New Rust Toolkit Supercharges Sex Determination Studies in Non-Model Organisms

A new open-source software package promises to transform how biologists hunt for the genetic switches that separate males from females in species with no reference genome. The tool, called rsx, was unveiled in the journal BMC Bioinformatics by Rohit Goswami of TurtleTech ehf and the École Polytechnique Fédérale de Lausanne, together with Ruhila Goswami of the University of Iceland. It is a complete re-engineering of RADSex, the most widely used computational workflow for finding sex-linked markers in restriction site-associated DNA sequencing data, and it delivers dramatic speed and memory gains while reproducing the original tool’s statistical calls with perfect fidelity. In benchmarks spanning four published RAD-seq panels and 41.9 billion sequenced bases, rsx was 8.38 times faster than RADSex v1.2.0 on average across 56 paired timings, and its optional graphics-card backend accelerated the most demanding step by nearly 30-fold.

The problem rsx tackles is deceptively simple to state but computationally brutal in practice. RAD-seq allows researchers to sample thousands of genomic positions scattered across the chromosomes of organisms whose genomes have never been sequenced in full, from fish and frogs to plants and fungi. By comparing how often each of these DNA fragments appears in males versus females, scientists can flag markers that show up almost exclusively in one sex, which are candidates for the sex-determining region. The standard analysis, implemented in the RADSex tool, builds an enormous table recording the sequencing depth of every marker in every individual, then tests each row for a statistically significant sex bias. As sequencing panels grow into the millions of distinct RAD tags, that table balloons, and the original C++ implementation becomes memory-hungry to the point of failure on ordinary workstations.

rsx attacks the memory bottleneck with a set of techniques drawn from modern systems programming and database engineering. DNA sequences are compressed into 2-bit packed keys, storing four nucleotides per byte with a one-byte header that preserves the exact sequence length. After the initial parallel reading of FASTQ files, every downstream command operates either on a memory-mapped marker table or on externally sorted, lz4-compressed chunks on disk, which means the tool never copies the whole dataset into RAM. The writable memory footprint is instead bounded by the number of individuals in the study, or by an explicit user-chosen buffer, and is completely independent of how many millions of markers the panel contains. This allows researchers to analyse tables with vast numbers of rare tags without discarding candidates, a down-sampling sacrifice the memory-hungry original sometimes forced.

Speed comes from a combination of parallelism and clever data structures. Group membership is encoded in bitsets built once from the population map, with one 64-bit machine word covering 64 individuals; intersecting a marker’s presence pattern with the male and female masks then costs 64 times fewer operations than scanning individuals one by one. Producer-consumer pipelines built on crossbeam channels stream parsed records from a parser thread to per-marker consumers, and external sorting with k-way merges handles order statistics and table merging without loading everything at once. Even the median depth, a routine quality-control statistic, is computed exactly from only the nonzero entries of each depth vector, exploiting the fact that a given RAD tag is absent in most individuals, which shrinks the sorting workload dramatically.

The most scientifically consequential addition, however, is statistical rather than computational. RADSex reported only frequentist significance calls, offering no measure of how strongly the evidence actually supports sex linkage. rsx augments every marker with conjugate Beta-Binomial Bayes factors and directional posterior probabilities, computed in closed form through log-gamma functions and log-sum-exp evaluation, so no sampling or numerical quadrature is needed. Each marker is then graded into one of three explicit tiers: a strict call meeting the conventional significance threshold, a hypothesis supported by posterior evidence, or a row supported only by its Bayes factor. A configurable prior probability that any given tag is sex-linked, along with adjustable Beta prior shapes on the group prevalence rates, lets researchers tailor the evidence model to their organism. Every output row also records the exact uncorrected p-value, the Bayes factor and the posterior, together with machine-readable provenance headers documenting the parameters used.

For dimensionality reduction, rsx streams a Gram matrix accumulation rather than ever materialising the full marker-by-individual matrix. By the fundamental theorem of the singular value decomposition, the eigenvectors of the n-individual by n-individual Gram matrix are exactly the principal component loadings that a full-matrix SVD would produce, so the result is identical up to sign and numerical tolerance while memory drops from scaling with both markers and individuals to scaling only with the square of the individual count. This identity, long used for out-of-core PCA on genotype matrices, is here derived and verified with SymPy scripts and a Lean model of the median selection rule, all shipped with the code. The authors also performed a forward-error analysis of the Yates-corrected chi-squared path, showing that 64-bit integer cross-products before floating-point conversion keep reported p-values accurate to a few units in the last place, and that p-values are clamped at ten to the minus sixteen to match the original reference and avoid spurious claims of exact zeros.

Validation was rigorous and, unusually for method papers, close to exhaustive. Across four published RAD-seq panels, rsx reproduced every call made by RADSex v1.2.0, recovering all Bonferroni-significant positive-control markers, and matched the original’s six-significant-digit output formatting exactly, including a small correction step for the decimal exponent that prevents off-by-one errors on powers of ten. Precision test suites check that the p-value evaluators track the complementary error function to one part in ten to the fifteenth and never leave the valid probability interval. The optional CUDA backend batches the per-marker arithmetic on the GPU, adding up to 29.86-fold acceleration on the p-value computation batch, while Python and C bindings expose the same core to notebooks and workflow engines, closing a long-standing gap since the original offered no programmatic interface.

The authors were careful to scope their comparisons honestly. Alternative tools such as SEX-DETector, which relies on RNA-seq segregation in controlled crosses, and findZX, which uses whole-genome coverage profiles, consume fundamentally different input data and cannot be benchmarked head-to-head on identical RAD-seq panels without manufacturing measurements that were never made. The team did execute both competitors on their authors’ own example datasets, confirming installation and end-to-end operation, but stopped short of claiming relative biological accuracy. What rsx offers instead is a drop-in replacement: the complete RADSex command set is preserved, command-line compatibility is maintained, and marker-table semantics are unchanged, so existing pipelines can adopt the new engine without modification while gaining the Bayesian evidence grades as a bonus.

Released under the GPL-3.0-or-later licence with a reproducibility archive covering every reported number, rsx arrives at a moment of surging interest in sex chromosome evolution across ecology and evolutionary biology. Researchers studying species with environmental or polygenic sex determination, or tracking the early stages of sex chromosome differentiation in recently diverged lineages, now have a tool that scales to contemporary sequencing volumes on laptops rather than clusters, and that reports not just a binary verdict but a graded, transparent accounting of the evidence for every marker. For a field long constrained by the computational ceiling of its standard workflow, the arrival of an allocation-bounded, statistically richer and GPU-accelerated successor marks a genuine widening of what non-model genomics can attempt.

Subject of Research: A high-performance Rust software toolkit for identifying sex-linked markers from RAD-seq data in non-model organisms

Article Title: rsx: a high-performance streaming toolkit for RAD-seq sex determination

Article References: Goswami, R., & Goswami, R. (2026). rsx: a high-performance streaming toolkit for RAD-seq sex determination. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06628-4

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06628-4

Keywords: RAD-seq, sex determination, Rust, bioinformatics, streaming algorithms, Bayesian statistics, GPU computing, non-model organisms, sex-linked markers, BMC Bioinformatics, genomics, high-performance