Hybrid genome assembly builds a genome from two kinds of sequencing data. Long reads (tens of thousands of letters, but error-prone) lay out the large-scale structure. Short reads (a few hundred letters, but very accurate) then "polish" that layout to fix base-level errors. Polishing breaks down inside repeats, meaning stretches of sequence that occur in several nearly identical copies. A short read from a repeat aligns equally well to every copy, so it says a lot about what the copies share and almost nothing about how they differ. Consider an optimizer that edits each copy separately, as most polishing loops and expectation-maximization schemes effectively do. It faces a badly conditioned problem: every copy-wise edit is immediately pulled back by evidence that refers to the copies jointly. The result is slow convergence, or copies collapsed toward a shared consensus. When those copies carry antibiotic-resistance genes or rDNA variants, that error matters biologically.
The proposal borrows one idea from incoherence processing, the family of quantization methods that rotate a neural network's weight coordinates before rounding. The idea is that the coordinate system is a free choice, and an orthogonal change of basis loses nothing and can be undone exactly. Our claim is that the copies of a repeat family should be re-expressed in a "consensus plus orthogonal contrasts" basis, built from a Hadamard-type orthogonal matrix whose first column is the uniform vector. Once this is done, the multi-platform likelihood of the assembly becomes nearly separable. Randomized block coordinate descent, which updates one small group of variables at a time, then converges at a rate that does not degrade with how ambiguous the short reads are. In plain terms, we optimize what the copies share and how they differ as separate, nearly independent variables, rather than optimizing each copy in turn.
Core Claim
Fix a repeat family with $k$ aligned copies and look at one aligned column. Each variable $x_j$ is a soft base score for copy $j$. Short reads that map ambiguously to all copies contribute curvature only along the shared direction. Reads that distinguish copies, such as long reads that bridge a repeat in HINGE's sense or rare unique-mapping short reads, contribute curvature copy by copy. Locally, the Hessian of the negative log-likelihood over the $k$ copies is approximately
$$H \approx \beta I_k + \frac{\alpha}{k}\,\mathbf{1}\mathbf{1}^{\top},$$
where $\alpha$ is the information from ambiguous reads, $\beta$ is the per-copy information from distinguishing reads, and $\mathbf{1}$ is the all-ones vector. When $\alpha \gg \beta$, which is the typical case for short-read-dominated coverage of a long repeat, single-copy coordinate updates converge at a rate controlled by roughly $(\alpha+\beta)/\beta$. This follows from the standard dependence of randomized coordinate descent on condition number reviewed in Coordinate Descent Algorithms. Now take any orthogonal $Q$ whose first column is $\mathbf{1}/\sqrt{k}$. The rotated Hessian $Q^{\top} H Q$ is exactly $\mathrm{diag}(\alpha+\beta, \beta, \dots, \beta)$. A normalized Walsh–Hadamard matrix has this property when $k$ is a power of two, and the Helmert matrix works for any $k$. Both apply in $O(k\log k)$ or $O(k)$ time.
The rotation does not use the random sign flips that QuaRot and QuIP rely on. Flipping the sign of the first column would destroy the consensus direction, so randomization is permitted only on the contrast columns. This is a deliberate departure from incoherence processing. Those methods rotate in order to spread energy across coordinates so that rounding is gentle. We rotate in order to align the coordinate blocks with the structure of the evidence so that block updates stop fighting each other.
What would be new is the following. Hybrid polishing is posed as continuous maximization of a GAML-style multi-platform likelihood. The variables are reparameterized per repeat family by a structured orthogonal transform. The problem is then solved by randomized block coordinate descent, with blocks defined as (genome window) × (consensus or contrast coordinates). To our knowledge, no work in the research file applies orthogonal reparameterization to assembly or polishing. Nor does any file paper connect block conditioning in coordinate descent to repeat ambiguity. The section of the file meant to list prior intersection work was truncated, and nothing visible in it covers this combination.
Relation to Prior Work
Two papers supply the objective and the structural signal. GAML defines a likelihood of several read sets given a candidate assembly and searches over discrete graph walks by simulated annealing. We keep its likelihood but fix the walk, taking the layout from a long-read assembler. We then optimize the base-level content continuously, replacing annealing with a solver that has a convergence rate. HINGE decides which repeats are bridged by at least one read. We use its bridged versus unbridged distinction to decide which repeat families get a full contrast basis, because their differences are identifiable. Unbridged families keep only the consensus coordinate, since their contrasts are not identifiable and should be left unresolved, not guessed.
Two lines of work supply the rotation machinery. QuIP shows that rounding error depends on how weights and their Hessian sit relative to the coordinate axes. It applies Kronecker-structured orthogonal transforms to make them incoherent. QuaRot uses fast Hadamard transforms and the invariance $Y = (XQ)(Q^{\top}W)$ to rotate at almost no cost. We reuse the fast structured transform and the exact-invertibility argument, but reverse the goal: we seek alignment with the Hessian, not incoherence. SpinQuant constrains our design. It found that random rotations vary by up to 13 zero-shot points across seeds and that learned rotations help. This motivates a learned-rotation ablation. It also motivates our fixed first column, which removes the seed sensitivity in the direction that matters most.
Three papers supply the optimizer. Randomized Block Cubic Newton Method gives block second-order updates with a cubic regularizer and reports a U-shaped runtime with optimal blocks of about 25–50 coordinates. This finding is central here. A family of 200 transposon copies across a 1 kb window is far too large for one Newton block. The rotation is what lets small blocks capture the coupling. Coordinate Descent Algorithms supplies the rates and warns that coordinate descent can stall on non-convex problems. We address that warning in the next section. Stochastic Primal-Dual Coordinate Method is a fallback if the read-likelihood terms are better handled in a dual form, with one dual variable per read. We do not expect to need it for the main result.
PECAT and Plassembler define the application regime. PECAT preserves heterozygous variants during long-read correction, a problem analogous to keeping repeat copies distinct. Plassembler isolates plasmids for focused hybrid assembly with Unicycler, which is exactly where resistance-gene repeat copies live. The learned assembly-graph methods, GatedGCN untangling and GraSSRep, work at the layout and repeat-detection stage, so they are upstream of and complementary to this work. GraSSRep's repeat calls could seed our family detection.
Research Program
The system has four stages. The first stage takes a long-read draft assembly and detects repeat families by self-alignment, producing aligned copy columns. The second labels each family as bridged or unbridged using HINGE-style read-bridging checks. The third parameterizes each column of each family as a profile over bases for each copy, under the Helmert or Hadamard rotation across copies. Columns in unique regions are left unrotated. The fourth runs randomized block coordinate descent on the negative log-likelihood. A block is one genome window crossed with one rotated coordinate type: the consensus, or a group of contrasts. Updates are cubic-regularized block Newton steps in the RBCN style, and block size is a tuned hyperparameter.
The objective is non-convex because each ambiguous read's likelihood is a mixture over copies. We handle this in two ways. First, we prove convergence of the rotated method on a convex surrogate obtained by linearizing the mixture weights. Second, we report empirical stationarity diagnostics, specifically gradient norms in both bases, on the true objective. The theory deliverable is narrow. For the two-component Hessian model above, plus a banded coupling between neighbouring columns created by reads spanning several columns, we will show that the rotated method's iteration count is independent of $\alpha/\beta$, while unrotated coordinate descent and expectation-maximization scale with it. We will then check whether this ratio, measured on real data, predicts the observed speedups.
Baselines
The naive reference point is the unpolished long-read draft, together with polishing that ignores repeats entirely by assigning each multi-mapping short read to a random best hit. This sets the floor for both error rate and copy collapse.
The strongest existing hybrid pipeline in the file is Unicycler as run inside Plassembler. It is the strongest because it is the tool practitioners actually use on the plasmid-rich bacterial genomes where repeat copies matter. We will run it with default and tuned settings on the same read sets and the same machine, and give it the same wall-clock budget as our method. We found no reported result on Lacuna for Unicycler's repeat-copy base accuracy. Plassembler's reported figures in the file concern plasmid recovery, not per-copy errors. Beating it means fewer substitution and indel errors per 100 kb inside repeat copies at matched error in unique regions, without more misassemblies.
The closest algorithmic competitor is GAML restricted to the same fixed layout, so that only base content is searched by its annealing moves. This is the strongest optimizer of the same objective. We will compare final log-likelihood and wall-clock time to a fixed likelihood gap. No reported result was found on Lacuna for GAML on repeat-copy polishing.
The most important baselines are three ablations that isolate the rotation itself. The first is identical block coordinate descent in the original per-copy basis. The second is expectation-maximization over read-to-copy assignments with exact per-copy updates. The third is a "whole family as one block" Newton solver for families small enough to permit it. If the rotated method does not beat the first two in iterations and time on families with high ambiguity, and match the third where the third is feasible, the core claim fails.
Why not just…?
Why not just run expectation-maximization, which already decouples copies once reads are assigned? Expectation-maximization does make each update trivially separable. However, its convergence rate is governed by the fraction of missing information, which here is exactly the fraction of reads whose copy of origin is ambiguous. In our model that fraction is the same $\alpha/\beta$ imbalance that slows coordinate descent. The prediction is that expectation-maximization behaves like unrotated coordinate ascent. Experiment 1 tests this directly by plotting iterations-to-tolerance against measured ambiguity for all three solvers. If expectation-maximization's curve is flat, the rotation adds nothing and we will say so.
Why not just put all copies of a family into one block and take an exact Newton step, which is rotation-invariant anyway? This is a real threat for small families, and we treat it as a baseline, not a strawman. Its limit is size. The block covers copies × window positions, and RBCN's own wall-clock profile favours blocks of tens of coordinates. A 30-copy insertion-sequence family over a 500-position window gives a block of 15,000 coordinates. The rotation achieves the decoupling with an $O(k\log k)$ transform and keeps blocks small. Experiment 2 sweeps family size to find the crossover point. If whole-family Newton wins up to the largest families seen in bacterial data, the proposal narrows to eukaryotic-scale repeats.
Why not just resolve repeats with long reads and then polish each copy using only uniquely mapping short reads? This discards most short-read evidence inside repeats, which is the evidence that pins down the shared sequence. It also fails exactly for unbridged repeats. Our method uses ambiguous reads for the consensus coordinate and distinguishing reads for contrasts, so it keeps both kinds of evidence. The unique-reads-only variant is cheap to run and is included as a baseline. It should match us on contrasts but lose on consensus accuracy in low-long-read-coverage settings.
Proposed Experiments
Experiment 1 tests the mechanism on simulated data. We take bacterial reference genomes with plasmids. Then we plant repeat families with controlled copy number ($k$ from 2 to 64), divergence (0.1%–2%), and long-read bridging fraction, and simulate short and long reads at varying error rates. We measure iterations and time to a fixed likelihood gap for the rotated method, the unrotated method, and expectation-maximization. This is the decisive check: the rotated method's cost should stay roughly flat as short-read ambiguity grows, while both baselines grow with the measured $\alpha/\beta$.
Experiment 2 is the block-size and family-size sweep against whole-family Newton described above. It reports RBCN-style U-curves in both bases.
Experiment 3 evaluates accuracy against the Plassembler/Unicycler pipeline and fixed-layout GAML. It uses simulated data and, where a curated reference exists, real hybrid bacterial datasets of the kind used in the file's resistance-plasmid studies, with the same reads, compute, and tuning budget for every method. The metrics are base errors per 100 kb split into repeat copies and unique sequence, the copy-collapse rate (copies whose true differences are erased), frameshifts in annotated genes inside repeats, and misassemblies.
Experiment 4 is the rotation ablation. It compares Helmert, Hadamard with random contrast signs, fully random orthogonal rotation (which breaks the consensus column), and a SpinQuant-style learned rotation. We expect the fully random rotation to lose most of the benefit and the learned rotation to recover little beyond the structured one. A large gain from the learned rotation would mean our two-component Hessian model is missing real structure, such as subfamilies within a repeat family, which a nested Helmert basis would then need to capture.
References
- Coordinate descent algorithms
- GAML: Genome Assembly by Maximum Likelihood
- HINGE: long-read assembly achieves optimal repeat resolution
- QuIP: 2-Bit Quantization of Large Language Models With Guarantees
- QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs
- SpinQuant: LLM Quantization with Learned Rotations
- Randomized Block Cubic Newton Method
- Stochastic Primal-Dual Coordinate Method for Regularized Empirical Risk Minimization
- De novo diploid genome assembly using long noisy reads
- Plassembler: an automated bacterial plasmid assembly tool
- Learning to Untangle Genome Assembly with Graph Convolutional Networks
- GraSSRep: Graph-Based Self-supervised Learning for Repeat Detection in Metagenomic Assembly