A Singapore-led research team is using AI to tackle one of genome assembly’s most expensive bottlenecks: correcting noisy nanopore sequencing reads without losing real genetic variation.
The team developed HERRO, a deep learning tool designed to improve the accuracy of Oxford Nanopore Simplex reads for genome assembly. In a Nature study published April 27, 2026, researchers reported read-accuracy gains of up to 100-fold and telomere-to-telomere genome assemblies using HERRO-corrected reads.
If replicated more widely, HERRO could reduce labs’ dependence on costly workflows that combine nanopore data with other sequencing technologies. For sequencing labs, the benefit is practical: fewer platforms, less genomic DNA, and simpler workflows for producing complete assemblies.
Why genome assembly still requires multiple platforms
The research team includes scientists from A*STAR’s Genome Institute of Singapore and Oxford Nanopore Technologies. The project also reflects Singapore’s investment in applied AI for life sciences research.
Genome assembly has long involved a tradeoff. Nanopore long reads can span repeats and structural variants that shorter reads struggle to resolve, but they also carry higher error rates. Many labs compensate by combining ultra-long nanopore reads with high-accuracy long reads, such as PacBio HiFi.
The Nature paper says multi-platform approaches can raise costs, require more genomic DNA, and add work across lab preparation and data analysis. HERRO is designed to reduce that burden by correcting Oxford Nanopore Simplex reads before they are passed to genome assemblers.
That makes HERRO a focused example of applied AI: it cleans specialized scientific data so downstream tools can work better. The same data-first constraint shows up in enterprise AI projects, where data pipelines can make or break performance.
HERRO’s correction task is difficult because real genetic variation can look like noise. Its public GitHub repository describes HERRO as a haplotype-aware tool for correcting Nanopore R10.4.1 Kit 14 reads, with an experimental model for R9.4.1 data and a recommended read length of at least 10,000 base pairs.
What HERRO’s benchmarks show
In the study’s tests, HERRO-corrected reads used with the Verkko assembler reached an NG50 of 146 Mb for both haplotypes in the I002C dataset at 71× correction coverage. At roughly 35× correction coverage, NG50 values ranged from about 90 Mb to 130 Mb.
For R10.4.1 ultra-long Oxford Nanopore datasets, more than half of the 46 chromosomes were resolved as telomere-to-telomere contigs or scaffolds. The researchers also reported runs in which the X and Y chromosomes were assembled end to end.
The early users are likely to be sequencing cores, population genomics programs, agricultural genomics teams, biodiversity projects, and research groups working with structurally complex genomes. Clinical diagnostics is a more distant use case; the paper does not validate HERRO as a diagnostic tool.
Independent labs still need to reproduce the results across more organisms, sample qualities, sequencing conditions, and cost models. The study’s competing-interest disclosure also matters because Oxford Nanopore Technologies and AI Singapore jointly funded the project.
If HERRO holds up, it could lower the infrastructure barrier for complete genome assemblies. Recent work on AI agent infrastructure and AI search platforms built around proprietary data points to the same broader shift toward specialized AI systems built around data quality and validation.
Also read: Microsoft’s New AI Agent Brings ‘Always-On’ Automation to Microsoft 365 for another example of AI moving deeper into operational workflows