Summary
A bioRxiv preprint describes LongVerse, a system that uses AI agents to configure a fixed pipeline for haplotype-resolved DNA methylation analysis from Nanopore or PacBio reads. In tests on the HG002 reference genome, the authors report calls identical to a human expert’s and comparisons with bisulfite sequencing.
Researchers at the University of Southern California describe LongVerse, a system in which AI agents turn a natural-language instruction into haplotype-resolved DNA methylation calls from Nanopore or PacBio sequencing reads. The results are reported in a bioRxiv preprint posted on October 19, 2026.
The authors say the agents inspect sequencing data and supply arguments to a pinned analysis pipeline, but do not write code. A pinned pipeline uses a specified version of the analysis workflow, keeping the agents’ role focused on configuring that workflow rather than generating new software.
Keeping the agents within a fixed workflow
DNA methylation is a chemical modification of DNA. Haplotype-resolved analysis assigns calls to particular chromosome copies, which can reveal when the two inherited copies carry different methylation patterns. Nanopore and PacBio are long-read sequencing technologies; LongVerse is designed to analyse reads from either platform.
The system is intended to make analysis accessible through a short English instruction while retaining an established pipeline for processing the data. The authors also frame it as a way to check agent outputs against biological data and expert analysis, rather than relying only on scoring rubrics or language-model judges.
Results reported on HG002
The evaluation used HG002, a healthy reference genome. The authors report that LongVerse and a human expert produced byte-identical calls, with a reported correlation of r = 1.000. In comparisons with bisulfite sequencing, the abstract gives correlations of r = 0.943 and r = 0.884. Bisulfite sequencing is a different method commonly used to measure DNA methylation, so the comparison provides a check against an independent measurement approach.
The abstract also reports that a consensus agent, which accepts calls from per-site callers, reduced error by 10.7% compared with the best published caller. It gives an AI conversation cost of $0.12 to $0.54 per run; this is the reported conversation cost, not a stated total cost for sequencing and analysis.
These findings are an initial evaluation on a reference genome, rather than a demonstration across a broad range of samples. The preprint’s abstract does not specify which of the two bisulfite-correlation values corresponds to each sequencing platform, or define the error metric behind the 10.7% reduction.