AVA: Agentic Variant Adjudication for Automated, Auditable Read-Level Review in Cancer Diagnostic Sequencing

arXiv:2611.14560 · q-bio.GN, cs.AI · Submitted 2026-11-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: I'm Ines.

Marcus: And I'm Marcus, here with Yuki.

Ines: We're looking at the paper: "AVA: Agentic Variant Adjudication for Automated, Auditable Read-Level Review in Cancer Diagnostic Sequencing".

Marcus: Here's the gist of the paper. Identifying genuine mutations from sequencing artifacts is error-prone in practice because fixed VAF and depth thresholds cannot separate them in the borderline low-VAF zone where they overlap This paper presents AVA, an agentic framework that automates read-level inspection of cancer diagnostic sequencing data by emulating a laboratory specialist to adjudicate variant calls.

Ines: The Problem Addressed The central issue is the manual inspection of read-level sequencing data required to distinguish genuine mutations from technical artifacts, a slow step that suffers from high inter-rater variability and causes delays in diagnosis The borderline zone, where true mutations and artifacts overlap according to VAF and depth thresholds, necessitates manual review by a laboratory specialist.

Marcus: This manual process is inefficient and variable across different specialists.

Ines: AVA Framework Operation AVA employs an agentic workflow where an orchestrator launches a sub-agent per borderline variant into a code-execution workspace.

Paper summary: Ines: So wrapping up this paper on AVA: Agentic Variant Adjudication for Automated, Auditable Read-Level Review in Cancer Diagnostic Sequencing. The core idea is using an agentic workflow to automate the laborious process of distinguishing genuine mutations from technical artifacts at the read level by letting specialized agents inspect sequence context around variants.

Marcus: It’s about moving away from fixed thresholds and toward a system that performs detailed, iterative reasoning on individual reads, which has shown performance gains over those simple statistical methods on benchmarks like SEQC2.

Yuki: From a population perspective, it suggests that the biological signal is often hidden in the local sequence context rather than just overall read counts or summary statistics. It points toward a more nuanced way to interpret sequencing data when trying to characterize cancer subtypes across different samples.

Ines: The implications are that this approach could significantly speed up diagnostic pipelines and reduce inter-rater variability by providing an auditable justification for every variant call, which is crucial for high-stakes clinical decisions.

Marcus: The authors acknowledge the limitations, pointing out that their current validation is on a single cell line set and that they still need to study the failure modes in the agent’s reasoning before deploying it broadly.

Yuki: Ultimately, AVA gives us a tool to probe those borderline low versus high variant allele frequency zones with more precision than standard methods allow, even if we need more data from real patient samples to truly see its full potential for widespread use.

Conclusion: Ines: So, we're wrapping up on AVA: Agentic Variant Adjudication for Automated, Auditable Read-Level Review in Cancer Diagnostic Sequencing. Basically, this paper introduces a system that uses AI agents to look at sequencing data and decide if a variant is real or just a technical error.

Marcus: It’s about automating the manual inspection that used to take lab specialists forever because they couldn't tell the difference when variants overlapped based on simple rules for read depth and frequency.

Yuki: The authors are using this agentic framework to simulate how a human expert would adjudicate those borderline calls by having an agent autonomously inspect the sequence context around the variant.

Ines: What this means is that instead of a slow, variable manual review process, you get a system that provides an auditable reason for every decision it makes about whether a mutation is genuine or just noise.

Marcus: The paper shows AVA models actually perform better than just using fixed thresholds on both DNA and RNA sequencing data on the SEQC2 benchmark. They hit an accuracy of zero point nine two seven plus or minus zero point zero zero eight, which beats the best threshold baseline of zero point eight one zero.

Yuki: That improvement is significant because it suggests that looking at individual reads and their local context gives us a better picture than just summarizing everything into a big count or frequency number. It points toward a more nuanced way to see cancer across different samples.

Ines: But the authors do have some caveats, like they're testing this on a single cell line set instead of real patient biopsies, and they still need to figure out exactly where the reasoning of these agents might go wrong sometimes.

Marcus: Yeah, and those are important limitations because we need to know how robust this is when you move from a controlled lab setting to messy, real-world clinical data.

Yuki: So it’s a strong tool for refining the diagnostic process right now, but we definitely need to see how it holds up when the complexity jumps up with more patient samples and real clinical workflows.

University of Toronto Centre for Addiction and Mental Health Canada The Hospital for Sick Children Canada Brain Health Imaging Centre Centre for Addiction and Mental Health Canada Department of Psychiatry University of Toronto

q-bio.GN, cs.AI

Submitted: 2026-11-05

Updated: 2026-11-05

Comments: 14 pages, 5 figures. Accepted at ML4H (Machine Learning for Health) 2026 Findings Track

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 97/100

The gist: Identifying genuine mutations from sequencing artifacts is error-prone in practice because fixed VAF and depth thresholds cannot separate them in the borderline low-VAF zone where they overlap This…

Key concepts

Borderline Zone
This is the area in sequencing data where true mutations and technical artifacts overlap when using standard metrics like VAF (Variant Allele Frequency) or depth thresholds. Because of this overlap, manual review by an expert is necessary to correctly identify the true mutation.
Agentic Workflow
AVA uses an agentic approach where a main orchestrator launches specialized sub-agents for each borderline variant. These agents autonomously examine the data, query supporting evidence, and iterate until they can make a final decision on whether the variant is real or an artifact.
Read-Level Inspection
Instead of relying only on summary statistics (like VAF or depth), AVA examines individual sequencing reads. It writes and executes custom code to analyze the local sequence context around a variant, allowing it to detect tumor-specific signals that might be missed by simpler counting methods.

Terminology

Summary

Identifying genuine mutations from sequencing artifacts is error-prone in practice because fixed VAF and depth thresholds cannot separate them in the borderline low-VAF zone where they overlap This paper presents AVA, an agentic framework that automates read-level inspection of cancer diagnostic sequencing data by emulating a laboratory specialist to adjudicate variant calls. p. 0

Where the paper says it

ca Department of Laboratory Medicine and Pathobiology, University of Toronto, Canada Division of Pathology, Department of Paediatric Laboratory Medicine, The Hospital for Sick Children, Canada The Arthur and Sonia Labatt Brain Tumour Research Centre, The Hospital for Sick Children, Canada Abstract Identifying a tumor’s DNA mutations is central to cancer diagnosis and treatment, but distinguishing genuine mutations from sequencing artifacts is error-prone. In practice, genuine mutations are identified using fixed variant-allelefrequency (VAF) and depth thresholds which cannot separate artifacts from genuine mutations in the borderline low-VAF zone where they overlap.

The Problem Addressed

The central issue is the manual inspection of read-level sequencing data required to distinguish genuine mutations from technical artifacts, a slow step that suffers from high inter-rater variability and causes delays in diagnosis The borderline zone, where true mutations and artifacts overlap according to VAF and depth thresholds, necessitates manual review by a laboratory specialist. This manual process is inefficient and variable across different specialists. p. 1

Where the paper says it

In this task of manually interpreting the detected variants, Sirohi et al. (2020) found that laboratory specialists and pathologists across institutions have significant interrater variability, indicating that manual review can both be inefficient and variable.

AVA Framework Operation

AVA employs an agentic workflow where an orchestrator launches a sub-agent per borderline variant into a code-execution workspace. Each sub-agent examines the variant pileup, autonomously queries aligned sequencing data for evidence, and iterates if necessary to return a real-or-artifact adjudication with an auditable justification. p. 2

Where the paper says it

Each sub-agent examines its variant pileup, autonomously queries the aligned sequencing data to gather additional evidence if needed, and ultimately classifies the variant as a true mutation or a technical artifact (Figure 2). As every verdict includes an auditable justification citing the specific features behind the call, AVA’s reasoning remains transparent.

The workflow involves writing and executing custom code to extract sequence context around the variant and comparing it to the reference sequence. p. 2
Where the paper says it

125 High base quality (Q42) and mapping quality (MQ60) 4 unique fragment starts, no soft clips, low NM (1-2) Strand skew: 0/4 (p=0. 113) Absent in normal (0 alt reads at 58x sequencing depth) reads = fetch_reads(bam, chr, pos) for read in reads:
 seq = get_aligned_pairs(read)
 window = get_sequence(seq, pos-10, pos+10)
 check_mismatches_indels(window) AVA writes a custom script to extract the ± 10 bp sequence around the variant from each supporting read and compare to the reference Iterative reasoning Ins pect results AVA examines the output to look for nearby mismatches, indels, or other artifact signatures R eason and decide If the question is resolved, AVA integrates the new evidence with all other signals.

Performance Benchmarking

AVA models outperform the best achievable VAF/depth threshold baseline across both DNA-seq and RNA-seq data on the SEQC2 benchmark. Specifically, the best performing AVA model GLM 5. p. 4

Where the paper says it

Across all 5 seeds, all AVA models exceed the VAF and depth optimistic upper-bound threshold rule in almost all reported metrics. Notably, the best performing AVA model GLM 5.

3 Flash achieved an accuracy of 0. p. 4
Where the paper says it

Notably, the best performing AVA model GLM 5. 3 Flash had accuracy 0.

927 ± 0.008 compared to the best threshold's accuracy of 0.810 on the borderline set. p. 3
Where the paper says it

Ground truth comes from the SEQC2 high-confidence single-nucleotide variant set, a consensus list of true variants for this cell line that the consortium established by cross-validating calls across many sequencing centers, sequencers, aligners, and callers, and which serves as the community gold standard. On SEQC2, we define the borderline zone as the VAF interval [0.

Furthermore, AVA models reduced false negatives (FNs) significantly compared to the threshold baseline in this setting. p. 4
Where the paper says it

008 vs the best threshold’s accuracy of 0. AVA models also reduced the number of false negatives (FNs) compared to the threshold in this setting (from 38 FNs to 1.

Qualitative Insights and Limitations

Qualitative results demonstrate that reading individual reads beats thresholding on summary statistics, as seen when the agent recovers a real mutation discarded by a count threshold because it inspected clean local context and tumor-specific signals. However, limitations exist, including the SEQC2 dataset being a single cell-line set which differs from real patient biopsies. p. 6

Where the paper says it

Limitations While our work shows promise in deploying agents in real-world settings for variant adjudication, there are a few key limitations in our findings. First, the SEQC2 dataset is a single public breast cancer cell-line dataset which differs from real patient biopsies.

Additionally, certain failure modes in AVA's reasoning bias have been observed and require further characterization.

Future Directions

Future work will focus on validating AVA on larger patient aligned sequencing data within real clinical workflows to elucidate its ability to provide clinical impact. There is also a proposal for in-depth analysis of prompt stability and optimization through prompt ablations and environment perturbation. p. 6

Where the paper says it

We also propose a more in-depth analysis of prompt stability and optimization by following an evaluation protocol similar to Rabanser et al. (2026) through prompt ablations and environment perturbation.

The framework's utility could extend to more complex multiagent architectures, such as a voting committee of agents for difficult variant adjudications.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Agentic variant adjudication can separate true variants from artifacts in exactly the borderline zone where fixed VAF/depth thresholds fail, exceeding them across accuracy and F1-score in both DNA-seq and real patient RNA-seq data.

  2. The improved system can function as an auditable second reader by emulating a specialist’s read-level review, recording a inspectable trace for each call which addresses the interrater variability discussed by Sirohi et al. (2020).

  3. The system can utilize a dynamic set of features much like a laboratory specialist, allowing it to perform toolcalling and code execution to investigate further read-level features autonomously if initial evidence is insufficient.

  4. The agentic framework can detect artifacts through specific read-level features such as Strand balance of variant-supporting reads, Base quality of the variant base, and Mapping quality (fraction with MAPQ 0).

  5. The system can perform a zero-shot adjudication, requiring no labeled variants or site-specific retraining, by analyzing read-level evidence from aligned sequencing data to return a verdict with an auditable justification.

  6. The system can be tuned to balance sensitivity and efficiency in clinical workflows by allowing the user to tune the adjudication process based on institutional preferences regarding false positive and false negative tradeoffs.

Abstract

Identifying a tumor's DNA mutations is central to cancer diagnosis and treatment, but distinguishing genuine mutations from sequencing artifacts is error-prone. In practice, genuine mutations are identified using fixed variant-allele-frequency (VAF) and depth thresholds which cannot separate artifacts from genuine mutations in the borderline low-VAF zone where they overlap. A laboratory specialist must manually inspect the read-level sequencing data, a slow step that has high inter-rater variability and causes diagnosis delays. We present AVA: Agentic Variant Adjudication; an agentic framework that automates this read-level inspection with subagents that reason over evidence, generate investigative code, and return a real-or-artifact adjudications with an auditable reasoning trace all without training on existing data. On both the SEQC2 DNA-seq benchmark and patient RNA-seq data, restricted to a subsample of borderline variants where thresholds fail, both closed-source (Claude Sonnet 5) and open-weight (GLM 5.3 Flash and Gemma 4 31b) agents outperform the best achievable VAF/depth thresholds. Ultimately, AVA is suited for privacy-constrained clinical deployment, since the framework can run on locally deployed open-weight models with descriptive reasoning traces to keep specialists in the loop.

Sources

Related papers