{"paper_id":"0499b330-a563-4b8d-a360-cb63e72c1590","body_text":"Phenome-wide genetic framework to identify mechanisms  \nof social effects \n \n \nHélène Tonnelé1,2, Francesco Paolo Casale3,4,5, Amelie Baud1,2,*  \n \n1Centre for Genomic Regulation, Barcelona Institute of Science and Technology, Spain;  \n2Universitat Pompeu Fabra, Barcelona, Spain  \n3Institute of AI for Health, Helmholtz Zentrum München – German Research Center for \nEnvironmental Health, Neuherberg,/i1 Germany  \n4Helmholtz Pioneer Campus, Helmholtz Zentrum München – German Research Center for \nEnvironmental Health, Neuherberg,/i1 Germany  \n5School of Computation, Information and Technology, Technical University of Munich, \nGarching,/i1 Germany  \n \nCorresponding author (*): Amelie Baud, amelie.baud@crg.eu \n  \n \nAbstract \n \nPhenotypes are shaped not only by an individual’s genotype (direct genetic effects, DGE) and \nenvironment but also by the genetic composition of social partners through indirect genetic \neffects (IGE). Although IGE have been detected across many traits and species, their \nmechanisms remain largely unknown, particularly for physiological traits. Here we introduce a \nphenome-wide genetic framework that identifies proxy phenotypes for the heritable traits of \nsocial partners mediating IGE by estimating genetic correlations between IGE on focal \nphenotypes and DGE on measured traits. Applying this approach to two large, outbred mouse \ndatasets comprising hundreds of behavioural and physiological phenotypes, we find that \nbehavioural traits are neither more affected by IGE nor better proxies for the traits mediating \nthem, challenging the prevailing behavioural-centric view of social effects. Instead, immune, \nmetabolic and growth phenotypes are both affected by IGE and informative proxies for their \nunderlying mechanisms, potentially reflecting the social transmission of gut microbes. Our \nframework provides a novel strategy to better understand the genetic basis of complex traits \nand uncover mechanisms of social effects. \n \nIntroduction \n \nPhenotypes do not arise in isolation but are shaped by the social environments in which \nindividuals develop and live. Across mammals, interactions with conspecifics influence a wide \nrange of biomedical traits, from early development to adult behaviour and physiology\n1. Yet, \nidentifying the mechanisms underlying social effects remains challenging, for several reasons: \nfirst, it is difficult to know a priori  all the traits of social partners that could influence the \nphenotype of interest. Secondly, socially interacting individuals tend to share resources and \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 11, 2026. ; https://doi.org/10.64898/2026.03.10.710784doi: bioRxiv preprint \n\n \nexposures, leading to pervasive correlations between their phenotypes 2. Disentangling genuine \nsocial influences from confounding effects therefore requires approaches that can move beyond \ndescriptive correlations. \n \nIndirect genetic effects 3,4 (IGE) are social effects of genetic origin. They arise when heritable \ntraits of social partners influence the phenotype of interest (Fig. 1A). There are several \nadvantages to using IGE to study social effects. First, IGE can be detected without specifying \nthe traits of social partners involved, by testing for associations between the phenotype of \ninterest and the genotypes of social partners\n5. Secondly, IGE can serve as causal anchors to \nestablish causal relationships since the genotypes of the social partners are assigned at birth \nand unaffected by environmental factors and phenotypes\n6. Finally, modelling IGE is essential to \nunderstand the evolution of socially affected phenotypes3,7.  \n \nFigure 1. Quantitative genetics framework to uncover new mechanisms of indirect \ngenetic effects. (A) Definition of key terms. The heritable trait(s) of social partners mediating \nIGE are unobserved. (B) Representation of the genetic correlation ρ between IGE on the \nphenotype of interest and DGE on a measured phenotype to identify good proxy phenotypes for \nthe heritable traits of social partners mediating IGE.  \n \nIGE have been uncovered across a variety of animal species, types of relationships (e.g. \nbetween parents and offspring, between males and females, between peers, etc.) and \nphenotypes\n4,8,9. Behavioural phenotypes are most commonly reported to be affected by IGE 9,10. \nStrong IGE are easy to understand for social behaviours (i.e. behaviours that are only \nexpressed when social partners are present) and behaviours that are “contagious”, as \nindividuals genetically predisposed to a high trait value can elicit higher trait values in their \npartners. For example, Peromyscus mice genetically predisposed to aggression can instigate \nhigher aggression in others 11 and an individual’s smoking behaviour is influenced by peer \nsmoking12.  \n \nBeyond behavioural traits, IGEs have been reported on a wide range of developmental, \nphysiological, morphological, reproduction and survival phenotypes\n9. In humans, robust IGE in \neighty thousand couples of the UK Biobank were reported for educational attainment and \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 11, 2026. ; https://doi.org/10.64898/2026.03.10.710784doi: bioRxiv preprint \n\n \narthrosis13. In laboratory mice, IGE were shown to affect biomedical markers, including immune, \nmetabolic and growth phenotypes 14,15. Although IGE are an important source of phenotypic \nvariation, the underlying mechanisms remain unknown, particularly where non-behavioural \nphenotypes are concerned.  \nTo address this gap and leverage the increasingly large number of biobanks available for \ngenetics research16–20, we introduce a genetic framework that identifies good proxy phenotypes \nfor the traits of social partners influencing a phenotype of interest on the basis of their genetic \ncorrelation ρ with the traits mediating IGE (Fig. 1B). Genetic correlations can arise through \ncausal relationships or pleiotropy and are widely used to identify functional relationships \nbetween phenotypes21. We apply this strategy to two datasets collected in thousands of “HS” \nand “CFW” outbred laboratory mice and including a wide range of behavioural, physiological, \nand morphological phenotypes relevant to human diseases\n22–24 (Fig. 2 and Supplementary \nTables 1 and 2). In these datasets, we previously detected IGE arising from the genotypes of \ncage mates on a wide range of phenotypes, many of which were not previously known to be \naffected by social effects 14,15. In most cases, the mechanisms of social effects are completely \nunknown. A better understanding of the traits of social partners mediating social effects could \nopen the door to preventive or therapeutic interventions based on a better management of \nsocial interactions\n25.  \n \n \n \nFigure 2. Phenome-wide data collected in HS and CFW mice.  Tests performed in both \npopulations are displayed along the grey timeline, while those specific to a single population are \nshown above (HS mice) or below (CFW mice) the line. The colour of the border line of each box \nindicates the category of the phenotypes collected (see also Supplementary Tables 1 and 2). \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 11, 2026. ; https://doi.org/10.64898/2026.03.10.710784doi: bioRxiv preprint \n\n \n \nResults \n \nRefining the leading view that behaviours are more affected by IGE than other \nphenotypes \n \nOur previous quantification of IGE in the HS and CFW mouse datasets suggested that \nbehaviours were not the class of phenotypes most strongly affected by IGE\n14,15, contrary to the \nprevailing view in the field and the results of a recent meta-analysis that did not include the HS \nand CFW datasets\n9. To address this apparent discrepancy, we systematically compared IGE on \nbehavioural and non-behavioural phenotypes in the mouse datasets and re-analysed the data of \nthe meta-analysis to distinguish between social and non-social behaviours, because the mouse \ndatasets did not include social behaviours. In mice, we found no evidence that behaviours were \nmore affected by IGE than non-behavioural phenotypes (one-tailed non-parametric Wilcoxon \nrank-sum test p-value = 0.91 for HS mice, p-value = 0.94 for CFW mice, Fig. 3A and 3B) . In the \nmeta-analysis, we found that the large IGE observed for behaviours were driven by social \nbehaviours (Fig. 3C), resolving the apparent discrepancy.  \n \n \nFigure 3. Comparison of the magnitude of indirect genetic effect (IGE) between \nbehavioural and non-behavioural phenotypes. (A) Our estimates from HS outbred laboratory \nmice. (B) Our estimates from CFW mice. Each dot represents an individual trait, coloured in \ngreen if non-behavioural or in pink if behavioural. (C) IGE estimates from Santostefano et al.\n9, \nshown for non-behavioural, non-social behavioural, social behavioural, and assumed social \nbehavioural traits; point size reflects estimate precision (inverse standard error). Dots are \ncoloured grey for phenotypic classes that are not present in our study.  \n \nValidation of a new implementation of bivariate IGE models \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 11, 2026. ; https://doi.org/10.64898/2026.03.10.710784doi: bioRxiv preprint \n\n \nTo identify good proxies for the traits of cage mates mediating IGE, we implemented a bivariate\nlinear mixed model that jointly models direct and indirect genetic effects across pairs of\nphenotypes. This model allows us to estimate the genetic correlation ρ between IGE on the\nphenotype of interest and DGE on each measured phenotype. A high correlation suggests that\nthe measured phenotype is genetically related to the unobserved traits of cage mates mediating\nIGE (Fig. 1B). We used simulations based on the real genotypes and cage structure of the HS\nand CFW datasets to confirm that the bivariate model yields unbiased genetic estimates and\nwell-calibrated p-values (Supplementary Figs. 1 and 2). \n \nBehaviours are not better proxies for the traits mediating IGE than non- behavioural\nphenotypes \n \nBecause social partners are often thought to exert their influence through behaviours , we first\ntested whether the behavioural phenotypes included in the HS and CFW datasets are better\nproxies for the traits mediating IGE than the other, non-behavioural phenotypes. We limited our\nanalysis to the phenotype pairs where IGE on one phenotype and DGE on the other phenotype\nwere strong enough (>5% of phenotypic variation explained). We saw no evidence of this, as\nthe correlations ρ  involving DGE on behavioural phenotypes were no greater than the\ncorrelations involving DGE on non-behavioural phenotypes (one-tailed non- parametric Wilcoxon\nrank-sum test p -value = 0.33 and 1.0 for behavioural and non-behavioural phenotypes ,\nrespectively, in HS mice; p-value = 0.57 and 0.85 for behavioural and non- behavioural\nphenotypes, respectively, in CFW mice; Fig. 4). Thus, behaviours are no better proxies for the\ntraits mediating IGE than non-behavioural phenotypes. \n \nFigure 4. Behavioural VS non-\nbehavioural phenotypes as proxies for the traits of cage\nmates mediating IGE. The comparison was done in both HS mice (A) and CFW mice (B) and\nseparately for IGE on behavioural phenotypes (top two rows) and IGE on non- behavioural\nphenotypes (bottom two rows). Each dot represents the IGE-DGE correlation ρ for a pair of\nphenotypes; the coloured dots and black lines show the mean and standard deviations for each\ndistribution. \n \nIdentifying good proxies for the heritable traits of cage mates mediating IGE \n \nte \nof \nhe \nat \nng \nS \nnd \nal \nrst \nter \nur \npe \nas \nhe \non \n, \nral \nhe \n \nge \nnd \nral \nof \nch \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 11, 2026. ; https://doi.org/10.64898/2026.03.10.710784doi: bioRxiv preprint \n\n \nTo gain insights into the traits of social partners mediating IGE, we visualized the IGE-DGE \ncorrelations ρ in a heat map where phenotypes of interest (rows) and potential proxy \nphenotypes (columns) were clustered based on the significance of the IGE-DGE correlation ( ρ ≠ \n0). We limited our analyses to the phenotype pairs where IGE on one phenotype and DGE on \nthe other phenotype were strong enough (>5% of phenotypic variation explained). In HS mice \n(Fig. 5A), we observed a cluster featuring immune, metabolic and growth phenotypes as both \ntraits affected by IGE and good proxies for the traits mediating these IGE. More specifically, the \nimmune phenotypes included counts and size of T and B lymphocytes, expression of CD4 at the \nsurface of T cells, and measures of airway resistance following immunization with ovalbumin. \nMetabolic traits included glycemia before and 15 minutes after intraperitoneal glucose injection, \nand serum ion levels. The growth phenotypes were measures of body weight collected \nthroughout the life of the mice, as well as adult body length. Because many of the immune and \nmetabolic phenotypes were blood and serum phenotypes, we wondered whether the clustering \nobserved could be due to batch effects associated with the blood draws and subsequent \nprocessing. Since the blood and serum phenotypes were derived from four separate blood \ndraws spread across the life of the mice and four separate analytical procedures (see Methods), \nwe considered that batch effects were extremely unlikely to explain the clustering and instead \nconcluded that the high genetic correlations observed arose from causal relationships or \npleiotropy, but not experimental confounding.  \nIn CFW mice, IGE-DGE correlations were generally less significant, but we nevertheless \nobserved two clusters (Fig. 5B). The larger one was highly heterogeneous in terms of the types \nof the phenotypes involved. The smaller one featured T lymphocytes counts and ratios as \nphenotypes affected by IGE and proxy phenotypes for the traits mediating these IGE (Fig. 5B). \nThus, the smaller cluster is reminiscent of the cluster observed in HS mice (Fig. 5A).  \nAltogether, these results indicate a shared genetic architecture between DGE and IGE affecting \nT lymphocyte phenotypes. In HS mice, these genetic effects also influence metabolic and \ngrowth phenotypes directly and indirectly (i.e. through social effects). \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 11, 2026. ; https://doi.org/10.64898/2026.03.10.710784doi: bioRxiv preprint \n\n \n \n \nFigure 5. Clustered heat maps of IGE-DGE correlations ρ in (A) HS and (B) CFW mice. \nPhenotypes affected by IGE (rows, >5% of phenotypic variation explained) and potential proxy \nphenotypes affected by DGE (columns, >5% of phenotypic variation explained) were clustered \nbased on the significance of the IGE-DGE correlation ( ρ ≠ 0), which was calculated using a \nlikelihood ratio test (see Methods). In HS mice (A) the black square highlights a cluster of \nimmune, metabolic and growth phenotypes; in CFW mice (B) the black square on the top left \nhighlights a large, heterogeneous cluster, whereas the square in the middle highlights a smaller \ncluster of T lymphocytes phenotypes.  \n \nSimilar traits mediate IGE in groups of males and groups of females \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 11, 2026. ; https://doi.org/10.64898/2026.03.10.710784doi: bioRxiv preprint \n\n \nIn our analyses we implicitly modelled IGE as arising from the same traits of cage mates in \ngroups of males and groups of females. To evaluate this assumption, focusing on a single \nphenotype of interest, we split the phenotype between a male phenotype and a female \nphenotype, assigning missing values to females for the male phenotype and vice versa. Doing \nso, the genetic correlation between IGE on the male phenotype and IGE on the female \nphenotype reflects the extent to which similar genetic variants, hence similar biological \npathways, give rise to the traits of social partners mediating IGE in groups of males and in \ngroups of females. After confirming using simulations that genetic estimates and p-values were \nalso valid in this analysis (Supplementary Fig. 3 and Supplementary Fig. 4), we estimated the \nmale-female genetic correlation for each phenotype included in the dataset and tested whether \nit was different from one, which would provide evidence that different traits of social partners \nunderlie IGE in groups of males and groups of females. The correlations ranged from 0.29±0.56 \nto 0.99±0.36 in HS mice (mean 0.87) and from -0.27±0.43 to 1±1.21 in CFW mice (mean 0.63). \nNo correlation was significantly different from 1 (FDR > 0.1), providing no evidence that different \ntraits of social partners mediate IGE in male and female groups.  \n \nDiscussion \nIn this study, we introduce a phenome-wide, genetics-based strategy to gain insights into the \ntraits of social partners mediating indirect genetic effects (IGE). By leveraging genetic \ncorrelations between IGE on phenotypes of interest and direct genetic effects (DGE) on a broad \narray of measured traits, we identify phenotypes that act as informative proxies for the heritable \ntraits of social partners influencing these phenotypes of interest. Applying this approach to two \nindependent outbred mouse populations provided evidence that similar genetic effects act on \nadaptive immunity phenotypes directly and indirectly. This signal was visible in both datasets, \nand in HS mice these genetic effects were also correlated with DGE and IGE on metabolic and \ngrowth phenotypes. \nRevisiting the behavioural-centric view of social genetic effects \nIGE are most intuitively understood for social behaviours (i.e. behaviours that are expressed \nonly in the presence of other individuals) and behaviours that are contagious, because \nindividuals genetically predisposed to a high trait value can elicit higher trait values in their \npartners (positive feedback loops). Consistent with this intuition, behavioural traits have \ndominated the IGE literature and were concluded to be particularly affected by IGE in a recent \nmeta-analysis. However, by re-analysing the data from the meta-analysis with a finer distinction \nbetween social and non-social behaviours, and by analysing two independent mouse datasets \nthat included contagious behaviours (anxiety and depressed mood/stress coping strategy 26) but \nno social behaviours, we showed that the tendency for behaviours to be strongly affected by \nIGE is driven by social behaviours. Our findings therefore refine our understanding of IGE on \nbehavioural phenotypes. Furthermore, in the mouse datasets non-behavioural phenotypes \nexhibited IGE of comparable—or even greater—magnitude as non-social behavioural \nphenotypes, and the best proxies for traits of social partners mediating IGE were non-\nbehavioural traits. These results call for a broader conceptualization of social genetic effects, in \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 11, 2026. ; https://doi.org/10.64898/2026.03.10.710784doi: bioRxiv preprint \n\n \nwhich behaviours represent only one of several possible conduits through which genetic \nvariation in social partners can influence focal individuals. Importantly, this broader view is \nnecessary to explain the strong and reproducible IGE observed for physiological phenotypes, \nsuch as immune traits, metabolism, and growth phenotypes. \nPhenome-wide genetic correlations as a window into social mechanisms \nThe central methodological advance of this work is the use of phenome-wide genetic \ncorrelations between IGE and DGE to identify measured traits that are good proxies for the \nunobserved traits mediating social effects. This strategy builds on the logic of genetic correlation \nanalyses widely used to study shared disease aetiology 21, which exploits the causal anchoring \nprovided by genotypes to reduce confounding relative to phenotypic correlations. Although \ngenetic correlations do not, by themselves, distinguish between causality and pleiotropy, both \nmechanisms are biologically meaningful in the context of IGE, as they reflect shared heritable \npathways that shape social environments and influence evolutionary dynamics. The validation \nof our bivariate IGE models through extensive simulations confirms that these correlations can \nbe estimated with minimal bias, even in complex social designs. This opens the door to \nsystematic, unbiased scans of phenome datasets to generate mechanistic hypotheses about \nsocial effects, without requiring prior knowledge of the relevant traits or behaviours. Such \nphenome datasets are increasingly available in humans, laboratory, wild and agricultural \npopulations. \nGut microbiota transmission as a plausible mechanism for physiological IGE \nAcross both mouse populations, behavioural phenotypes were not better proxies for the traits \nmediating IGE than non-behavioural phenotypes. Instead, the most robust signal emerged from \nT lymphocytes phenotypes, which were affected by IGE and good proxies for the traits of social \npartners mediating these IGE in both datasets. In HS mice, DGE and IGE on metabolic and \ngrowth phenotypes also shared a genetic basis with DGE and IGE on immune phenotypes. \nBlood lymphocytes are very unlikely to be passed between mice sharing a cage, and growth \nphenotypes simply cannot. Hence, these phenotypes are merely proxies for the traits mediating \nIGE: they are related to those traits through causal relationships or pleiotropy. Which factors, \nthen, can be passed between mice sharing a cage and affect blood lymphocytes levels and, \nalso, metabolic and growth phenotypes? A plausible hypothesis is commensal gut bacteria, \nwhich are well-known to be transferred between mice sharing a cage through allo-coprophagy \nand affect the immunity, metabolism, and growth of their host (Fig. 6). This hypothesis is further \nsupported by recent evidence from our lab that demonstrated IGE on a subset of gut bacteria in \nlaboratory rats27. Alternatively, stress and its transmission could explain these signals. The lack \nof correlation between DGE on the anxiety and stress-coping phenotypes included in both \nmouse datasets28 and IGE on immune, metabolic and growth phenotypes, however, does not \nsupport this alternative hypothesis. Future work integrating experimental manipulations of the \ntraits suspected to mediate IGE—such as targeted manipulations of the gut microbiota of one \nmouse and evaluation of the immune, metabolic and growth phenotypes of the other mice in the \ncage—will be essential to directly test the proposed mechanisms.  \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 11, 2026. ; https://doi.org/10.64898/2026.03.10.710784doi: bioRxiv preprint \n\n \nMany animals besides rodents are coprophagic and ingest faeces from conspecifics 29. Humans \nare not coprophagic, yet there is ever increasing evidence of pervasive horizontal transmission \nof commensal gut microbes in our species, including within the household and at nurseries 30–33. \nThe social transmission of gut microbes could mediate social effects, including IGE, on immune, \nmetabolic and growth phenotypes in other animal species, including ours.  \n \nFigure 6. Hypothesized mechanism for the high correlations between IGE and DGE on \nimmune phenotypes (dashed lines): gut microbiome transfers through allo-coprophagy.  \nSimilar social mechanisms in males and females \nWe formally compared IGE in male-only and female-only groups and found no evidence that \ndifferent mechanisms are at play in the two sexes, justifying the analysis of IGE in males and \nfemales jointly. There are well-established differences in social behaviours between males and \nfemales in mice, including in the laboratory setting, but since (social) behaviours do not seem to \nbe the main mechanisms for IGE in our study, the lack of sex differences in the mechanisms of \nIGE is not unexpected. \nLimitations and future directions \nSeveral limitations of this study warrant consideration. First, the phenotypes available in HS and \nCFW mice do not include measures of social behaviours such as aggression or affiliation, \nlimiting our ability to implicate social behaviours in IGE. Secondly, albeit similar, the phenotypes \nincluded in the HS and CFW mouse datasets were not the same, limiting our ability to identify \nmechanisms that generalise across genetic backgrounds and experimental designs, hence are \nmore robust. Finally, in the HS dataset only, DGE and IGE were confounded with maternal \n(genetic) effects. We checked that the IGE-DGE correlations analysed in this study were largely \nunaffected by this confounding (Supplementary Fig. 5).  \nDespite these limitations, our results demons trate the ability of phenome-wide, genetics-based \napproaches to fundamentally expand the exploration of social effects beyond behaviour. As \nincreasingly rich phenotypic and multi-omics datasets become available in both animal models \nand humans, this strategy will become widely applicable, with potential  implications for \nmedicine, conservation, and animal breeding.  \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 11, 2026. ; https://doi.org/10.64898/2026.03.10.710784doi: bioRxiv preprint \n\n \n \nMethods \n \nStatistical models with direct and indirect genetic effects  \n \nUnivariate models (used for Fig. 3): \nThe following model, which is the same as the models we used in two of our previous studies of \nIGE41,42, was used to quantify DGE, IGE and their correlation, while accounting for non-genetic \nfactors contributing to phenotypic variation: \n/g1877 /g3404  /g1850/g1854 /g3397 /g1853 /g3005  /g3397 /g1852 /g1853 /g3020  /g3397 /g1857 /g3005  /g3397/g1852 /g1857 /g3020 /g3397 /g1849 /g1855                              (1) \n/g1877  is the vector of phenotypic residuals (phenotype of interest), /g1850  is a vector describing how \nmany rats are in each cage and /g1854  the corresponding estimated fixed effect, /g1849  is the matrix of \ncage assignments and /g1855  the corresponding vector of random cage effects. /g1853 /g3005  is the vector of \nrandom additive DGE, /g1853 /g3020  is the vector of random additive IGE and /g1852  is the matrix indicating, for \neach mouse, which are the cage mates (importantly /g1852 /g3036,/g3036  /g3404  0 ). /g1857 /g3005  and /g1857 /g3020 , also random effects, \nrefer to the non-genetic component of direct and indirect effects. \nThe joint distribution of all random effects was defined as: \n \n/g1743\n/g1742\n/g1742\n/g1742\n/g1741\n/g1853 /g3005\n/g1853 /g3020\n/g1857 /g3005\n/g1857 /g3020\n/g1855 /g1746\n/g1745\n/g1745\n/g1745\n/g1744\n~ /g1839/g1848/g1840 \n/g1737\n/g1736/g1736/g1736\n/g1735\n0,\n/g1743\n/g1742\n/g1742\n/g1742\n/g1742\n/g1742\n/g1741 /g2026 /g3002 /g3253\n/g2870 /g1827/g2026 /g3002 /g3253/g3268/g18270 0 0\n/g2026 /g3002 /g3253/g3268/g1827/g2026 /g3002 /g3268\n/g2870 /g18270 0 0\n00 /g2026 /g3006 /g3253\n/g2870 /g1835/g2026 /g3006 /g3253/g3268/g18350\n00 /g2026 /g3006 /g3253/g3268/g1835/g2026 /g3006 /g3268\n/g2870 /g18350\n00 0 0 /g2026 /g3004\n/g2870 /g1835 /g1746\n/g1745\n/g1745\n/g1745\n/g1745\n/g1745\n/g1744\n \n/g1740\n/g1739/g1739/g1739\n/g1738\n    (2) \n \n/g2026 /g3002 /g3268\n/g2870  reflects the proportion of phenotypic variance explained by IGE and is the estimate reported \nin Fig. 3.  \n \nBivariate models (used for Figs. 4 and 5): \nThe main estimate of interest in this study is the genetic correlation ρ between IGE on a \nphenotype of interest and DGE on a measured phenotype that is being evaluated as a potential \nproxy phenotype. The joint distribution of all random effects in the bivariate model was defined \nas: \n \n/g1743\n/g1742\n/g1742\n/g1742\n/g1742\n/g1742\n/g1742\n/g1742\n/g1742\n/g1742\n/g1741\n/g1853 /g3005/g2869\n/g1853 /g3005/g2870\n/g1853 /g3020/g2869\n/g1853 /g3020/g2870\n/g1857 /g3005/g2869\n/g1857 /g3005/g2870\n/g1857 /g3020/g2869\n/g1857 /g3020/g2870\n/g1855 /g2869\n/g1855 /g2870/g1746\n/g1745\n/g1745\n/g1745\n/g1745\n/g1745\n/g1745\n/g1745\n/g1745\n/g1745\n/g1744\n ~ /g1839/g1848/g1840\n/g1737\n/g1736\n/g1736\n/g1736\n/g1736\n/g1736\n/g1736\n/g1736\n/g1736\n/g1736\n/g1735\n0,\n/g1743\n/g1742\n/g1742\n/g1742\n/g1742\n/g1742\n/g1742\n/g1742\n/g1742\n/g1742\n/g1742\n/g1742\n/g1741 /g2026 /g3002/g3253/g3117\n/g2870/g1827/g2026 /g3002/g3253/g3117/g3253/g3118/g1827/g2026 /g3002/g3253/g3117/g3268/g3117/g1827/g2026 /g3002/g3253/g3117/g3268/g3118/g18270 0 0 0 0 0\n/g2026 /g3002/g3253/g3117/g3253/g3118/g1827/g2026 /g3002/g3253/g3118\n/g2870/g1827 /g2026 /g3002/g3253/g3118/g3268/g3117/g1827/g2026 /g3002/g3253/g3118/g3268/g3118/g18270 0 0 0 0 0\n/g2026 /g3002/g3253/g3117/g3268/g3117/g1827/g2026 /g3002/g3253/g3118/g3268/g3117/g1827/g2026 /g3002/g3268/g3117\n/g2870/g1827/g2026 /g3002/g3268/g3117/g3268/g3118/g18270 0 0 0 0 0\n/g2026 /g3002/g3253/g3117/g3268/g3118/g1827/g2026 /g3002/g3253/g3118/g3268/g3118/g1827/g2026 /g3002/g3268/g3117/g3268/g3118/g1827/g2026 /g3002/g3268/g3118\n/g2870/g18270 0 0 0 00\n000 0 /g2026 /g3006/g3253/g3117\n/g2870/g1835/g2026 /g3006/g3253/g3117/g3253/g3118/g1835/g2026 /g3006/g3253/g3117/g3268/g3117/g1835/g2026 /g3006/g3253/g3117/g3268/g3118/g18350 0\n000 0 /g2026 /g3006/g3253/g3117/g3253/g3118/g1835/g2026 /g3006/g3253/g3118\n/g2870/g1835/g2026 /g3006/g3253/g3118/g3268/g3117/g1835/g2026 /g3006/g3253/g3118/g3268/g3118/g18350 0\n000 0 /g2026 /g3006/g3253/g3117/g3268/g3117/g1835/g2026 /g3006/g3253/g3118/g3268/g3117/g1835/g2026 /g3006/g3268/g3117\n/g2870/g1835/g2026 /g3006/g3268/g3117/g3268/g3118/g18350 0\n000 0 /g2026 /g3006/g3253/g3117/g3268/g3118/g1835/g2026 /g3006/g3253/g3118/g3268/g3118/g1835/g2026 /g3006/g3268/g3117/g3268/g3118/g1835/g2026 /g3006/g3268/g3118\n/g2870/g18350 0\n000 0 0 0 0 0 /g2026 /g3004/g3117\n/g2870/g1835/g2026 /g3004/g3117/g3118/g1835\n000 0 0 0 0 0 /g2026 /g3004/g3117/g3118/g1835/g2026 /g3004/g3118\n/g2870/g1835 /g1746\n/g1745\n/g1745\n/g1745\n/g1745\n/g1745\n/g1745\n/g1745\n/g1745\n/g1745\n/g1745\n/g1745\n/g1744\n/g1740\n/g1739\n/g1739\n/g1739\n/g1739\n/g1739\n/g1739\n/g1739\n/g1739\n/g1739\n/g1738\n   (3)  \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 11, 2026. ; https://doi.org/10.64898/2026.03.10.710784doi: bioRxiv preprint \n\n \n/g2026 /g3002 /g3253/g3118/g3268/g3117 corresponds to the correlation of interest, which is reported in Fig. 4 and Fig. 5 and called \nρ throughout this manuscript. \n \nTo evaluate whether IGE in groups of males and groups of females arise from different \nmechanisms, we used the same bivariate model but, in that case, one of the outcomes ( /g1877 /g2869 ) was \nthe female phenotype (males had missing values) and the other ( /g1877 /g2870 ) the male phenotype \n(females had missing values). For this analysis only, /g2026 /g3002 /g3268/g3117/g3268/g3118 was the estimate of interest. \n \nThe model was developed within a flexible linear mixed-model framework 34,35 and parameter \nestimation was performed by restricted maximum likelihood using gradient-based optimisation. \n \nSignificance values for the correlations were obtained using likelihood ratio tests, comparing a \nmodel with unconstrained correlations to a model in which the correlation of interest was fixed to \nthe value tested (0 for \n/g2026 /g3002 /g3253/g3118/g3268/g3117 and 1 for /g2026 /g3002 /g3268/g3117/g3268/g3118). The LRT statistic for each test was compared to \nthe chi-square distribution with one degree of freedom to get the p-value. To account for \nmultiple testing, we controlled the false discovery rate using the function p.adjust from R stats \npackage. \n \nDescription of the mouse datasets  \nHeterogeneous Stock ( HS) mice: we used the same cages, phenotypes, and genetic \nrelatedness matrix as in Baud et al.\n14. The data were originally published by Valdar et al. 22,23 \n(genotypes and organismal phenotypes) and Huang et al. 36 (organismal phenotype: cellular \nproliferation in the subgranular zone of the dentate gyrus, a measure of adult neurogenesis). In \nbrief, 2,448 male and female mice were included in this study. They were related at various \nlevels, housed in same-sex groups of 2 to 7 mice (but mostly 3 and 4 mice), and the groups \nincluded many (but not only) full siblings. Genotypes at 13,459 single nucleotide polymorphisms \nwere available for a subset of 1,940 mice, and pedigree data for all mice. The genetic \nrelatedness matrix is the single-step (or H) matrix, constructed using both the genotypes and \nthe pedigree. Some phenotype data were available for all 2,448 mice, but the number of mice \nphenotyped for each organismal phenotype varied. Covariates such as sex, body weight, and \ngroup size were considered and phenotypes normalised using a covariate-aware Box-Cox \ntransformation. The covariates were fitted as fixed effects in all analyses of HS data. \n \nSwiss Webster (Crl:CFW(SW)-US_P08, CFW) mice: we used the same cages, phenotypes and \ngenetic relatedness matrix as in Baud et al.\n15. The data were originally published by Nicod et \nal.24. In brief, 1,812 male and female mice were included in this study. They were unrelated to \none another (beyond the baseline relatedness associated with the maintenance of the colony) \nand housed in same-sex groups of 3 mice. Genotypes at 353,697 LD-pruned single nucleotide \npolymorphisms were available for all mice and used to build the genetic relatedness matrix. \nSome phenotype data were available for all mice, but the number of mice phenotyped for each \norganismal phenotype varied. Covariates such as sex, body weight and group size were \nconsidered and phenotypes normalised using a covariate-aware Box-Cox transformation. The \nresiduals of a linear model including the covariates as fixed effects were used in all analyses of \nCFW data. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 11, 2026. ; https://doi.org/10.64898/2026.03.10.710784doi: bioRxiv preprint \n\n \n \nSpecific information about the blood and serum phenotypes collected in HS and CFW \nmice \nIn HS mice,  blood was collected, first, in naive mice and analysed by FACS (“Imm.” \nphenotypes); then just before and at regular intervals after intraperitoneal glucose injection, and \nanalysed with a glucose analyser (“Glucose” phenotypes); and finally after cardiac puncture for \nanalysed using a medical-grade haematology analyser (“Haem” phenotypes) and an automated \nclinical chemistry analyser (“Biochem” phenotypes). \nIn CFW mice, blood was collected after cardiac puncture, for FACS, haematology and \nbiochemistry analysis.  \n \nExperimental design and confounding \nConfounding between DGE, IGE and cage effects (in HS mice only): Because the average \nrelatedness of co-housed mice is greater than the average relatedness of non co-housed mice \nin the HS dataset, DGE and IGE are partially confounded. To the extent that the cage \nenvironment affects the phenotypes studied here, they are also partially confounded with cage \neffects. However, we demonstrated both before\n14 and again in this study that accounting for \nDGE, IGE and cage effects jointly in our models of phenotypic variation yields unbiased \nestimates of DGE, IGE and IGE-DGE univariate and bivariate genetic correlations. \nConfounding between DGE and parental effects (in HS and CFW mice):  DGE could be partially \nconfounded with parental effects, to the extent that the pre-weaning parental environment \naffects the adult phenotypes studied here. \nConfounding between IGE and parental effects (in HS mice only): Because the average \nrelatedness of co-housed mice is greater than the average relatedness of non co-housed mice \nin the HS dataset, IGE are partially confounded with parental effects, to the extent that the pre-\nweaning parental environment affects the adult phenotypes studied here. We checked that the \nIGE-DGE correlations analysed in this study were largely unaffected by this confounding \n(Supplementary Fig. 5). \n \nSimulations to validate our new implementation of bivariate IGE-DGE models \n(Supplementary Table 3 and Supplementary Fig. 1) \nFor each dataset, we simulated 1,000 pairs of “mock” phenotypes using the bivariate version of \nmodel (1), the real genetic relatedness matrix and cage assignments, and parameter values for \nthe phenotypic covariance (3) that mimicked a favourable scenario with relatively large DGE \nand IGE, and intermediate correlation values (Supplementary Table 3). To do so, we used the \nmvn() function from the R MASS package. \nWe analysed each pair of simulated phenotypes the same way we analysed the real \nphenotypes and, for each parameter, examined the difference between the simulated and the \nestimated value of the parameter (Supplementary Fig. 1). \n \nNull simulations to validate p-values for \n/g2251 /g2157 /g2160/g2779/g2175/g2778≠  0 (Supplementary Table 4 and \nSupplementary Fig. 2) \nWe selected five phenotypic pairs from CFW mice with increasing p-values: “Haem.MPV”- \n“Haem.Large_PLT” (p-value of 1·10 -4), “Bioch.LDL_B Wcorr”-“Haem.Large_PLT” (p-value of \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 11, 2026. ; https://doi.org/10.64898/2026.03.10.710784doi: bioRxiv preprint \n\n \n0.001), “Hypoxia.TV_SHR_BW corr”-“Neuro.Ki67_BWcorr” (p-val ue of 0.01), “Haem.MPV”-\n“Bioch.Iron_BWcorr” (p-value of 0.1), and “Bioch.CreatinineEnzymatic_BWcorr”-\n“Bioch.ALP_BWcorr” (p-value of 0.99). The smallest p-value, 10 -4, is close to the FDR-adjusted \nsignificance threshold (p ≤  2.4·10-4). We first fitted the null model with /g2025 /g3002 /g3253/g3118/g3268/g3117/g34040   and generated \nnull simulations, as described in the previous section, setting the parameter values to the \nparameter values estimates obtained from these null models. We simulated 10,000 pairs for the \nphenotype pair with the 1·10\n-4 p-value and 1,000 pairs for the other four phenotype pairs. \nWe analysed each pair of simulated phenotypes the same way we analysed the real \nphenotypes and derived LRT statistics and p-values for the hypothesis that /g2025 /g3002 /g3253/g3118/g3268/g3117/g34050 . \nTo evaluate our p-value, we compared in a quantile-quantile plot the p-values from the null \nsimulations to a uniform distribution of values between 0 and 1, which is the expected \ndistribution of p-values if \n/g2025 /g3002 /g3253/g3118/g3268/g3117/g34040  (Supplementary Fig. 2).  \n \nSimulations to validate sex-specific bivariate IGE-DGE models (Supplementary Table 3 \nand Supplementary Fig. 3) \nFor each dataset, we simulated 1,000 “mock” phenotypes using a similar strategy to the one \ndescribed above. We assigned missing values to males to simulate the female phenotype (\n/g1877 /g2869 ), \nmissing values to females for the male phenotype ( /g1877 /g2870 ) and finally concatenated them in a single \nphenotype. We used the same favourable scenario with large DGE and IGE, and intermediate \ncorrelation values (Supplementary Table 3). \nWe analysed each pair of simulated phenotypes the same way we analysed the real \nphenotypes and, for each parameter, examined the difference between the simulated and the \nestimated value of the parameter (Supplementary Fig. 3). \n \nNull simulations to validate p-values for \n/g2251 /g2157 /g2175/g2778/g2175/g2779 ≠  1 (Supplementary Table 5 and \nSupplementary Fig. 4) \nWe selected the phenotype from CFW mice with the smallest p-value: “Haem.EOS_percent” (p \nof 0.004). We first fitted the null model with /g2025 /g3002 /g3268/g3117/g3268/g3118/g34041   and generated null simulations, as \ndescribed in the previous section, setting the parameter values to the parameter values \nestimates obtained from the null model of the real phenotype. \nWe analysed each phenotype the same way we analysed the real phenotypes and derived LRT \nstatistics and p-values for the hypothesis that \n/g2025 /g3002 /g3268/g3117/g3268/g3118/g34051 . \nTo evaluate our p-value, we compared in a quantile-quantile plot the p-values from the null \nsimulations to a uniform distribution of values between 0 and 1, which is the expected \ndistribution of p-values if \n/g2025 /g3002 /g3268/g3117/g3268/g3118/g34041  (Supplementary Fig. 4).  \n \nComputational performance \n \nThe analysis of the phenotype pairs from the HS mice dataset took between 12 minutes and 19 \nhours and 25 minutes, using 1.2 to 4.9 GB of RAM. For CFW mice, runtime ranged from 14 \nminutes to 3 hours and 8 minutes, with RAM usage between 1.3 and 3.0 GB. \n \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 11, 2026. ; https://doi.org/10.64898/2026.03.10.710784doi: bioRxiv preprint \n\n \nData availability \n \nAll data are available from published datasets. Details are provided on the mouse data used in \nthis study in the Methods section. Data from Santostefano et al.\n9 were also re-analysed. \n \nCode availability \n \nThe computational pipeline used in this study is available from https://github.com/Baud-\nlab/CoreQuantGen. \n \nREFERENCES \n1. Lemonnier, C. et al. Effects of the social environment on vertebrate fitness and health in \nnature: Moving beyond the stress axis. Horm. Behav. 145, 105232 (2022). \n2. Cohen-Cole, E. & Fletcher, J. M. Detecting implausible social network effects in acne, \nheight, and headaches: longitudinal analysis. BMJ 337, a2533 (2008). \n3. Wolf, J. B., Iii, E. D. B., Cheverud, J. M., Moore, A. J. & Wade, M. J. Evolutionary \nconsequences of indirect genetic effects. Trends Ecol. Evol. 13, 64–69 (1998). \n4. Baud, A., McPeek, S., Chen, N. & Hughes, K. A. Indirect Genetic Effects: A Cross-\ndisciplinary Perspective on Empirical Studies. J. Hered. 113, 1–15 (2022). \n5. Bijma, P. The quantitative genetics of indirect genetic effects: a selective review of \nmodelling issues. Heredity 112, 61–69 (2014). \n6. Burgess, S. & Thompson, S. G. Mendelian Randomization: Methods for Causal Inference \nUsing Genetic Variants. (CRC Press, 2021). \n7. Bijma, P. & Wade, M. J. The joint effects of kin, multilevel selection and indirect genetic \neffects on response to genetic selection. J. Evol. Biol. 21, 1175–1188 (2008). \n8. Bailey, N. W. & Desjonquères, C. The Indirect Genetic Effect Interaction Coefficient ψ : \nTheoretically Essential and Empirically Neglected. J. Hered. 113, 79–90 (2022). \n9. Santostefano, F., Moiron, M., Sánchez-Tójar, A. & Fisher, D. N. Indirect genetic effects \nincrease the heritable variation available to selection and are largest for behaviors: a meta-\nanalysis. Evol. Lett. 9, 89–104 (2025). \n10. Bailey, N. W., Marie-Orleach, L. & Moore, A. J. Indirect genetic effects in behavioral \necology: does behavior play a special role in evolution? Behav. Ecol. 29, 1–11 (2018). \n11. Wilson, A. J., Gelin, U., Perron, M.-C. & Réale, D. Indirect genetic effects and the evolution \nof aggression in a vertebrate system. Proc. R. Soc. B Biol. Sci. 276, 533–541 (2008). \n12. Sotoudeh, R., Harris, K. M. & Conley, D. Effects of the peer metagenomic environment on \nsmoking behavior. Proc. Natl. Acad. Sci. 116, 16302–16307 (2019). \n13. Xia, C., Canela-Xandri, O., Rawlik, K. & Tenesa, A. Evidence of horizontal indirect genetic \neffects in humans. Nat. Hum. Behav. 5, 399–406 (2021). \n14. Baud, A. et al. Genetic Variation in the Social Environment Contributes to Health and \nDisease. PLOS Genet. 13, e1006498 (2017). \n15. Baud, A. et al. Dissecting indirect genetic effects from peers in laboratory mice. Genome \nBiol. 22, 216 (2021). \n16. Gallagher, C. S., Ginsburg, G. S. & Musick, A. Biobanking with genetics shapes precision \nmedicine and global health. Nat. Rev. Genet. 26, 191–202 (2025). \n17. Li, H. & Auwerx, J. Mouse Systems Genetics as a Prelude to Precision Medicine. Trends \nGenet. 36, 259–272 (2020). \n18. Baud, A. et al. Combined sequence-based and genetic mapping analysis of complex traits \nin outbred rats. Nat. Genet. 45, 767–775 (2013). \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 11, 2026. ; https://doi.org/10.64898/2026.03.10.710784doi: bioRxiv preprint \n\n \n19. Dwinell, M. R. et al. Establishing the hybrid rat diversity program: a resource for dissecting \ncomplex traits. Mamm. Genome 36, 25–37 (2025). \n20. Sparks, A. M. et al. The genetic architecture of helminth-specific immune responses in a \nwild population of Soay sheep (Ovis aries). PLOS Genet. 15, e1008461 (2019). \n21. Bulik-Sullivan, B. et al. An atlas of genetic correlations across human diseases and traits. \nNat. Genet. 47, 1236–1241 (2015). \n22. Valdar, W. et al. Genome-wide genetic association of complex traits in heterogeneous stock \nmice. Nat. Genet. 38, 879–887 (2006). \n23. Solberg, L. C. et al. A protocol for high-throughput phenotyping, suitable for quantitative trait \nanalysis in mice. Mamm. Genome 17, 129–146 (2006). \n24. Nicod, J. et al. Genome-wide association of multiple complex traits in outbred mice by ultra-\nlow-coverage sequencing. Nat. Genet. 48, 912–918 (2016). \n25. https://biologue.plos.org/2017/03/16/understanding-images-how-genetic-makeup-of-a-\nroommate-can-influence-health/. \n26. Haeffel, G. J. & Hames, J. L. Cognitive Vulnerability to Depression Can Be Contagious. Clin. \nPsychol. Sci. 2, 75–85 (2014). \n27. Tonnelé, H. et al. Genetic architecture and mechanisms of host-microbiome interactions \nfrom a multi-cohort analysis of outbred laboratory rats. Nat. Commun. 16, 10126 (2025). \n28. Packard, A. E. B., Egan, A. E. & Ulrich-Lai, Y. M. HPA Axis Interactions with Behavioral \nSystems. Compr. Physiol. 6, 1897–1934 (2016). \n29. Power, E. J., Bornbusch, S. L. & Kendrick, E. L. Faeces as food: a framework for adaptive \nnutritional coprophagy in vertebrates. Anim. Behav. 218, 75–86 (2024). \n30. Sarkar, A. et al. Microbial transmission in the social microbiome and host health and \ndisease. Cell 187, 17–43 (2024). \n31. Heidrich, V., Valles-Colomer, M. & Segata, N. Human microbiome acquisition and \ntransmission. Nat. Rev. Microbiol. 23, 568–584 (2025). \n32. Valles-Colomer, M. et al. The person-to-person transmission landscape of the gut and oral \nmicrobiomes. Nature 614, 125–135 (2023). \n33. Ricci, L. et al. Baby-to-baby strain transmission shapes the developing gut microbiome. \nNature 651, 191–200 (2026). \n34. Lippert, C., Casale, F. P., Rakitsch, B. & Stegle, O. LIMIX: genetic analysis of multiple traits. \n003905 Preprint at https://doi.org/10.1101/003905 (2014). \n35. Casale, F. P., Rakitsch, B., Lippert, C. & Stegle, O. Efficient set tests for the genetic \nanalysis of correlated traits. Nat. Methods 12, 755–758 (2015). \n36. Huang, G.-J. et al. A Genetic and Functional Relationship between T Cells and Cellular \nProliferation in the Adult Hippocampus. PLOS Biol. 8, e1000561 (2010). \n \n \nAcknowledgements \n \nThis work was supported by PID2021-122651NA-I00 and PRE2021-097413 contracts funded by \nMCIN/AEI/10.13039/501100011033 and FSE+ to AB and HT. We acknowledge support of the \nSpanish Ministry of Science and Innovation through the Centro de Excelencia Severo Ochoa \n(CEX2020-001049-S, MCIN/AEI /10.13039/501100011033), the Generalitat de Catalunya \nthrough the CERCA programme and to the EMBL partnership. FPC was funded by the Free \nState of Bavaria’s Hightech Agenda through the Institute of AI for Health (AIH). \n \nAuthor contributions \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 11, 2026. ; https://doi.org/10.64898/2026.03.10.710784doi: bioRxiv preprint \n\n \n \nAB designed the study, with support from FPC. HT, FPC and AB developed the code. HT \nperformed the analyses. HT, FPC and AB wrote the manuscript. \n \nCompeting interests \n \nThe authors declare no competing interests. \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted March 11, 2026. ; https://doi.org/10.64898/2026.03.10.710784doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}