Questions on applying RAFT to human genome assembly with corrected ONT simplex reads

[kuanchiun]

Dear developers,

Thank you very much for developing RAFT and for the clear and insightful paper.
I am still relatively new to genome assembly, and I am currently assembling two haplotype-resolved human genomes. I would like to ask whether RAFT would be beneficial in my case.

Data Description:

• ~45× ONT R10.4.1 simplex reads (basecalled & corrected with Dorado, filtered by length > 10 kb using seqkit), N50 ≈ 54K
• ~20× Hi-C reads (for scaffolding)
• ~75x Illumina WGS reads (for polishing)

Assembly workflow:
HiFiasm (assembly) -> HapHic (scaffolding) -> Dorado polish + NextPolish2 (polishing)

From your paper, I understand that:

• Contained-read removal can be more problematic for ONT simplex data than for HiFi.
• Higher coverage helps reduce the risk of gaps caused by contained reads.
• RAFT mitigates such gaps by fragmenting reads to a more uniform length.

My questions are:

1. Given that I already have ~45× Dorado-corrected ONT simplex reads, would you recommend running RAFT before HiFiasm?
2. I plan to increase coverage from 45× to at least 80×. In this case, would RAFT still be recommended before HiFiasm?
3. Are there specific coverage thresholds or data characteristics where RAFT provides the most benefit (e.g., >30× simplex, UL vs duplex, etc.)?
4. If I use RAFT, should I keep the default parameters (20 kb fragment length, µ=1.5), or do you suggest tuning them for human genome + simplex data?

Any guidance would be very helpful, especially for deciding whether RAFT should be integrated into my standard pipeline.
Thank you again for your work and for making RAFT available to the community!

Best Regards,
Kuan Chiun

[ssk5e]

Hi Kuan!
Sorry for the late response.

1. When we worked on RAFT, Dorado-corrected ONT reads weren't available. We have not experimented with Dorado corrected reads. You can try to feed Dorado-corrected reads to RAFT. If you later pass the reads to Hifiasm, you might want to consider how many rounds of error correction you perform.
2. RAFT should work fine even at 80x coverage. One of the datasets we experimented on had similarly high coverage, and we got good results. What worked for 45x data should work for 80x data as well.
3. The results from using RAFT are only as good as the error-correction and overlap computation. Hifiasm recommends at least 15x coverage. We did not experiment with very low coverage data.
4. During our experiments, we found that 20kb was a good length for the datasets on which we experimented. The µ parameter is supposed to help differentiate a repetitive region in the genome from a non-repetitive one. It doesn't vary with input data. We did not vary it for our experiments on HiFi, ONT Duplex, and ONT Simplex data. If you are assembling human genomes, tuning this parameter should not be necessary.

Please feel free to ask any other questions you may have.

Best,
Sudhanva