Found In Translation (FIT; Normand et al., Nature Methods 2018) trains a per-gene regularized linear model on 170 curated mouse-human disease expression pairings to translate a new mouse experiment into predicted human differential-expression signatures, improving recovery of human-relevant genes by 20-50% where it applies. The paper itself left the limits of this paradigm unquantified: how achievable translation accuracy depends on training-compendium scale and diversity, why translation succeeds in inflammatory and infectious diseases but fails in domains such as cancer, whether model classes beyond the per-gene linear map close the residual gap, and how to evaluate translation when human reference data are themselves inter-study discordant (the silver-standard problem). The successor literature - TransCompR, AutoTransOP, FlowTransOP, congruence scoring, and single-cell comparative frameworks - has advanced the model-class and applicability facets, yet no study quantifies the scaling of translation accuracy with compendium size and diversity, explains domain-specific failure such as cancer mechanistically, resolves how much of the residual cross-species gap each model class removes, or establishes an evaluation protocol robust to the choice of discordant human references. The limit-quantification problem therefore remains open.

Background

A central obstacle in translational medicine is the cross-species gap: findings in mouse disease models frequently fail to recapitulate human biology, and cross-species differences have not been systematically incorporated into the interpretation of animal-model data. At the level of gene expression, the basic readout of such experiments is the set of differentially expressed genes (DEGs, genes whose expression differs significantly between disease and control samples), usually summarized per gene by an effect size such as a fold-change or z-statistic. Because orthologous mouse and human genes (genes derived from the same ancestral gene in the two species) can diverge in regulation, simply carrying mouse effect sizes over to human gene symbols (direct extrapolation) produces many false leads and misses many true human disease genes.

Found In Translation (FIT; Normand et al., 2018) introduced a data-reuse paradigm addressing this gap. The authors manually curated a compendium of 170 Cross-Species Pairings (CSPs, matched disease-vs-control expression datasets from mouse models and the corresponding human condition) covering 28 human diseases. For each gene gg and each pairing pp, FIT learns a regularized linear model ZpH(g)=αg+βgZpM(g)Z^{H}_{p}(g)=\alpha_{g}+\beta_{g}Z^{M}_{p}(g) linking mouse to human effect sizes, fit by lasso regression (an L1-penalized least-squares method that shrinks weak coefficients to zero) with penalties that shrink the slope toward 1 and the intercept toward 0, so that the mouse value is retained unless the training evidence supports a shift. The fitted models convert a new mouse experiment's effect sizes into predicted human effect sizes. Across the compendium, FIT increased the overlap with human DEGs by 20-50% relative to direct-from-mouse analysis, predominantly in infectious and inflammatory diseases, while performing poorly in cancer; an SVM applicability classifier (a support vector machine trained on the first 50 principal components of the mouse expression data) predicted on average in 80% of test cases whether FIT would improve a given dataset.

FIT's own Discussion flags what remains unknown. Performance is stated to be a complex function of the training compendium's size, the disease, the tissues, and the experiment's similarity to the compendium, with performance and applicable breadth expected to rise as the compendium grows - but no quantitative relationship is established. Proposed augmentations (explicit modeling of gene relations such as pathways, disease-disease relations, more sophisticated weighting of training comparisons) were left undeveloped. Evaluation relies on human data that the authors call a silver standard: the most logical comparison target, yet internally inconsistent, since different human datasets of the same disease report widely different DEGs. The paradigm's extension to other data types and species was proposed but not demonstrated.

Subsequent work has built a successor landscape rather than closing these questions. TransCompR (Brubaker et al., 2019) translates genomic responses via semi-supervised neural-network components and was applied to inflammatory bowel disease (Brubaker et al., 2020), osteoarthritis (Frost et al., 2026), tuberculosis infection (Pullen et al., 2025), and BCG vaccination (Bridges et al., 2025). AutoTransOP (Meimetis et al., 2024) maps omics profiles into a shared autoencoder latent space (a compressed representation learned by a neural network) without requiring one-to-one orthologues and outperforms FIT on L1000 benchmark data (L1000 is a high-throughput gene-expression profiling platform); FlowTransOP (Meimetis et al., 2026, preprint) aligns full distributional differences via constrained deep flow matching (a deep-learning technique that learns to transform one full data distribution into another) without paired samples and provides a model-selection guideline across four translational regimes. Network-aware multi-species methods (Mancuso et al., 2024), genome-wide functional-conservation scoring (Kwon and Ernst, 2021), transcriptomic congruence analysis for evaluating model organisms (Zong et al., 2023), the single-cell comparative method singIST (Moruno-Cuenca et al., 2026), and a review of transfer learning from preclinical molecular data (Kowald et al., 2022) address adjacent facets, and the field review of Yuan et al. (2026) synthesizes this landscape while listing benchmarking and disease-facet capture as key remaining challenges. None of these works quantifies how translation accuracy scales with compendium size and diversity, attributes domain-specific failures such as cancer to their causes, or establishes validation robust to discordant human references.

Problem Statement

Establish, quantitatively and mechanistically, the achievable limits of compendium-based mouse-to-human transcriptomic translation - the regime in which a model trained on public paired mouse-human disease expression data (as in FIT, the Found In Translation model) recovers human disease-associated signals from new mouse experiments better than direct cross-species extrapolation (carrying mouse effect sizes over to the corresponding human genes unchanged). The required components of this objective are: (1) Scaling - how does achievable translation accuracy, measured on held-out conditions, vary with the size and diversity of the training compendium of cross-species pairings, and where do diminishing returns set in? (2) Failure domains - why does translation succeed in infectious and inflammatory diseases yet fail in domains such as cancer, and can the causes (inadequate mouse models, model-class mismatch, or data limitations) be separated and remedied? (3) Model class - do models that go beyond the per-gene regularized linear map, such as network- or pathway-aware models, orthologue-free latent-space models, or distribution-level translation, systematically close the residual cross-species gap relative to FIT and to one another? (4) Validation - how should translation quality be evaluated when human reference datasets of the same disease are sparse and mutually discordant (the silver-standard problem), such that conclusions are robust to the choice of reference data? The extension of this paradigm to other species and data modalities is an optional extension rather than a required component.

References

  1. Normand R, Du W, Briller M, Gaujoux R, Starosvetsky E, Ziv-Kenet A, Shalev-Malul G, Tibshirani RJ, Shen-Orr SS. Found In Translation: a machine learning model for mouse-to-human inference. Nature Methods 15(12):1067-1073 (2018). DOI: 10.1038/s41592-018-0214-9. https://doi.org/10.1038/s41592-018-0214-9
  2. Brubaker DK, Proctor EA, Haigis KM, Lauffenburger DA. Computational translation of genomic responses from experimental model systems to humans. PLOS Computational Biology 15(1):e1006286 (2019). DOI: 10.1371/journal.pcbi.1006286. https://doi.org/10.1371/journal.pcbi.1006286
  3. Brubaker DK, Kumar MP, Chiswick EL, Gregg C, Starchenko A, Vega PN, Southard-Smith AN, Simmons AJ, Scoville EA, Coburn LA, Wilson KT, Lau KS, Lauffenburger DA. An interspecies translation model implicates integrin signaling in infliximab-resistant inflammatory bowel disease. Science Signaling 13(643):eaay3258 (2020). DOI: 10.1126/scisignal.aay3258. https://doi.org/10.1126/scisignal.aay3258
  4. Kwon SB, Ernst J. Learning a genome-wide score of human-mouse conservation at the functional genomics level. Nature Communications 12:2495 (2021). DOI: 10.1038/s41467-021-22653-8. https://doi.org/10.1038/s41467-021-22653-8
  5. Kowald A, Barrantes I, Moller S, Palmer D, Murua Escobar H, Schwerk A, Fuellen G. Transfer learning of clinical outcomes from preclinical molecular data, principles and perspectives. Briefings in Bioinformatics 23(3):bbac133 (2022). DOI: 10.1093/bib/bbac133. https://doi.org/10.1093/bib/bbac133
  6. Zong W, Rahman T, Zhu L, Zeng X, Zhang Y, Zou J, Liu S, Ren Z, Li JJ, Sibille E, Lee AV, Oesterreich S, Ma T, Tseng GC. Transcriptomic congruence analysis for evaluating model organisms. Proceedings of the National Academy of Sciences 120(6):e2202584120 (2023). DOI: 10.1073/pnas.2202584120. https://doi.org/10.1073/pnas.2202584120
  7. Mancuso CA, Johnson KA, Liu R, Krishnan A. Joint representation of molecular networks from multiple species improves gene classification. PLOS Computational Biology 20(1):e1011773 (2024). DOI: 10.1371/journal.pcbi.1011773. https://doi.org/10.1371/journal.pcbi.1011773
  8. Meimetis N, Pullen KM, Zhu DY, Nilsson A, Hoang TN, Magliacane S, Lauffenburger DA. AutoTransOP: translating omics signatures without orthologue requirements using deep learning. npj Systems Biology and Applications 10:13 (2024). DOI: 10.1038/s41540-024-00341-9. https://doi.org/10.1038/s41540-024-00341-9
  9. Su Z, Fang M, Smolnikov A, Vafaee F, Dinger ME, Oates EC. Genes in Humans and Mice: Insights from Deep learning of 777K Bulk Transcriptomes. bioRxiv (2024). DOI: 10.1101/2024.04.01.587517. https://doi.org/10.1101/2024.04.01.587517
  10. Pullen KM, Finethy R, Ko SHB, Reames CJ, Sassetti CM, Lauffenburger DA. Cross-species transcriptomics translation reveals a role for the unfolded protein response in Mycobacterium tuberculosis infection. npj Systems Biology and Applications 11:19 (2025). DOI: 10.1038/s41540-024-00487-6. https://doi.org/10.1038/s41540-024-00487-6
  11. Bridges K, Awany D, Gela A, Mwambene TD, Kurtz SL, Baker RE, Elkins KL, Sassetti CM, Scriba TJ, Lauffenburger DA. Cross-species blood transcriptional correlates of BCG-mediated protection against tuberculosis include innate and adaptive immune processes. JCI Insight 10(24):e194450 (2025). DOI: 10.1172/jci.insight.194450. https://doi.org/10.1172/jci.insight.194450
  12. Yuan H, Mancuso CA, Johnson K, Braasch I, Krishnan A. Computational strategies for cross-species knowledge transfer. Nature Methods 23(2):312-327 (2026; online 2025-12-12). DOI: 10.1038/s41592-025-02931-9. https://doi.org/10.1038/s41592-025-02931-9
  13. Frost MR, Ball BK, Pendyala M, Douglas SR, Brubaker DK, Chan DD. Computational translation of mouse models of osteoarthritis predicts human disease. Osteoarthritis and Cartilage 34(3):484-495 (2026). DOI: 10.1016/j.joca.2025.09.010. https://doi.org/10.1016/j.joca.2025.09.010
  14. Moruno-Cuenca A, Picart-Armada S, Bogle R, Fox J, Tsoi LC, Gudjonsson JE, Perera-Lluna A, Fernandez-Albert F. singIST: An integrative method for comparative single-cell transcriptomics between disease models and humans. PLOS Computational Biology 22(3):e1014002 (2026). DOI: 10.1371/journal.pcbi.1014002. https://doi.org/10.1371/journal.pcbi.1014002
  15. Meimetis N, Hoang TN, Magliacane S, Lauffenburger DA. FlowTransOP: Distributional Translation of Omics Signatures via Constrained Deep Flow Matching. Preprint, posted content (2026). DOI: 10.64898/2026.05.27.728305. https://doi.org/10.64898/2026.05.27.728305