

AVA is suitable for clinical use where privacy is a concern. It can run on locally deployed open-weight models. The system provides descriptive reasoning traces to keep specialists involved in the process.

Key facts

  • AVA automates the review of DNA sequencing data to identify true mutations.
  • The system performs better than fixed VAF and depth thresholds in the borderline zone.
  • AVA can run on local models to protect patient privacy.
  • The framework provides an auditable reasoning trace for every decision made.

All sections

Edition