Random-Forest Phasing Tool Nearly Doubles Contiguity of Diploid Genome Assemblies

A new computational method called RFhap delivers sharper, more complete reconstructions of the two haplotypes that make up every human genome. Developed by Damariz González, Gabriel Cabas and colleagues, and published in BMC Bioinformatics, the tool tackles a stubborn bottleneck in long-read genome sequencing: correctly sorting individual sequencing reads into the maternal and paternal copies of the genome before assembly begins. In benchmark tests on human trios, RFhap nearly doubled a key measure of assembly contiguity compared with the widely used Hifiasm-Trio pipeline, while cutting a major class of phasing error roughly in half.

Haplotype-resolved de novo assembly — building a reference-free reconstruction of both parental chromosome sets — matters because many disease-relevant variants only make sense when maternal and paternal versions are separated genome-wide. Complex regions such as interspersed repeats can scramble traditional assemblies, mixing pieces inherited from the two parents. Trio-aware assemblers like Hifiasm-Trio address this by mining short reads from both parents to derive parent-specific k-mers, short DNA words unique to each side of the family, and using them to guide phasing inside the long-read assembly graph. That approach, however, depends on a single fixed k-mer length and heuristic decisions that can falter when sequencing errors accumulate or repetitive regions tangle the graph.

RFhap replaces that single-scale strategy with two combined ingredients. First, it generates multi-k-mer parent-specific markers at several k-mer lengths, giving the classifier more robust evidence when any one word length is unreliable. Second, it feeds these markers through an alignment-free k-mer lookup engine into a random forest classifier, a machine-learning ensemble that assigns each long read to the maternal haplotype, the paternal haplotype, or the unknown category before assembly ever starts.

The benchmark covered four human trio datasets from the Human Pangenome Project, sequenced with two Oxford Nanopore chemistries, R10.4 and R9.4.1, a useful breadth given how differently these platforms behave. Evaluating downstream assemblies with the Merqury tool, RFhap raised corrected haplotype NG50 — a contiguity metric for the haplotype-resolved assembly — to a mean of 24.3 megabases versus 13.1 megabases for Hifiasm-Trio, close to a doubling.

Switch error rates, which count how often an assembly erroneously jumps between haplotypes, fell from a mean of 0.236 percent to 0.111 percent. Long-switch errors dropped roughly threefold across the datasets, with the largest gains occurring inside interspersed repeats, precisely the regions where fixed-k-mer heuristics tend to break down.

Accuracy gains sometimes come at the cost of other quality metrics, but the authors report that RFhap remained competitive in consensus QV, an error-rate measure of base-level accuracy, and in parental k-mer completeness, which tracks how much of the parents’ distinct sequence is retained in the final assembly.

The team characterizes RFhap as a step toward accurate and automated diploid human de novo assembly from long reads. Because the method works from standard trio data — parental short reads plus a child’s long reads — it slots into existing sequencing workflows without requiring additional laboratory protocols.

As with any benchmark study, the results reflect the four tested trios and the two nanopore chemistries evaluated, and performance on other datasets remains to be demonstrated. Within that scope, the reported gains in haplotype contiguity and phasing accuracy mark a measurable advance for trio-based long-read assembly pipelines.

Subject of Research: A random-forest-based, alignment-free long-read phasing method for haplotype-resolved diploid genome assembly

Article Title: RFhap: accurate alignment-free long-read phasing for improved diploid de novo assembly

Article References: González, D., Cabas, G., Miquel, J. F., Moraga, C., Salas-Pérez, F., & Di Genova, A. (2026). RFhap: accurate alignment-free long-read phasing for improved diploid de novo assembly. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06735-2

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06735-2

Keywords: genomics, phasing, long-read sequencing, de novo assembly, bioinformatics, random forest, k-mers, haplotype, Human Pangenome Project, Oxford Nanopore, machine learning, RFhap