DeepPolisher

A modern diploid genome assembly built from PacBio HiFi reads looks finished — contiguous, phased, in places telomere-to-telomere — and still carries on the order of eight base-level errors per megabase, clustered in homopolymers and tandem repeats where the sequencing chemistry itself is biased. A consensus step cannot remove a mistake that most of the reads agree on, and the polishing passes that try face a bind: edit aggressively and you introduce new errors, edit conservatively and you leave most in place.

DeepPolisher, from the genomics team at Google with the Genomics Institute at UC Santa Cruz, sidesteps that trade-off by treating polishing as sequence prediction rather than variant calling. Read alignments are partitioned into windows of at most 100 bp and encoded as a tensor whose channels are the read base, whether it matches the assembly, the base quality and the mapping quality. An encoder-only transformer reads the window and emits the sequence the reads actually support; wherever that prediction disagrees with the assembly, the difference is written out as a candidate correction in VCF, applied with BCFtools. It appeared as a preprint in September 2024 and in Genome Research in 2025, with the Human Pangenome Reference Consortium as a corporate author.

Prediction is only as good as the alignments behind it, and assemblers routinely collapse heterozygous sequence into a single falsely homozygous copy. When such a stretch runs longer than a HiFi read — up to roughly 25 kb — the aligner cannot tell which haplotype a read came from and assigns it arbitrarily, so no heterozygous edit can be made there. The companion PHARAOH procedure (PHAsing Reads in Areas Of Homozygosity) fixes the input rather than the model: it phases those regions with Oxford Nanopore ultra-long reads over 100 kb, then reassigns each HiFi read to the haplotype it matches by edit distance.

#Key Features

• Sequence prediction over variant calling: An edit is proposed only where the reads imply a different string, not where a genotyping model clears a threshold.
• Diploid-aware: Both haplotypes are polished from haplotype-assigned alignments, under a diploid alignment loss.
• Phasing through false homozygosity: PHARAOH recovers read haplotypes in long homozygous stretches, adding a 9% error reduction on HG002 and 2.5% on HG005.
• Restrained editing: Of the polishers compared it introduces the fewest new error k-mers, with genotype-quality cutoffs tuned separately by edit type.
• Transfers to unseen genomes: One checkpoint was applied across 180 human assemblies and two marmoset assemblies without retraining.

#Technical Details

The network is an encoder-only transformer with learned positional encoding, trained with the alignment loss introduced for DeepConsensus, which scores a prediction against truth through an alignment rather than position by position. The release model was trained on an HG002 hifiasm v0.19.5 diploid assembly using the Q100 consortium's HG002 T2T v0.9 assembly as truth — contigs mapping to chromosomes 1–19 for training, 21–22 for tuning, chromosome 20 held out; a second model was trained the same way on a Verkko assembly. On HG005, a sample absent from training, it cut variant-calling errors against the Genome in a Bottle v4.2.1 call set from 20,274 to 11,750, or 8.14 to 4.72 errors per megabase, against 5.51 for DeepVariant on the same reads, 6.68 for NextPolish2 and 7.99 for the T2T consortium pipeline. Across eight held-out HPRC samples it improved k-mer quality value by 3.95 on average versus 2.26 for DeepVariant, while inducing error k-mers in 4% of its edits against DeepVariant's 9%. Coverage of 40× is optimal, though gains persist down to 10×.

#Applications

Its production use is population-scale assembly: it polished 180 assemblies for the second HPRC data release, lifting average quality value within high-confidence regions from 66.66 to 70.05 and reducing switch and hamming error for all 111 samples with trio data. At gene level it removes false frameshifts and premature stop codons that would otherwise be read as biology. It generalizes past human — on two marmoset assemblies it reduced whole-genome errors by 9% and 11% — and ships as code, Docker images and WDL workflows, so a group can polish its own HiFi assemblies without retraining.

#Impact

DeepPolisher makes base-level accuracy an automated step in reference-grade assembly rather than a curation problem, and its place in the HPRC release pipeline puts its edits into sequences that downstream analyses treat as ground truth. The work also sharpens how polishing is measured: k-mer quality values, compared against alignment-based assessment over the same regions, miss at least 90% of residual errors, so a headline QV overstates how finished an assembly is. The limits are explicit — the model is trained on human HiFi data and needs retraining for distant clades, regions covered only by ONT reads cannot be polished, and training needs a truth assembly, available for only a handful of benchmark samples.