henhaobio
Sign in

bioRxivBioinformaticsPreprint

CycloCross: Adversarial single-cell RNA-seq data translation across species

Olympio Hacquard, Yuya TokutaiD, Yusuke ImotoiD, Tomonori Nakamura, Shuto DeguchiiD, Masahiro NaganoiD, Mitinori SaitouiD, Yasuaki HiraokaiD

PDFDOI
In-depth full-text digest available · Jump to digest ↓

Abstract

Single-cell RNA sequencing (scRNA-seq) has transformed our understanding of cellular heterogeneity, yet most datasets are limited to a handful of model organisms, leaving critical gaps in cross-species biology. Existing computational methods for integrating multi-species scRNA-seq data often rely on one-to-one ortholog mapping, which fails to account for gene duplications, losses, or functional divergences. To address these challenges, we introduce CycloCross, an adversarial method based on the CycleGAN architecture, designed to translate scRNA-seq data between species without requiring a priori gene correspondences. Unlike generative models that synthesize data from noise, CycloCross leverages real data from a source species, enabling more realistic and data-efficient translations. We demonstrate its effectiveness in germ cell development, translating data between mouse, macaque, and human with conserved yet divergent transcriptional programs. CycloCross not only preserves temporal dynamics but also enables interpolation and extrapolation of gene expression at missing time points, offering predictions for experimentally inaccessible states. Furthermore, CycloCross generates biologically plausible samples even with limited data, outperforming traditional generative models.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (30 Sept 2026) · CC BY 4.0

Translating cells between mouse, macaque and human scRNA-seq with a CycleGAN, without ortholog mapping, preserved developmental time structure, imputed missing states and produced more realistic synthetic cells than scGAN.

As the authors describe it, cross-species methods depend on gene correspondences (one-to-one orthologs for Icebear; protein language models plus precomputed annotations for SATURN and UNIFY), and scRNA-seq generators build cells from latent noise. CycloCross instead translates real cells of one species directly into another species' expression space without a priori gene matching, even when the two gene sets differ.

Key results

  1. SuggestedIn primordial germ cell-like cell (PGCLC) induction time courses, time-conditioned translation kept cells grouped by time stamp near real cells of related stages for the human-macaque and human-mouse pairs shown, even with the small mouse set, and reproduced species-specific trends of pluripotency factors.

    UMAPs of real and translated cells (qualitative). SOX2 falls in human and macaque but rises to a day-4 peak in mouse, and human-mouse translation reproduces these opposite trends; the POU5F1 rise and late human decline are captured; the NANOG day-1 minimum is missed for macaque and only partly reproduced for human. Datasets: 11,277 human, 8,432 macaque, 1,761 mouse cells. · CycloCross translates temporal data with…

  2. ShownTranslating a source-species state, with the matching target state held out, predicted the target's pseudo-bulk expression: clearly for human embryonic stem cells (ESCs) from macaque ESCs, only marginally for human mesoderm-like cells (iMeLCs) from mouse epiblast-like cells (EpiLCs).

    Across 1,144 transcription factors (TFs) with one-to-one human-macaque orthologs, CycloCross was significantly better (no test named) than cubic-spline extrapolation from the other human time points and than naive reuse of macaque ESC values, with no aberrant values; across 742 orthologous TFs for iMeLC it was only slightly better than spline interpolation. Errors are shown as box plots; no values are printed. · CycloCross translates temporal data with…; Pseudo-bulk gene expression interpolation; Figure legends

  3. ShownTranslating real mouse or macaque cells into human space gave more realistic synthetic human PGCLC-induction cells than scGAN, a Wasserstein GAN that samples from noise; mouse-sourced cells beat macaque-sourced cells on every metric.

    With 1,000 human cells held out, 1,000 generated cells were scored by Euclidean and cosine distance between real and generated centroids and by random-forest AUROC on 50 principal components (0.5 = indistinguishable), averaged over five splits and five generations. CycloCross beat scGAN on all three; mouse-translated distances neared the control of resampled training cells. Values appear only in figures. · CycloCross generates realistic scRNA-seq data; Resemblance of real and simulated cells; Figure legends

  4. ShownApplied within one species, translation acts as batch correction: it ranks among the top methods on a PBMC benchmark, with a tunable trade-off between batch mixing and cell-type separation, and merges two-lab human PGCLC data that Harmony and bbkNN left separated.

    PBMC 10x 3' vs 5' (8,098 and 7,378 cells; 1,000 variable genes) vs 13 methods on average silhouette width and adjusted Rand index (published scores) and kBET: λCT = 2 made batches indistinguishable by k-means but mixed CD4 and CD8 T cells; λCT = 10 separated them. PGCLC (15,203 vs 11,277 cells, λCT = 5): all time-point clusters merged, unlike Harmony and bbkNN on 966 shared TFs. · CycloCross removes batch effects through…; Batch integration; Tables

  5. SuggestedA ridge classifier separating real human cells from mouse-translated human-like cells ranks the genes the translation fails to reproduce, flagging candidate human-mouse divergence genes such as HMGA1, POU5F1, FOXM1 and MESP1.

    About 10% of total importance falls in the top 30 genes; importance correlates with mean expression (r = 0.588), and the top three are also the most highly expressed. MESP1 and GATA6 rank 8th and 76th, while EOMES, SOX17, TFAP2C and PRDM1 rank 317, 610, 1,229 and 789. No experimental validation; the authors note that conserved hits (YBX1, CENPX) may be model artifacts. · Identification of contrastive genes between…

Methods, data and reproducibility

Two fully connected generators (hidden layers 1024-2048-4096-4096) and two discriminators are trained adversarially with a cycle-consistency loss on unpaired, log-normalized transcription-factor expression from two species, with time or cell type appended as a condition vector weighted by λCT. New mouse PGCLC-induction 10x scRNA-seq was generated for the study.

  • Human PGCLC induction time course: ESC, iMeLC, PGCLC days 1-4 (Clark lab) · human · n = 11,277 cells
  • Macaque PGCLC induction: ESC and PGCLC days 1, 2, 4, 6, 8 (private until journal publication) · macaque · n = 8,432 cells; 1,180 transcription factors
  • Mouse EpiLC and day 2, 4, 6 PGCLC generated here from BDF1-2-1 ESCs with Blimp1-mVenus/Stella-ECFP reporters, 10x 3' v3.1 (private until journal publication) · mouse · n = 1,761 cells; 786 transcription factors
  • Human PGCLC induction from induced pluripotent stem cells (Saitou lab), used for two-lab batch integration · human · n = 15,203 cells
  • Human peripheral blood mononuclear cells (PBMC), 10x Genomics 3' and 5' batches from one donor, nine cell types, 17,430 genes · human · n = 8,098 and 7,378 cells
Methods & tools
CycleGAN, scGAN, RECODE, Seurat v3, Harmony, bbkNN, kBET, GroundGAN evaluation pipeline, Cell Ranger v6.0.1, Scrublet v0.2.3, 10x Genomics Chromium Single Cell 3' v3.1, Illumina NovaSeq 6000
Reagents & environment
Versions are given for preprocessing tools only (Cell Ranger v6.0.1, Scrublet v0.2.3, Seurat v3); no deep-learning framework, container or requirements file is stated.
Experimental conditions
About an hour to integrate two batches of roughly 10,000 cells and 1,000 genes, versus minutes for Harmony and bbkNN; hardware not stated.
Reproducibility gaps
Code and preprocessed data are said to be freely available, but no link or accession is printed; macaque and mouse PGCLC data stay private until journal publication; the PBMC data are described only as available online.
Limitations (6)
  • Costlier and less interpretable than standard integrators: about an hour for two batches of roughly 10,000 cells versus minutes for Harmony or bbkNN, and B > 2 batches would need 2(B - 1) generators and B discriminators. · stated by the authors
  • Translation yields only as many target cells as there are source cells; producing more requires first simulating source cells with a latent-space generator. · stated by the authors
  • No comparison with other cross-species methods: SATURN and CONCORD lack a translation pipeline, and the public Icebear code failed to run. · stated by the authors
  • Networks are retrained from scratch for each species pair, and reuse of a pre-trained model on new data is untested; integration is asymmetric, and the choice of reference dataset affects quality. · stated by the authors
  • Despite the no-gene-matching framing, preprocessing kept only transcription factors from a human list applied to macaque and mouse by gene name; hyperparameters were hand-tuned on the Saitou/Clark PGCLC integration task that is also used for evaluation. · noted in this summary
  • Cross-species evidence comes from one system (PGCLC induction) with a 1,761-cell mouse set; the PBMC ASW/ARI comparison reuses published scores for 13 methods rather than rerunning them, and imputation errors and AUROCs are not printed in the text. · noted in this summary

Read this if you need to move scRNA-seq time-course data between species with non-matching gene sets, impute an unmeasured developmental state from a related species, or integrate batches with large lab-specific effects.

Citations
1

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISGuo '26Riva '25Suthahar '25Wu '25Nouri '25Zhong '25
64% semantic matchBioinformatics
Zhen-Hao Guo, De-Shuang Huang, Shihua Zhang·Nature Communications·Feb 21, 2026· 4 citations