fastrho

Pedigrees and crossing experiments observe crossovers directly, the gold standard for recombination mapping, but they need hundreds of typed individuals and exist for only a handful of species. Statistical methods read recombination out of linkage disequilibrium (LD) instead, and those scale to any sequenced cohort — but LD is shaped by demography, mating system and structural variation as well as by the crossover rate. So the established LD tools ask the user to commit to a demographic model first: pyrho builds a lookup table under an assumed history, and ReLERNN simulates and trains a network for each target dataset. Every new population means another simulation-and-fitting cycle.

fastrho, from Kevin Korfmann and Sara Mathieson at University of Pennsylvania with Neda Rahnamae at Heinrich Heine University Düsseldorf, pays that cost once, at training time. It is a state-space neural network fitted to coalescent simulations drawn from a deliberately broad prior — constant, sawtooth, island and bottleneck histories across many population and sample sizes — and then frozen. Inference on a new cohort is a forward pass over a VCF: the checkpoint emits interval estimates of the population-scaled rate ρ = 4Ner without being told a demographic history, and the absolute rate follows as r = ρ/(4Ne) from a supplied or auxiliary Ne. That separation is what keeps map shape robust to demographic misspecification while confining the assumption to the final rescaling. It joins Popformer, from the same senior author, among neural population-genetic models that learn one reusable backbone rather than a per-dataset estimator.

#Key Features

• Amortized across datasets: One fixed checkpoint maps new cohorts, chromosomes and species with no target-specific retraining; the training cost is paid once.
• Bidirectional state-space encoder: Mamba-2 blocks scan the SNP-interval sequence forward and backward, collecting left and right LD context in linear time.
• Phase-agnostic input: A domain-randomized checkpoint accepts phased haplotypes, unphased genotypes and unpolarized dosages, substituting composite dosage LD and minor-allele folding when phase or ancestral state is missing.
• Uncertainty and an Ne head: Each interval gets a mean and variance of log ρ, and the decoder returns a pooled effective-population-size estimate used for absolute rescaling.
• Regime specialists: Separate checkpoints cover predominant selfing, structured Arabis panels, severe canid bottlenecks and very large Ne.

#Technical Details

Every biallelic segregating site becomes a 17-value token describing the interval to its right: geometry and allele-frequency features, mean r² at cumulative radii of 5, 25 and 50 kb on both flanks, adjacent-two-locus haplotype frequencies, and local diversity over a ±8-SNP neighbourhood. Tokens pass through parallel convolutions of width 3, 7 and 15 into 256 channels, then six encoder and four decoder bidirectional Mamba-2 blocks with hidden width 256, state size 64 and residual 256→1024→256 feed-forward layers. Context length is 1,024 intervals; longer sequences are chunked with 256-interval overlap and stitched under Hann weights. Training used 15,000 msprime-simulated regions under a beta-weighted Gaussian negative log-likelihood. At 25 kb resolution across six demographic scenarios, Pearson correlation with the true map ranged from 0.604 to 0.886 for fastrho and 0.444 to 0.840 for pyrho; in the paired bottleneck and expansion arms pyrho used lookup tables built under the simulation-generating history, which fastrho never saw. Window-averaged correlations were 0.942–0.978 for fastrho, 0.943–0.990 for pyrho and 0.847–0.946 for ReLERNN. Nominal 95% intervals covered the true rate for 95.4% of 365,280 held-out SNP intervals given the simulated Ne, and the full dataset-specific workflow over twenty 10-Mb regions took 113 s against 308 s for pyrho and 3.81 h for ReLERNN.

#Applications

The headline application is a recombination atlas covering five chromosome arms across 13 MalariaGEN Ag3.0 Anopheles gambiae, A. coluzzii and A. arabiensis populations, built with the high-Ne checkpoint and independently supported by crossover maps from 15 Ag3 colony pedigrees (Spearman rs = 0.692 at 5 Mb). Further analyses recover map shape in selfing Arabidopsis thaliana against meiotic references, compare Arabis population maps to a 2,082-marker hybrid cross, and dissect a ~55-Mb redpoll supergene where arrangement mixing creates an apparent recombination cold block. Any group holding a population VCF can map its own species.

#Impact

fastrho makes fine-scale recombination mapping a routine annotation step rather than a project, and it does so for the non-model organisms where crossing experiments are impractical. The authors are candid about the boundaries: the output is an LD-derived historical map, not a meiotic crossover map; absolute scaling stays conditional on Ne and the assumed mutation rate; gene conversion inflated the inferred rate in simulations that included it; and selfing, population structure and inversions each required a trained specialist rather than being absorbed by the broad prior. The released model cards add that validation seeds overlapped the first training seeds, so the recorded selection score is not evidence of generalization. What the work establishes is that amortization holds for recombination mapping: one training run, many genomes.