Long-Read Assembly (Noisy Reads) - Usage Guide

Overview

This skill builds de novo genome assemblies from noisy long reads - Oxford Nanopore (R9.4.1, R10.4.1, Dorado simplex/duplex) and PacBio CLR. It covers the modern long-read assemblers (Flye, Canu, NextDenovo, Shasta, Raven, wtdbg2, miniasm) and the bacterial multi-assembler consensus tools (Trycycler, Autocycler). The two decisions it exists to get right are: matching the assembler input flag to the basecaller error regime (the wrong flag silently over-collapses repeats while raising N50), and treating the assembler as the halfway point because a raw long-read assembly is contiguous but low-QV and is not finished until it is polished and its QV is measured. PacBio HiFi reads are accurate, not noisy - use the hifi-assembly skill for those.

Prerequisites

conda install -c bioconda flye canu nextdenovo shasta raven-assembler wtdbg miniasm minimap2 purge_dups porechop_abi trycycler

Genome-size/coverage estimation (GenomeScope2 + jellyfish/KMC) is best run on accurate short reads, not raw ONT, because errors inflate unique k-mers. Polishing (Medaka/Racon) and QV assessment (Merqury) are separate downstream tools - see the assembly-polishing and assembly-qc skills.

Quick Start

Tell your AI agent what you want to do: - "Assemble my ONT R10 reads with Flye" - "I have Nanopore reads basecalled with Dorado SUP - which Flye flag do I use?" - "Assemble a bacterial genome reliably from Nanopore reads" - "My assembly is twice the expected size - what went wrong?" - "Assemble a large plant genome from noisy long reads"

Example Prompts

Matching the flag to the basecaller era

"I have Nanopore reads from an R10.4.1 flow cell basecalled with Dorado SUP. Assemble them with Flye using the correct input mode and explain why --nano-raw would be wrong here."

Bacterial reliability

"Assemble my bacterial isolate from ONT reads as reliably as possible. Use a multi-assembler consensus rather than trusting one assembler, and tell me which contigs are circular."

Large eukaryote

"Assemble a 3 Gb repetitive plant genome from noisy long reads. Pick the most memory-efficient contiguity-first assembler and downsample to the longest reads for the initial assembly."

Diagnosing duplication

"My Flye assembly of a diploid genome is 1.8x the expected genome size and BUSCO-Duplicated is high. Diagnose whether this is haplotig duplication and run purge_dups, checking the coverage cutoffs before trusting them."

Pre-assembly cleanup

"Before assembling my Nanopore reads, check for and remove mid-read adapter chimeras so I don't get structural mis-joins."

What the Agent Will Do

Establish the platform and basecaller era (ONT R9 vs R10/Dorado vs PacBio CLR) - this is an assembly parameter, not optional metadata.
De-chimerize internal adapters with Porechop_ABI and confirm read length/quality and coverage (~30-60x) and read-N50 are adequate.
Select an assembler and the input mode matching the error regime (Flye --nano-hq for modern ONT; Trycycler/Autocycler for bacterial reliability; NextDenovo for large repetitive genomes).
Run the assembly, downsampling to the longest reads (--asm-coverage) on deep large-genome data.
Check for haplotig false-duplication (size vs expected, half-coverage depth peak, BUSCO-Duplicated) and run purge_dups if warranted.
State that the assembly is unpolished and low-QV, and hand off to polishing then QV assessment - the deliverable is a polished assembly with a measured QV, not a high-N50 FASTA.

Tips

The most dangerous mistake is telling the assembler the reads are noisier than they are: --nano-raw on R10/Dorado data over-collapses repeats and raises N50 as the assembly gets worse - nothing crashes.
Record pore + kit + basecaller model before assembling; if the chemistry is unknown the flag cannot be chosen.
A high N50 is not a finished assembly. Raw consensus is Q20-Q30; indels frameshift genes. Polish, then measure QV (Merqury Q40 ~ 1 error/10 kb).
For bacteria, prefer a multi-assembler consensus (Trycycler/Autocycler) over agonizing about which single assembler - any one makes structural errors polishing cannot fix.
Coverage and read-N50 are non-substitutable: 200x of 5 kb reads gives great consensus and a fragmented assembly. Spend effort on read length, not just gigabases.
Filter conservatively - the longest reads are often the lowest quality, and aggressive length filters discard the reads that span hard repeats.
wtdbg2 needs the separate wtpoa-cns consensus step; miniasm does no consensus at all and its output must be polished.
PacBio CLR is legacy (superseded by HiFi). Do not recommend generating new CLR; treat existing CLR like noisy ONT and polish hard.
Pre-2022 tutorials assume R9 and tend to default to --nano-raw plus an unnecessary short-read polish - check the chemistry the advice was written for before applying it.

hifi-assembly - PacBio HiFi (Q30+) phased haplotype-resolved assembly with hifiasm
assembly-polishing - Polishes the low-QV contigs this skill produces
assembly-qc - QUAST/BUSCO/Merqury QV assessment; the QV is the real deliverable
long-read-sequencing/long-read-qc - Read QC and conservative filtering before assembly
long-read-sequencing/basecalling - Basecaller era and model that set the assembler flag
read-qc/contamination-screening - Screen reads for contamination before assembling
workflows/genome-assembly-pipeline - End-to-end QC -> assemble -> polish -> scaffold -> QC