Integrated in vivo combinatorial functional genomics and spatial transcriptomics of tumours to decode genotype-to-phenotype relationships

Breinig, Marco; Lomakin, Artem; Heidari, Elyas; Ritter, Michael; Rukhovich, Gleb; Böse, Lio; Butthof, Luise; Wendler-Link, Lena; Wiethoff, Hendrik; Poth, Tanja; Sahm, Felix; Schirmacher, Peter; Stegle, Oliver; Gerstung, Moritz; Tschaharganeh, Darjus F.

doi:10.1038/s41551-025-01437-1

Download PDF

Article
Open access
Published: 28 July 2025

Integrated in vivo combinatorial functional genomics and spatial transcriptomics of tumours to decode genotype-to-phenotype relationships

Marco Breinig ORCID: orcid.org/0009-0000-9740-2986¹^na1,
Artem Lomakin²^na1^nAff11,
Elyas Heidari^2,3^na1,
Michael Ritter^4,5,
Gleb Rukhovich²,
Lio Böse¹,
Luise Butthof¹,
Lena Wendler-Link¹,
Hendrik Wiethoff ORCID: orcid.org/0000-0002-8450-653X¹,
Tanja Poth⁶,
Felix Sahm ORCID: orcid.org/0000-0001-5441-1962^4,5,
Peter Schirmacher¹,
Oliver Stegle³,
Moritz Gerstung ORCID: orcid.org/0000-0001-6709-963X^2,7,8,9 &
…
Darjus F. Tschaharganeh^1,10

Nature Biomedical Engineering (2025)Cite this article

10k Accesses
18 Altmetric
Metrics details

Subjects

Abstract

Advancing spatially resolved in vivo functional genomes will link complex genetic alterations prevalent in cancer to critical disease phenotypes within tumour ecosystems. To this end, we developed PERTURB-CAST, a method to streamline the identification of perturbations at the tissue level. By adapting RNA-templated ligation probes, PERTURB-CAST leverages commercial 10X Visium spatial transcriptomics to integrate perturbation mapping with transcriptome-wide phenotyping in the same tissue section using a widely available single-readout platform. In addition, we present CHOCOLAT-G2P, a scalable framework designed to study higher-order combinatorial perturbations that mimic tumour heterogeneity. We apply it to investigate tissue-level phenotypic effects of combinatorial perturbations that induce autochthonous mosaic liver tumours.

Live cell tagging tracking and isolation for spatial transcriptomics using photoactivatable cell dyes

Article Open access 17 August 2021

Spatial deconvolution of HER2-positive breast cancer delineates tumor-associated cell type interactions

Article Open access 14 October 2021

Spatial genomics maps the structure, nature and evolution of cancer clones

Article Open access 09 November 2022

Main

Cancer, like many other complex diseases, is caused by a combination of multiple genetic alterations^1,2. Transitioning from portraying these genetic changes to understanding their phenotypic consequences by comparing human samples is, however, constrained by environmental influences, genetic diversity between patients, pervasive epistasis and the complexity of multicellular tissue structure^3,4,5. Consequently, it remains poorly understood how combinations of alterations reprogramme cells and their interactions with the tissue environment.

Genetic screens conducted in model systems have proven valuable for decoding genotype–phenotype relationships in controlled settings^6,7 However, the presently available approaches for cancer-relevant in vivo functional genomics mostly investigate the effect of singular or pairwise alterations on proliferation and tumorigenesis without considering spatial niches in which cancer cells competitively develop and grow^8,9,10,11,12. Although emerging studies are beginning to explore the impact of individual tumour alterations on their immunological microenvironments^13,14,15, gaps remain in our understanding of how genetic changes jointly rewire tumour cells and their surrounding cellular ecosystems through epistasis. Overcoming this challenge calls for experimental approaches and modelling strategies that leverage tissue-level analyses and can effectively scale to handle higher-order combinations³. For instance, testing all combinations of four alterations in standard rodent models with four mice per group requires 64 animals. This need escalates rapidly: six perturbations would demand 256 and eight perturbations would require >1,000 animals.

Here, to meet this challenge, we introduce a scalable experimental framework designed to facilitate the functional exploration of complex genotype–phenotype associations at the tissue level, which we termed charting higher-order combinations leveraging analysis of tissue to investigate genotype-to-phenotype relationships (CHOCOLAT-G2P; hereafter referred to as C-G2P). C-G2P is based on a mouse model of autochthonous tumour development, where combinations of barcoded perturbation plasmids randomly integrate into cells within their native environment (Random Unique Barcode Integration Combinatorics, RUBIX). RUBIX thus generates mosaics of genetically heterogeneous tumour clones in a single tissue. To streamline C-G2P spatially resolved in vivo functional genomics, we further developed Perturbation Barcode Capture Spatial Transcriptomics (PERTURB-CAST), a method that seamlessly integrates perturbation mapping with the standardized and commercially available 10X Visium spatial transcriptomics platform. We applied C-G2P to investigate phenotypic effects of eight combinatorial perturbations that induce liver tumours sampled from 256 possible genotypes.

Results

A framework to spatially map engineered tumour heterogeneity

Human tumours frequently present combinations of genetic alterations. With the aim to link complex genetic alterations prevalent in cancer to critical disease phenotypes within tumour ecosystems, we developed C-G2P. C-G2P allows induction and mapping of combinatorial perturbations in murine tissue and simultaneous characterization of the resulting neoplastic phenotypes on the same sample using a single spatial transcriptomics readout platform. Therefore, C-G2P merges and advances available technologies, including multiplexed perturbation in vivo functional genomics, molecular barcoding and spatial omics^5,7,9,14,16 (Fig. 1).

Previous in vivo approaches to spatially map perturbations relevant to cancer employed ex vivo-manipulated cells that were subsequently injected into animals^14,15,17. For C-G2P, we aimed to leverage an in vivo setting that more closely resembles tumour heterogeneity and tumorigenesis by direct genetic modification of cells embedded in their native tissue environment⁷. We therefore modified an autochthonous murine mosaic liver cancer model^18,19,20 to allow for the creation of coexisting genetically diverse tumours. This approach (RUBIX) relies on hydrodynamic-tail-vein (HDTV) injection of pooled molecular-barcoded plasmids and sleeping beauty transposon-based methods to stably integrate traceable higher-order combinations of genetic alterations in hepatocytes within their tissue context. Consequently, RUBIX offers the possibility to generate mosaics of genetically heterogeneous tumour clones in a single tissue (Fig. 1a(top) and Methods).

Presently available approaches to spatially map perturbations within tissue engage custom protocols and orthogonal readouts to also obtain transcriptomics-based phenotypic profiles, such as sequential antibody-based barcode detection and 10X Visium spatial transcriptomics on an additional tissue sample¹⁴. We developed PERTURB-CAST to address the constraints of existing methods and streamline the identification of perturbations and comprehensive tissue-level phenotypic information. PERTURB-CAST leverages spatial transcriptomics based on targeted transcript capture via RNA-templated ligation (RTL) probes²¹ that are commercially available with 10X Visium for formalin-fixed paraffin-embedded (FFPE) samples. To detect the introduced perturbations, PERTURB-CAST engages perturbation plasmids extended with 50-nucleotide (nt) barcodes amenable to RTL-probe capture (Fig. 1a(bottom)). Importantly, to ensure immediate compatibility with default commercial kits and circumvent modifications to the standard protocol, we redeployed 10X Visium RTL probes capturing chemosensory receptor transcripts that are not expressed in mouse liver for barcode identification. Specifically, we exploited 50-nt RTL-probe capture sequences of olfactory-, taste- and vomeronasal-receptor transcripts as molecular barcodes (Fig. 1b, Supplementary Fig. 1 and Methods). To achieve robustness, triplet barcode arrays were included in each perturbation construct to enable detection by three individual RTL probes that are pre-existing components of 10X Visium spatial transcriptomics kits (Fig. 1b, Extended Data Fig. 1, Supplementary Fig. 1 and Methods).

RUBIX generates higher-order combinatorial perturbations

Related to the genetic complexity in human liver cancer, we found that approximately 30% of tumours simultaneously present seven established cancer-driving alterations (including gain- and loss-of-function mutations as well as somatic copy number alterations such as amplifications and heterozygous losses; Fig. 2a and Supplementary Fig. 2) that can reveal varying combinatorial patterns in individual tumour samples (Fig. 2b)¹. For C-G2P proof-of-concept, we therefore modelled complex cancer genetics by not solely focusing on loss-of-function mutations and multiplex CRISPR knockouts that may induce chromosomal rearrangements and cellular toxicity due to multiple double-stranded breaks⁸. We instead concentrated on combinations of alterations associated with liver cancer, including overexpression of oncogenic drivers (Myc, mutant Ctnnb1 (mtCtnnb1), Vegfa and NICD) and silencing of tumour suppressors (Trp53, Pten and Kmt2c) with short hairpin RNA (shRNA) alongside a frequently used Renilla luciferase (shRen)-targeting control construct^13,22. We used RUBIX with a mix of eight barcoded perturbation plasmids for HTDV injection to generate a spectrum of combinatorial alterations relevant to liver cancer (Fig. 2b,c) and waited until tumours were palpable. Defined by combinations of these eight perturbations, we consequently anticipated testing 2⁸ = 256 possible cancer-driving genotypes in a single experiment (Fig. 1a). In our pilot experiments, we distributed a total of 38 redeployed barcodes amenable to RTL-probe capture across eight perturbation plasmids (including variations in position and promotors) and further included complementary barcodes (for example, peptides)^14,23 for orthogonal readouts (Extended Data Fig. 1 and Methods).

**Fig. 2: Modelling tumour heterogeneity.**

PERTURB-CAST hijacks probe-based transcript capture for barcode mapping

C-G2P liver samples were collected ten weeks after HTDV injection and processed (FFPE; Extended Data Fig. 2a). For spatial transcriptomics, six topographically separated regions of interest were selected based on histopathological assessment of haematoxylin and eosin (H&E)-stained sections (Fig. 3a and Extended Data Fig. 2a–e). Both 10X Visium and 10X CytAssist were conducted for a total of 12 samples covering these six regions, including serial sections as technical replicates (Extended Data Fig. 2f–h). A total of 324 tumour nodules were identified across the six segregated sections (Fig. 3a). Notably, spatial transcriptomics helped distinguish overlapping nodules that seemed to be single lesions from the histopathological perspective (Supplementary Fig. 3).

**Fig. 3: PERTURB-CAST spatially resolves multiplexed genetic perturbations in hundreds of coexisting cancer clones.**

Assessing the feasibility of PERTURB-CAST and our barcoding strategy, we observed that most barcode signals were readily detected in the C-G2P liver samples. Strikingly, barcode signals were spatially confined and closely tracking the areas of microscopic tumour nodules, as revealed by H&E staining (Fig. 3b). In contrast, we observed that none of the 38 redeployed barcode sequences were detected in publicly available murine liver 10X Visium datasets²⁴ (Fig. 3b(left)). Reassuringly, with few exceptions (for example, Olfr1033 and Olfr1358), chemosensory receptor transcripts that provided the repertoire for barcode redeployment (n = 1,216) were generally not detected by 10X Visium in murine livers (Extended Data Fig. 3a,b). Overall, 5/38 redeployed barcodes had insufficient signal across all samples investigated (log(1p)-transformed average expression, <0.05 counts per 1 × 10⁴; Extended Data Fig. 3). Notably, redeployed barcodes expressed using a Pol III promoter (hU6; Extended Data Fig. 3c) were detectable but we noticed a trend where detection became weaker as the barcode was positioned farther from the 5′ end (Extended Data Fig. 3c). Although detection strength of individual barcodes varied, barcode-triplets enabled us to spatially identify all eight perturbations (Fig. 3b and Extended Data Figs. 3,4).

Spatial perturbation mapping across tissue

Given the observed uncertainties related to individual barcode readouts, we used a variational Bayesian model, which accounts for multiple sources of variability (including, for example, local 10X Visium spot sensitivity) to assign perturbations to each nodule (Fig. 3c, Extended Data Fig. 5 and Methods). Notably, nodule-level predictions correlated across serial tissue sections and between 10X Visium and 10X CytAssist replicate experiments (Pearson’s correlation coefficient (r) = 0.63–0.78 depending on perturbation; Extended Data Fig. 6), demonstrating the quantitative reproducibility of the approach.

Investigation of the expression levels of the genes targeted by each perturbation provides an orthogonal readout of inferred perturbation-plasmid integration. Accordingly, across lesions, a generalized linear model (GLM; Methods) confirmed the expected trends of overexpression or silencing based on the corresponding perturbation, including increased expression of Notch1 and reduced expression of Pten (Fig. 3d) as well as Kmt2c and Trp53 (Extended Data Fig. 7a,b). Further validation of mtCtnnb1 expression was achieved by a similar analysis of glutamine synthetase (GS) immunohistochemistry (IHC), which indicates hepatic WNT–Ctnnb1 signalling activity²⁵ (Fig. 3d). Last, the plasmids targeting Trp53 and Kmt2c contained green and red fluorescent protein barcodes, respectively, providing additional IHC-based validation (Extended Data Fig. 7c–e).

Thus, the probe-based barcode capture of PERTURB-CAST spatially maps combinatorial perturbations within hundreds of coexisting tumours generated by RUBIX and provides a foundation to comprehensively chart tumour genotypes (interactive maps at https://chocolat-g2p.dkfz.de/).

Comparative analyses in hundreds of coexisting cancer clones

Across the entirety of the 324 identified nodules, the Bayesian model calculates the probabilities for all 2⁸ = 256 possible genotypes defined by the combinations of eight perturbations, thereby converting spatial barcode signals into genotypically defined clonal maps (Fig. 4a where individual perturbation probabilities for one nodule are highlighted as an example and Extended Data Fig. 8).

**Fig. 4: C-G2P enables comparative genotype analyses and disentangles context-dependent genetic interactions.**

We observed that tumour clones established in our initial C-G2P experiments typically exhibited combinatorial alterations with quintets being the most prevalent (approximately 30%; Fig. 4b,c). The absence of nodules with low integration numbers and the overall tendency towards multiple perturbations corroborates expected and previously described genetic interactions of oncogenes and tumour suppressors, defined simply by pairwise cooperation inferred from individual experiments^19,20. Furthermore, alterations of well-recognized oncogenes, for example, Myc (82% of all nodules) and mtCtnnb1 (80%) occurred most frequently, indicating strong clonal selection, whereas VEGFA (53%) and shRen (46%) were less frequent, possibly reflecting low tumorigenic potential (Fig. 4d).

The frequency at which perturbations are observed across nodules is dependent on the rate of successful integrations and the neoplastic potential of the combinatorial perturbation. Assuming a fixed integration rate for each perturbation allows modelling of an expected distribution of combinatorial events and assessment of whether specific observed combinations are enriched, suggesting higher tumorigenic potential (orange), or depleted and thus indicating disadvantageous effects (blue) independent of technical influences (Fig. 4b(bottom) and Methods). Notably, among genotypes with fewer combinations, the triplet comprising Myc, mtCtnnb1 and NICD emerged as frequent (n = 5, p(O > E) = 0.87; solid line and arrow in Fig. 4a and b(top), respectively), which suggests a strong association of this specific compound genotype with tumorigenesis. Interestingly, although septets seemed to be generally prevalent, the specific septet devoid of VEGFA (dashed line and arrow in Fig. 4a and b(top), respectively) demonstrated enrichment (n = 13, p(O > E) = 0.90) comparable to the complete octet (n = 12, p(O > E) = 0.81), whereas septets lacking mtCtnnb1 or Myc, for example, were less enriched. This in turn suggested a diminished cancer-driving effect of VEGFA in the setting of the combinatorial alterations tested.

Cancer-driving co-dependencies and potential context dependencies

To pinpoint which perturbations contributed to the observed patterns of enriched and depleted genotypes, we conducted co-occurrence odds ratio (OR) analysis. This analysis measures second-order epistatic interactions, which quantifies the deviations from purely additive effects in a commonly used multiplicative model^6,26 (Fig. 4e,f and Methods). Our results revealed co-dependency patterns for Myc, mtCtnnb1 and NICD as well as shTrp53 and shPten, aligning with previous observations^18,19,20. The combination of Myc and NICD exhibited the most pronounced effect (OR = 1.64; P = 0.067). In contrast, we observed a tendency towards mutual exclusivity between VEGFA and NICD (OR = 0.68; P = 0.036) and between VEGFA and mtCtnnb1 (OR = 0.67; P = 0.064; Fig. 4e,f).

Together, these observations indicate a context-dependent oncogenic effect of VEGFA.

Phenotypic landscapes of engineered tumour heterogeneity

To elucidate spatial phenotypes and enable subsequent genotype-to-phenotype analyses (Fig. 5), we leveraged tissue-wide transcriptional signatures and defined sets of transcripts that characterized prevalent cell states (Supplementary Figs. 4 and 5 and Methods). To finally map the complexity of tumour ecosystems, we visualized phenotype-associated transcriptional signatures within their spatial context (Fig. 5a and Supplementary Figs. 6–18). Furthermore, to highlight associations for nodule-intrinsic phenotypes as well as those related to the tumour microenvironment (TME), we used transcript-resolved heatmap presentations (Fig. 5b–d). Thereby, C-G2P allowed us to chart the heterogeneous phenotypic landscape of hundreds of coexisting genotypically defined tumours and their surrounding tissue environment (interactive maps at: https://chocolat-g2p.dkfz.de/).

**Fig. 5: C-G2P maps tumour ecosystems comprising hundreds of coexisting cancer clones.**

Stratification of coexisting liver tumour subtypes

Our approach readily distinguished prominent subtypes of liver tumours (Fig. 5a(top)). First, we observed nodules with cholangiocyte-like transcriptional signatures (for example, Krt19⁺Cldn7⁺) indicative of cholangiocarcinoma (CCA)^27,28 (Fig. 5a,b and Supplementary Fig. 6). Microscopy inspection indeed classified these tumour nodules as CCAs exhibiting a glandular growth pattern and stroma deposition as well as CK19-protein expression by IHC (Supplementary Fig. 7). Cholangiocarcinoma is the second-most common type of liver cancer following hepatocellular carcinoma (HCC) and both tumour types can develop from hepatocytes in the HDTV injection-based mouse model^27,28. Interestingly, C-G2P pinpointed, among others, expression of solute carrier family 15 member 2 (Slc15a2) for which genomic variants have been linked to sorafenib-therapy response²⁹ as well as pancreatic glycoprotein 2 (Gp2) as being associated with CCA (Fig. 5b and Supplementary Fig. 6). The latter observation harmonizes with earlier research suggesting that anti-GP2 IgA autoantibodies enable early CCA detection in subsets of human patients³⁰, hence indicating that our C-G2P approach captures key elements of CCA biology that have so far not been observed in animal models.

Moreover, given the spatial resolution of C-G2P, we could immediately relate the prominent second and third cluster of tumour nodules to metabolic liver zonation. Spatial division of metabolic functions is not only central to liver-tissue organization under physiological conditions^31,32 but has been proposed to enable molecular classification of human HCC^33,34,35. In alignment with this zonation-based molecular classification, C-G2P enabled us to stratify nodules either as portal-like (for example, Sds⁺Sdsl⁺) or central-like (for example, Cyp2e1⁺Oat⁺)^31,32, the latter being the most abundant tumour class observed (Fig. 5a,b and Supplementary Figs. 8 and 9). Interestingly, recent findings from zonation fate-mapping animal models suggest liver cancer prevention strategies that leverage central-zonation-dependent mechanisms, particularly targeting Gstm3, which we also identified as a central-like tumour marker (Fig. 5b)³⁶.

Last, focusing on the tumours that could not readily be assigned to the aforementioned subtypes, we identified a fraction of nodules that revealed enrichment of hepcidin antimicrobial peptide (Hamp), Hamp2 and uridine phosphorylase 2 (Upp2) expression (Hamp2+Upp2⁺; Fig. 5a and Supplementary Fig. 10). Hamp and its paralogue Hamp2 have both been associated with midlobular zonation³², a feature of liver structure important for regeneration³⁷. Upp2, on the other hand, is involved in pyrimidine salvage, which fuels glycolysis and enables growth of cancer cells under nutrient-limited conditions^38,39.

We further identified subgroups of nodules sharing cholangiocytic as well as portal-like features (Fig. 5b), an observation in agreement with a proposed hybrid periportal hepatocyte cell type⁴⁰. Similarly, subsets of nodules from the major classes, with the exception of central-like nodules, shared striking enrichment of numerous histone-associated transcripts (Fig. 5b and Supplementary Fig. 11). Upregulation of genes encoding histone proteins is described as the most prominent gene regulatory programme at the G1–S phase transition in human pluripotent cells⁴¹.

Tumour–stroma and tumour–immune cell connections

Next, by focusing on cellular ecosystems of the liver TME (Fig. 5a(bottom) and Fig. 5d), we identified prominent fibroblast-associated transcriptional signatures (for example, Col1a1⁺Col3a1⁺)^42,43 at the tumour–stroma border. We further observed regionally segregated expression patterns associated with haematopoietic/immune cell clusters (Fig. 5a). These included signatures likely to be associated with erythroblasts (for example, Hbb-bt⁺Slc4a1⁺)^44,45, platelets (for example, Pf4⁺Itga2⁺)⁴⁶, mast cells (for example, Cpa3⁺Cma⁺)^47,48, B cells (Jchain⁺Igkc⁺)⁴⁹ and neutrophil subpopulations (for example, Elane⁺Mpo⁺ and Ngp⁺Camp⁺)^45,50,51. Signatures associated with Kupffer cells/macrophages (for example, Marco⁺Clec4f⁺ and Csf1r⁺C1qa⁺)^3,24 were primarily detected within the non-tumour compartment (Supplementary Figs. 12–18).

Our approach immediately revealed connections between tumour-intrinsic cell states and the microenvironment, such as a notable link between CCA and fibroblast-like signatures (Fig. 5c). This observation aligns with human data indicating that cancer-associated fibroblasts are the major cellular component of CCA-associated desmoplastic stroma⁴³. Our approach indeed grouped fibroblast-like signatures alongside growth arrest-specific 6 (Gas6) and thrombospondin 1 (Thbs1), both of which were previously identified as marker transcripts for a mechanoresponsive cancer-associated fibroblast subpopulation⁴² (Fig. 5d and Supplementary Fig. 12). Our results further pointed towards additional associations such as a link between CCA and macrophages (for example, Csf1⁺C1q⁺) as well as a connection between enriched erythroblast (Hbb-b⁺Slc4a1⁺) occurrence and the histone-associated subgroup of nodules (Fig. 5c).

Complex genotype–phenotype relationships

Using spatial maps that combine phenotypic and genotypic data from the same tissue sections enables detailed investigation of phenotype–genotype relationships (Figs. 4a and 5a–d). We therefore assigned binary phenotype labels to nodules (Extended Data Fig. 9 and Methods) and calculated the OR values to assess the connection of each perturbation to specific tumour-intrinsic and microenvironmental phenotypic groups (Fig. 6a). Remarkably, in the setting of the combinatorial perturbations tested, our findings indicated that CCA reveal a strong positive association with VEGFA, exceeding any other observed linkage, as well as a strong negative association with mtCtnnb1 (Fig. 6a). Furthermore, consistent with the central role of WNT signalling in liver zonation⁵², portal-like tumours revealed negative associations, whereas central-like nodules revealed positive associations with mtCtnnb1 (Fig. 6a). Notably, these genotype–phenotype observations align well with the aforementioned zonation-based classification of human HCCs and single-cell RNA-sequencing (scRNA-seq) data from human liver cancer, which revealed that the central-like HCC subtype is associated with Ctnnb1 mutations^33,34,35, hence indicating that our C-G2P approach mirrors features of human HCC biology.

**Fig. 6: C-G2P decodes relationships between complex genotypes and tumour-intrinsic and microenvironmental phenotypes.**

Focussing on relationships between tumour genotypes and TME phenotypes, we observed strong positive associations for VEGFA with fibroblast signatures, alongside a negative association with mtCtnnb1 (Fig. 6a), largely resembling the patterns observed for CCA and being in line with their prominent spatial association (Fig. 5c). A similarly positive association for VEGFA was observed for B cell-like signatures alongside negative association for Myc and mtCtnnb1, and positive associations for shTrp53 and shPten (Fig. 6a). Reflective of our findings, immune cell infiltration, as evaluated by CD45-positive IHC, was reported in a compound Trp53 and Pten-knockout HDTV injection model of liver cancer²⁰, whereas immune cell exclusion has been observed in a corresponding Myc–mtCtnnb1 model¹⁹.

To broaden our analysis beyond binary phenotypes, we leveraged spot-level continuous expression of phenotype-associated marker transcripts. We therefore calculated associations for each perturbation using the aforementioned GLM analyses (Fig. 6b and Extended Data Fig. 10a,b). Despite its limited sensitivity to identify associations for transcripts that reveal sparse spatial expression (Methods), this analysis supported the identified associations for VEGFA and mtCtnnb1 for cholangiocyte-associated transcripts such as Krt19 and Cldn7 (Fig. 6b).

Similarly, GLM analyses substantiated the observed inverse mtCtnnb1 associations for portal-like markers (that is, negative association for Sds and Sdsl) versus central-like markers (that is, positive associations for Oat and Gulo; Extended Data Fig. 10a). Notably, GLM analyses uncovered additional transcript-specific contributions of perturbations not readily apparent using the nodule phenotype-binarization approach (Fig. 6a). For example, we observed a marked relation of the cholangiocyte-associated transcripts Krt19 and Gp2 with shKmt2c perturbation (Fig. 6b and Extended Data Fig. 10a,b).

Given that multiple transcripts associated with predefined phenotype signatures shared similar ‘GLM-patterns’ (Extended Data Fig. 10a,b), we finally interrogated perturbation–phenotype associations on a transcriptome-wide scale (Methods). Focusing on 1,283 genes that showed associations with perturbations, we observed clusters of transcripts that correspond to specific cell states (Extended Data Fig. 10c). For example, global-GLM analysis grouped together fibroblast-associated transcripts such as Col1a1, Col1a2, Gas6 and Thbs1 or transcripts related to the aforementioned histone-enriched phenotype. Similarly, GLM analysis aggregated Krt19, Epcam, Gp2 and Krt7, all of which defined the cholangiocytic phenotype that was linked to positive associations with VEGFA and negative associations with mtCtnnb1 (Extended Data Fig. 10c and Fig. 6a,b).

Epistatic regulation of CCA development

Our C-G2P experiments suggested that VEGFA and mtCTNNB1 mediate opposing epistasis effects relevant to CCA development within the genetically heterogenous tumour ecosystems analysed, as nodules of this tumour class were consistently negative for mtCTNNB1 and positive for VEGFA (Fig. 7a). To substantiate this finding, we took advantage of the flexibility of our RUBIX mouse model and adjusted combinatorial complexity of genetic perturbations to directly assess the individual contributions to tumour development and phenotype. Specifically, we employed a ‘leave-one-out’ strategy, where we compared animal cohorts that received a full mix of all eight perturbation plasmids to cohorts where either VEGFA or mtCTNNB1 was deliberately omitted (Fig. 7b). In mice that received all eight perturbation plasmids, all animals developed tumours within eight weeks and histopathology revealed multiple tumour nodules, a fraction of which identified as CCA (HCC, 42 ± 17.7 versus CCA, 5 ± 1.96 per section). Immunohistochemistry for CK19 confirmed the CCA classification (Fig. 7c–e), matching our earlier C-G2P results (Fig. 5). In the cohort missing VEGFA, tumours developed faster and all animals presented tumours within five weeks. The majority of nodules identified as HCC (HCC, 63 ± 13.6 versus CCA, 1 ± 0.756 per section). Using IHC, we observed that tumour nodules were generally negative for CK19, confirming the absence of CCA, although CK19 was still detectable in normal bile ducts (Fig. 7c–e). In contrast, the absence of mtCTNNB1 delayed tumour development, extending the time for all animals to present tumours to 11 weeks. Observed tumours were predominantly CK19-positive CCAs (HCC, 1 ± 1.51 versus CCA, 12 ± 4.71 per section; Fig. 7c–e). Together, these results confirm predictions of C-G2P-based findings revealing that whereas VEGFA expression essentially contributes to CCA development in the setting of combinatorial alterations investigated, mtCTNNB1 elicits a dominant epistasis-masking effect on this particular liver tumour subclass.

**Fig. 7: VEGFA and mtCTNNB1 confer epistasis control of CCA development within heterogeneous tumour ecosystems.**

Discussion

Here we introduced PERTURB-CAST, an approach that seamlessly integrates perturbation mapping for in vivo functional genomics with spatial transcriptomics interrogation by redeploying RTL probes from commercial technology for molecular barcode identification. Combined with RUBIX, which allows induction of hundreds of tumours, each with distinct combinations of alterations, in a single tissue, our CHOCOLAT-G2P framework offers the capability to characterize tumour gene expression and cellular microenvironments, and helps address the long-standing question of how multiple genetic changes interact to shape disease phenotypes within cellular ecosystems.

By applying C-G2P in an autochthonous mouse model of liver cancer, we explored a wide range of cancer-driving combinatorial genotypes sampled from nearly all of the 2⁸ combinations of perturbations present in hundreds of coexisting tumours. The integration of PERTURB-CAST for spatial transcriptomics enabled simultaneous mapping of the genotype of each nodule alongside tumour-intrinsic and microenvironment-related phenotypes on the same tissue sample. PERTURB-CAST eliminated the need for complementary readouts as well as the requirement for analyses on serial tissue samples, thereby preserving spatial relationships and providing the basis for detailed genotype–phenotype analyses.

Interrogating 324 liver tumour nodules from a single C-G2P experiment revealed mutual exclusivity between mtCtnnb1 and VEGFA, indicating epistatic fitness effects of these two alterations. C-G2P revealed that their exclusivity was further underscored by phenotypic divergence. Specifically, VEGFA induced a cholangiocytic histology and gene expression profile. VEGFA-perturbed nodules also exhibited a greater abundance of cancer-associated fibroblasts compared with nodules with mtCtnnb1, indicating that genetic alterations also shape, and possibly co-opt, their TME. In contrast, mtCtnnb1, which we identified as a crucial contributor to overall liver tumour occurrence, masked the emergence of the CCA subtype, thus exemplifying Bateson’s classical definition of epistasis²⁶.

C-G2P can be applied and extended in a number of ways. First, it is straightforward to adjust combinatorial complexity and exchange the alterations to additional cancer drivers¹ and other perturbations, potentially within a compressed screening framework¹⁶. C-G2P may also be conducted in different mouse strains or growth conditions to model interactions and selective forces between tumour genomes and host genetics, immunocompetence, environmental exposures and therapeutic interventions^7,53. Second, we envision the applicability of PERTURB-CAST and C-G2P beyond liver cancer. Currently available autochthonous animal disease models that similarly build on stable integration of perturbations include diverse tumour types such as lung, pancreas, stomach and soft tissue cancers^7,54,55. Moreover, the perturbation plasmids we employed here (Extended Data Fig. 1 and Methods) are expected to be compatible with scRNA-seq readouts and imaging-based screening platforms^{14,16,17,56,57,58} that could in the future enable complementary insights.

In summary, C-G2P provides a multiplexed approach for higher-order combinatorial cancer screens in an individual mouse. Its design—built on PERTURB-CAST, which uses off-the-shelf spatial transcriptomics protocols—facilitates comprehensive readouts of tumour genotypes and spatial phenotypes from the same tissue sample.

As C-G2P may be extended to other disease models, perturbations and (spatial) omics technologies, we envisage a broad range of applications to decode the relationships between complex genotypes and phenotypes⁵⁹ within the holistic context of tissue and the entire organism.

There are some limitations of the study. First, RUBIX currently generates tumour heterogeneity by establishing random combinatorial alterations simultaneously. Although this scalable approach could aid to initially explore a vastly unknown epistatic interaction space, this contrasts tumour development in humans where, in most cases, cells sequentially acquire alterations¹. To address this, experimental strategies that enable stepwise introduction of genetic alterations require further exploration. Inducible perturbation systems, previously used in the liver cancer mouse model employed here, may provide a potential solution. Alternatively, we envision that serial injections of differentially barcoded plasmid pools could capture spatially resolved in vivo functional genomics data across temporal alteration trajectories. Although the random simultaneous introduction of alterations and establishment of hundreds of coexisting tumour nodules could provide a valuable opportunity for studying interclonal interactions, the complexity of our current model may as such not fully recapitulate the ancestral lineage, clonal evolution and genetic epistasis seen in human tumours. To more systematically investigate genetic interactions, we speculate that the multiplexing capabilities offered by MultiMir combinatorial RNA interference⁶⁰ as well as CRISPR/Cas12a⁶¹ warrant further investigation. For multiplex CRISPR perturbations, it may however, be necessary to employ modifications that avoid double-stranded breaks, such as CRISPR interference, to minimize the risk of unwanted chromosomal rearrangements and other detrimental effects¹⁶.

Second, PERTURB-CAST hinges on the availability of robust RNA-detection probes, a requirement that is not always satisfied, as evidenced by the failure to detect 5/38 barcodes tested in this study. We anticipate that prospective massively parallel assays⁶¹ based on C-G2P could be leveraged to select reliable probes.

Third, the moderate resolution of the 10X Visium platform we employed does not allow for single-cell analyses. Additional computational approaches and next-generation spatial transcriptomics platforms could address this shortcoming^62,63,64,65. PERTURB-CAST could indeed be integrated with complementary probe-based single-cell spatial omics technologies such as the recently introduced CosMx WTx assay, which uses two in situ hybridization probes per transcript with 35–50 nt RNA-targeting domains⁶⁶. Finally, PERTURB-CAST may offer immediate implementation into high-resolution 10X Visium HD, which operates on the same RTL-probe-based transcript capture technology⁶⁷.

Methods

Animal experiments

Group size was determined on the basis of our experience in previous experiments^68,69. For HDTV injections, eight-week-old female C57Bl/6 animals were purchased from Envigo. The mice were injected (into the lateral tail vein in 5–7 s) with 5 μg DNA of each of a total of eight perturbation plasmids (40 µg total DNA) mixed with 20 µg CMV-SB13 transposase (1:2 ratio) prepared in sterile 0.9% sodium chloride solution in a total volume corresponding to 10% of the body weight, as described before. Two animals were used. We labelled this approach RUBIX as the perturbation plasmids are equipped with molecular barcodes (Extended Data Fig. 1) and the plasmid mixtures injected allow for all possible combinations to become integrated in the genome of hepatocytes (Fig. 1a). All animals were monitored twice weekly and animal experiments were performed in compliance with all relevant ethical regulations determined in the animal permit. After tumours were palpable (10 weeks), the animals were euthanized and their livers were harvested (Extended Data Fig. 2). As a control group, two animals were injected with 40 µg pT3-EF1-shRen and 20 µg CMV-SB13 transposase (1:2 ratio) prepared in sterile 0.9% sodium chloride. For fixation, livers were incubated in 4% paraformaldehyde for 48 h. The sample processing procedure is illustrated in Extended Data Fig. 4. For the leave-one-out experiments (Fig. 7), HDTV injections were performed essentially as described before. For omission of either VEGFA or mtCTNNB1 perturbation from the plasmid pools, the respective plasmids were replaced with a control plasmid. Housing conditions for the mice were: 12 h light–12 h dark cycle, an ambient temperature of 20–24 °C and relative humidity of 45–65%. All animal experiments were approved by the regional board Karlsruhe, Germany.

10X Visium for FFPE spatial transcriptomics

Spatial transcriptomics were performed using the manual 10X Visium workflow for samples embedded in paraffin blocks or the 10X Visium CytAssist workflow for samples already placed on glass slides and stained with H&E (Extended Data Fig. 2). Both workflows were carried out according to the manufacturer’s protocol (CytAssist, CG000495, RevC; manual Visium, CG000407, RevD). Briefly, slices of approximately 5 µm were cut from FFPE blocks using a microtome and floated onto a water bath at 42 °C until all wrinkles were resolved. For the manual Visium workflow the slice was then placed inside the capture frame of the spatial transcriptomics slide (M.R.). Slices used for the CytAssist workflow were placed on frosted glass slides. Deparaffinization and staining of the slides was similar between both workflows. After drying the slide, paraffin was removed through incubation at 60 °C for 2 h and a subsequent incubation in xylol. Rehydration was done by sequential washes with decreasing ethanol concentrations. After rehydration, the tissue was stained with H&E and imaged using a Leica Aperio AT2 microscope at ×40 magnification. Following imaging, the slides were destained by incubation in 0.1 N HCl and formalin crosslinks were removed by incubation with TE buffer pH 9.0 for 1 h at 70 °C (manual workflow) or decrosslinking buffer for 1 h at 95 °C (CytAssist workflow.) The tissue was then permeabilized and incubated with RTL probes at 50 °C for approximately 20 h. Free probes were washed away and the bound probes were ligated, followed by washing steps to remove unligated probes. For the manual workflow, the probes were released by treating the slices with RNase and a permeabilization enzyme. For the CytAssist workflow, the slices were stained with a diluted eosin solution and placed in the CytAssist together with the Visium spatial transcriptomics slides and incubated for 30 min at 37 °C with RNase and permeabilization enzyme. For both protocols, the spatial barcode was added to the probes by extending them and the probes were released using a 0.08 M potassium hydroxide solution. For the CytAssist workflow, a pre-amplification PCR consisting of eight cycles was performed. After clean-up with 1.2× SPRIselect beads, 25% of the product was used as input for the index PCR. For both protocols, a quantitative PCR was used to select the number of cycles for the index PCR. To reduce PCR duplicates and avoid overamplification, cycle number at a C_q value of 10% was used for the index PCR. The PCR product was purified using 0.85× SPRIselect beads.

For samples already stained and mounted on a slide, the slides were first imaged and then incubated in xylol to remove the coverslip. Sample rehydration was done by sequential washes with decreasing ethanol concentrations. The slides were destained and decrosslinking was performed by incubating with a decrosslinking buffer at 95 °C for 1 h. After decrosslinking, the samples were incubated with probes for 20 h at 50 °C. Excess probes were washed away and the probes were ligated. Thereafter, the unligated probes were washed away. The samples were stained again with eosin and placed in the 10X CytAssist together with the spatial transcriptomics slides. A mixture of RNase and permeabilization enzyme was added to the spatial transcriptomics slides and the 10X CytAssist was started. After incubation, the spatial transcriptomics slides were removed and the enzymes were washed off. The spatial barcodes were attached to the probes with an extension enzyme. Probes were released using 0.08 M potassium hydroxide solution. The probes were then amplified through eight PCR cycles; 25% of the purified PCR products were used as input for the index PCR. The cycle number of the index PCR was determined using the cycle number at a C_q value of 10%.

For all samples, the final sample concentration was determined using Agilent Tapestation 4150 with D1000 HS tapes. Sequencing for both protocols was performed on an Illumina NovaSeq6000 system. Four samples were pooled on one SP flow cell with 100 cycles to aim for a read count of 250 × 10⁶ reads per sample. The FASTQ files and the alignment were done using spaceranger 2.0.1. A total of 12 spatial transcriptomics datasets were generated (Extended Data Fig. 2). Note that the utility of sample ML-II_B_2Cyt is constrained by tissue detachment of the sample during the processing for 10X Visium CytAssist.

The 10X Visium for FFPE protocol engages RTL probes that capture a 50-nt sequence specific to endogenous transcripts (note that we used Visium mouse transcriptome probe set v1). We leveraged this strategy to likewise capture molecular barcodes via RTL probes (Fig. 1). We hence labelled this approach PERTURB-CAST.

Economized molecular barcode selection

To enable PERTURB-CAST, we aimed to avoid additional expenses and protocol modifications by redeploying RTL probes from commercially available 10X Visium reagents (originally designed to detect endogenous transcripts) as barcode identification reagents, provided that the selected transcripts are not expressed in mouse liver. We named this strategy redeploy probes for barcode capture (REDPRO-BC; Fig. 1 and Supplementary Fig. 1). To this end, we initially analysed a publicly available bulk RNA-seq dataset including a total of 128 murine liver samples (GSE137385)⁷⁰ to identify transcripts that were generally not detected (fragments per kilobase of transcript per million mapped reads = 0 over all samples). Note that this approach can be error-prone due to the initial source data. Consequently, we went on to validate non-expression of selected transcripts (olfactory, vomeronasal and taste receptors) in additional datasets (including bulk RNA-seq from GSE148379)¹⁹, information provided in MGI GXD⁷¹ as well as 10X Visium data²⁴. The endogenous transcripts associated with the REDPRO-BCs used in this study are illustrated in Extended Data Fig. 3 and the respective nucleotide sequences for 10X Visium RTL-probe capture barcodes (reverse complement to RTL-probe sequence provided by 10X Genomics) are listed under the section ‘Molecular cloning’. Note that we used Visium mouse transcriptome probe set v1. Visium mouse transcriptome probe set v2 is not compatible with the barcodes employed in this study.

Molecular cloning

The transposon plasmids used in this study including overexpression constructs for NICD, mutant human CTNNB1 (T41A) and human MYC and potent shRNA constructs targeting Trp53, Pten, Kmt2c and Renilla luciferase were previously described and validated in animal experiments^19,68,72. For spatial transcriptomics, note that NICD overexpression can be investigated via Notch1 expression given that the 10X Visium RTL probe identifies the exogenous transcript introduced. However, endogenous Notch1 is similarly identified. VEGFA overexpression plasmid was cloned by insertion of a codon-optimized gene fragment (gBlock, IDT) based on Vegfa NCBI reference sequence NP_033531.3 by replacing hMYC from a previously validated expression plasmid⁶⁸ using NEBuilder HiFi-DNA assembly according to the manufacturer’s protocol (NEB). All plasmids were individually modified to express molecular barcodes. Specifically, fluorescent protein-based peptide barcodes (mKate2, mOrange2 and mWasabi) were ordered as codon-optimized gene fragments (gBlock, IDT) and cloned into previously validated shRNA expression plasmids to replace GFP⁶⁸ using NEBuilder HiFi-DNA assembly according to the manufacturer’s protocol. Small-peptide barcodes (for example, AU1, AU5 and so on), as described previously¹⁴, were ordered as oligonucleotides (Sigma), annealed and cloned using NEBuilder HiFi-DNA assembly according to the manufacturer’s protocol. Long RNA barcodes (stretches of at least 650 nt derived from the combination of multiple oligonucleotide–miner probe sequences designed against Arabidopsis thaliana Chr1 (ref. ⁷³) were ordered as gene fragments (gBlock, IDT) and cloned using NEBuilder HiFi-DNA assembly according to the manufacturer’s protocol. REDPRO-BC triplet arrays were ordered as gene fragments (gBlock, IDT) and cloned using NEBuilder HiFi-DNA assembly according to the manufacturer’s protocol. Briefly, each REDPRO-BC triplet array incorporates one 50 nt sequence derived from olfactory receptors, one 50 nt sequence derived from taste receptors and one 50 nt sequence derived from vomeronasal receptors (based on 10X Genomics RTL-probe sequences against murine transcripts; note that we used Visium mouse transcriptome probe set v1; sequences below), separated and flanked by spacer sequences of approximately 20 nt to avoid potential steric hindrance during hybridization. The spacer sequences used were derived from T7 and T3 promoters (described in ref. ⁵⁶) and/or AsCas12a-DR sequences (described in ref. ⁶¹), and/or the 10X Capture sequences cs1 and cs2 (described in ref. ⁵⁷), and as such provide additional functionality, which was not tested in this study. Single 50-nt REDPRO-BCs were ordered as oligonucleotides (Sigma), annealed and cloned using NEBuilder HiFi-DNA assembly according to the manufacturer’s protocol. Note that the REDPRO-BC length should enable straightforward en masse cloning engaging commercially available oligonucleotide pools (such as in ref. ⁶¹), which was not tested in this study. Peptide barcodes were integrated in frame with the respective coding regions. RNA barcodes were integrated in the 3′ untranslated region of coding regions expressed under the control of a polymerase II promoter (EF1), unless otherwise specified. Subsets of perturbation plasmids were equipped with REDPRO-BC arrays either 5′ and 3′ of the shRNA expression cassette (mir-E-based) to account for mir-E processing⁷⁴, or with additional REDPRO-BC arrays driven by a polymerase III promoter (hU6) in reverse orientation to the EF1-driven transcript. Extended Data Figs. 1 and 3 provide simplified illustrations of the plasmid design and barcode position. Respective FASTA sequences of plasmids are available on request. Plasmids were validated by restriction digest and Sanger sequencing (Microsynth).

Selected REDPRO-BC barcode sequences based on Visium mouse transcriptome probe set v1 were as follows.

Myc

Olfr103, TGGGAGTGAGAGACATACAAGAACCACAGCCCTTTCTCTTTGCTATTTTC; Tas2r102, AACACAAGTGTGAATACCATGAGCAATGACCTTGCAATGTGGACCGAGCT; Vmn1r1, TAAAAGGCAGTGTCAGTACCTTCACAACACCAGCATTTCCCGCAAAGCAT; Olfr1018, CAGTTCCATGGTTATCAATGTTCTCACCTTGAGTTTGCCCTACTGTGGAC; Tas2r118, TTATTGGCACTGTGTTTGATAAGAAATCTTGGTTCTGGGTCTGCGAAGCT; Vmn1r174, ACTTCAACCAGAGGCCAGAGCAGCAAACACAATTCTCATGCTGATGATCA; Olfr1, TGGCCAGCATCTTTCTTGTCCTTCCATTTGCACTCATTACCATGTCCTAT.

mtCtnnb1

Olfr1000, GGCACAGTAGGTATGTTCACTGGTCTGATAATTCTGGGGTCCTATGTATG; Tas2r103, TGTCACTAATCACAGGGTTCTTGGTATCATTATTGGACCCAGCTTTATTG; Vmn1r178, GTCTCTTCATGAGTCATTTCAGTAAAGTTTTTGCTGCAGGATTCCCCACT; Olfr1019, TGCTTGGTCCTAATGCTGGGCTCTTACTTCGCTGGCCTAGTGAGTTTAGT; Tas2r119, GATATCCAGGTTGGTGCCATGGCTGATCCTGGCATCTGTGGTCTATGTAA; Vmn1r175, AGTACAAACATGTGCTCCACCTGCTTTCTGAGCACTTATCAGCTTGTCAC.

NICD

Olfr1006, AGGGAACATGTTGCTGGTTGTTTTAATCCGAATTGATTCTAGACTGCATA; Tas2r105, GACCTCGGAGATGTACTGGGAGAAAAGGCAATTCACTATTAACTACGTTT; Vmn1r139, AAGCATTGGCAAGTCACAGGCAAAGAGTGACACAGAGACGTTCCTCAATT.

VEGFA

Olfr1002, AGGCCTTATAAGCACTGTGGTCCATACTACTTCTGCATTTATTCTTCCAT; Tas2r104, TAACGTGGCTAGCTTCCTTTCCGCTAGCTGTGAAGGTCATTAAAGATGTT; Vmn1r12, ACTACATTGTCAGGAGCTTGATTTTAACTGTGACAACTTCCAGGGATATG.

shPTEN

Olfr1013, GTACACATTGACTTTGATGGGAAATAGCTCCCTCATTATGTTAATCTGCA; Tas2r110, ACTAGTGAATATCATGGACTGGACCAAGAGAAGAAGCATTTCATCAGCGG; Vmn1r170, TGATTCTCCTGAACAGACACCACCACAGACTGCAGCATATTCAATCCACA; Olfr1015, TGTCTATGTGAAAATCCTTTCCAGTATGGTGGGCTTCACTGTCCTCTCAA; Tas2r114, TGTAATTTGTCTGTTAATCCCAGAAAGCAACTTGTTATTCATGTTTGGTT; Vmn1r172, GGAAGTAAATGCCCAGAGAGTCTTCAAAGGAAGACAGTCATAGCTGTTTT.

shREN

Olfr1008, CCAGGCTCTGCTATTCACCAGTAAAATTTTCACATTAACTTTCTGTGGCT; Tas2r106, AAGGCACTGAAGCAATTAAAATGCCATAAGAAAGACAAGGACGTCAGAGT; Vmn1r157, CAGATCCTCTTGCTTTGCCATTTTGAGGTTGGGACCGTGGCCAATGTCTT.

shTRP53

Olfr1009, CCAGAGACTCTGCATACAGCTGGTGATCGGACCCTATGCTGTTGGCTTTT; Tas2r107, GCTCTCTAAGATCGGTTTCATTCTCATTGGCTTGGCGATTTCCAGAATTG; Vmn1r167, GTTTCAGTATAGGCATGCGCATCTTATCATTTGCCCATGATGGAGTGTTC; Olfr1014, TTGCTGTGTATGCATTAACTGTGTTAGGAAACAGCACCCTCATTGTGTTG; Tas2r113, GATCAATCATTGTAACTTTTGGCTTACTGCAAACTTGAGCATCCTTTATT; Vmn1r171, AACAGCACTGCCCTCATGATCACTATTCCGTTGACCAATGAAGTTGTCTC; Olfr107, TTACTGCTTTCTTGCTCAGACACTCACCTCAGTGAGGGCCTGATGATGGC.

shKMT2C

Olfr1012, ATCTACTCTCGGCCAAGTTCCAGTTATTCCTTGGAAAGGGATAAAATGGT; Tas2r109, TTCTAGAATTTTCCTGCTCTGGTTCATGCTAGTAGGTTTTCCAATTAGCT; Vmn1r169, GGTACCTGGGGTAGGGTGATGCTCCATGGAAGAGCCCCCAAATTTGTGAG.

Histopathology

After fixation, representative specimens of the liver were routinely dehydrated, embedded in paraffin and cut into 4-μm-thick sections. The tissue sections were stained with H&E according to standard protocols. Slides were scanned using a SCN400 slide scanner (Leica Biosystems) at ×20 magnification.

Nodule annotation was initially performed by experienced pathologists (D.F.T. and H.W.) based on H&E-stained sections using the quPath software and the 10X Loupe Browser software. Nodule annotation was further refined based on specific transcript expression (example provided in Supplementary Fig. 3) using the 10X Loupe Browser software (H.W. and M.B.).

IHC

After heat-induced antigen retrieval at pH 6 or pH 9, FFPE tissue sections were incubated overnight with the primary antibody and blocked with hydrogen peroxide if necessary. Depending on the primary antibody, an anti-mouse or anti-rabbit secondary antibody conjugated to horseradish peroxidase (HRP) and alkaline phosphatase (AP), respectively, was applied (PolyviewPlus, ENZO Life Sciences GmbH). The signal was visualized using either 3,3′-diaminobenzidine (Dako liquid DAB+ substrate, Agilent Technologies, Inc.) or AP (Permanent AP red, Zytomed Systems) as a chromogen. Details are given in Table 1.

Table 1 Details of antibodies used in IHC

Full size table

Slides were scanned using a SCN400 slide scanner (Leica Biosystems) at ×20 magnification. The individual histochemical GFP, RFP and GS staining was evaluated using the quPath software as: high, very intense uniform staining; moderate, moderate intensity or intense non-uniform staining; and low, low intensity and non-uniform staining. Mapping of barcode signals to respective IHC data was performed by manual assessment of marker staining on IHC images using the quPath software (H.W. and M.B.). Next, corresponding tumour nodules were selected, categorized and stratified using the 10X Loupe Browser software (H.W. and M.B.). Note that this approach can be error-prone due to shifts in the z-plane based on serial sectioning for each IHC sample and samples used for 10X Visium.

Computational data analysis

We used Python (v3.9.12) and the packages anndata (v0.11), scanpy (v1.9.8), squidpy (v1.4.1), sagenet (v1.1.0), Cell2module (GitHub version, retrieved in February 2024), pandas (v2.0.3), Torch (v2.1.1), Numpy (v1.24.3), Matplotlib (v3.7.2), Pyro (v1.8.6), SciPy (v1.11.3) and alpha_shape (GitHub clone c171a7d). We used R (v4.3.0) and the packages SingleCellExperiment (v1.24.0), ZellKonverter (v1.12.1), scater (v1.30.1), ComplexHeatmap (v2.16.0), glasso (v1.11), FSA (0.9.6), dplyr(1.1.4), ggplot2 (v3.5.1), igraph (v2.0.1.1) and scran (v1.28.2).

Genotyping

Barcode expression pre-preprocessing

Before analysing the 10X Visium data, we applied a filtering criterion of unique molecular identifier counts of >5,000. For the CytAssist platform, we excluded the outermost layer of spots due to unexpectedly high unique molecular identifier counts. In addition to manually identifying cancerous nodule regions, we annotated ‘normal tissue’ regions to acquire representation of areas without any cancerous cells. The selection of normal regions was based on a minimum distance of 250–700 µm from the nearest tumour, depending on the tissue section, to minimize contamination from adjacent tumour regions. Tumour nodules were defined as described in the ‘Histopathology’ section.

Bayesian modelling of perturbation probabilities from barcode counts

In our model, the observed expression count matrix D_s,b (spots s by barcode genes b) is assumed to follow a negative binomial distribution. This matrix has a mean λ_s,b and overdispersion ϕ_b. The overdispersion parameter ϕ is sampled from a Gamma distribution (shape = 1,000; rate = 0.03) skewed towards higher values to encourage the likelihood to approximate a Poisson distribution in the absence of overdispersion evidence.

The mean expression for each spot is calculated as:

$${{\rm{\lambda }}}_{{s},{g}}={{{\mu }}}_{{s}}{\sum }_{{r}}{{A}}_{{s},{r}}{\sum }_{{g}}{{G}}_{{r},{g}}{{B}}_{{g},{b}}{{\kappa}_{{b}}}+{{\rm{\xi }}}_b$$

where μ_s represents the sensitivity of each spot, A_s,r maps spots to clonal nodules r, G_r,g estimates the expected number of integrated copies of plasmid g, B_g,b links plasmids to their corresponding barcodes and κ_b is the barcode expression rate; ξ_b is a barcode-specific additive noise term. The per-nodule plasmid integration number (G_r,g) is modelled as an expected count of integration events, described by F_r,g,o. Here F_r,g,o captures the probability of no integration and higher indices reflect the integration of increasing numbers of copies. This is modelled using a Dirichlet distribution with a uniform concentration parameter and an order o = 6. This assumes the maximum of six copies of the same plasmid per clone, balancing the need to capture dosage-dependent variation with the practicality of limiting the number of parameters to be learnt. For normal tissue regions, the probability of perturbation presence was fixed to 1 × 10⁻³ to indicate near absence, but not zero, to prevent numerical instabilities.

Both κ_g and ξ_g are sampled from a weakly regularized exponential distribution with a rate of one. Spot sensitivity μ_s is modelled by a gamma distribution (shape = 3; rate = 0.3) that is weakly centred around one. This parameter accounts for both the sensitivity variability across 10X Visium spots and the dilution effects on the barcode signal due to varying tumour purity.

Perturbation probability model inference

We infer our Bayesian model using a variational posterior approximation. Specifically, we employ a log-normal guide distribution to approximate the parameters that have exponential and gamma distributed priors. In addition, we use a Dirichlet approximation for the posterior of F_r,g,o. The model and its inference framework are implemented in Pyro (v1.8.6)⁷⁵. The variational approximation is conducted via the stochastic variational inference method⁷⁶, employing the Adam optimizer set at a learning rate of 0.01 and using three samples for Kullback–Leibler divergence estimation. We perform inference over 10,000 gradient steps, monitoring the evidence lower bound to assess convergence.

Occurrence of individual perturbation combinations

Considering the probabilistic nature of our estimates for F_r,g,o, we utilized samples from the estimated posterior to analyse tumour populations. Due to uncertain integration copy number estimates, we focused on presence/absence categories. These probabilities were computed as 1 − F_r,g,o, and representative genotypes were sampled with Bernoulli distribution for each region and perturbation. We aggregated the data across 324 nodules into 256 possible genotype states, which allows us to compute medians and CIs for marginal integration numbers (indicating the count of different plasmids integrated) and individual perturbation for frequencies as well as frequencies of individual genotypes.

Although it may be tempting to interpret high frequencies of genotype occurrence as advantageous for tumour proliferation, such raw frequencies could be confounded by technical factors such as initial plasmid concentration and integration rate. To address this, we constructed a null hypothesis (H0) over the 256 individual genotype numbers. This hypothesis holds the marginal expected number of integrations and perturbation frequencies constant across the population, attributing variations solely to technical effects, and assumes that the genotypes are independently distributed.

In practice, we adjust the observed perturbation frequencies to account for technical biases by normalizing these frequencies—dividing the average observed perturbation frequency for each perturbation by the sum of all plasmid frequencies and multiplying by the expected number of integrations. We then simulate the distribution of genotypes by drawing Bernoulli samples using these rescaled probabilities for each perturbation. This process is repeated for the number of nodules (324) and the results are aggregated back into the 256 genotype states to create a sampling strategy that reflects the desired properties. By comparing deviations between 5,000 samples drawn from both the inferred posterior (observed) and the simulated H0 (expected), we can identify genotypes with significant tumorigenic effects (observed > expected) or disadvantageous effects (observed < expected).

Co-occurrence ORs and model of second-order interaction effect

With posterior estimates of the genotypes within the tumour population, we can test for interaction effects between individual perturbations. Wrange of variables accountinge categorize the frequencies of perturbations A and B into four groups: p₀₀(A−B−), p₀₁(A− and B+), p₁₀ (A+B−) and p₁₁(A+B+). The system can be described using a softmax linear model expressed as:

$${{p}}_{{i},\;{j}}=\exp ({{\rm{\theta }}}_{00}+{{i}{\rm{\theta}}}_{10}+{{j}{\rm{\theta }}}_{01}+{{ij}{\rm{\theta }}}_{11})/{\sum }_{ij}\exp ({{\rm{\theta }}}_{00}+{{i}{\rm{\theta }}}_{10}+{{j}{\rm{\theta }}}_{01}+{{ij}{\rm{\theta }}}_{11})$$

Here, θ₀₀ and θ₀₁ represent the effects of individual perturbations and θ₁₁ is the interaction effect. By setting θ₀₀ to zero to eliminate softmax non-identifiability and using Z as the normalization constant ${\Sigma}_{i,\;j}\;{\rm{e}}^{\theta_{i,\;j}}$, we derive the following relationships:

$${p}_{10}/{p}_{00}=[{e}^{{\theta }_{10}}/Z\;]/[1/Z\;]={e}^{{\theta }_{10}}$$

similarly,

$${p}_{01}/{p}_{00}={e}^{{\rm{\theta }}_{01}}$$

and

$${{{p}}}_{11}/{{{p}}}_{00}={{{e}}}^{{\rm{\theta }}_{10}+{\rm{\theta }}_{01}+{\rm{\theta }}_{11}}$$

Thus, computing the pairwise odds ratios (OR) effectively determines the interaction effect θ₁₁:

$$\begin{array}{l}{\rm{OR}}={{{p}}}_{11}{{{p}}}_{00}/{{{p}}}_{10}{{{p}}}_{01}={{{p}}}_{11}{{{p}}}_{00}/[{{{p}}}_{10}/{{{p}}}_{01}][{{{p}}}_{01}/{{{p}}}_{00}]\\\quad\quad={{{e}}}^{{\rm{\theta }}_{10}+{\rm{\theta }}_{01}+{\rm{\theta }}_{11}}/{{{e}}}^{{\rm{\theta }}_{10}}\,{{{e}}}^{{\rm{\theta}}_{01}}={{{e}}}^{{\rm{\theta }}_{11}}\end{array}$$

For each gene pair, we estimated the ORs and assessed their significance by drawing 2,000 samples from the posterior probability of perturbation presence for each nodule. By fitting a softmax linear model directly to the data and setting the interaction effect θ₁₁ to zero (OR = 1), we simulated the expected probabilities for the pairwise groups under the assumption of no interaction.

Genotype-to-phenotype GLM

To explore the relationships between inferred perturbation probabilities and phenotypic features—specifically, IHC staining status and gene expression—we employed a GLM model. Here the inferred perturbation probabilities serve as the explanatory variables X.

For the IHC staining analysis conducted at the nodule level, we used binary annotations (positive/negative) and modelled the outcomes with a Bernoulli distribution. The staining status for each nodule, Y_r,m,k (r, region; m, gene; k, sample) is modelled as:

$${{{Y}}}_{{{r}},{{m}},{{k}}} \sim {\rm{Bernoulli}}({\rm{\sigma }}({\sum }_{{{g}}}{{{X}}}_{{{r}},{{g}}}{{{w}}}_{{{g}},{{m}}})+{{{z}}}_{{{k}}})$$

where σ(x) =1 / 1 + e^−x is the sigmoid link function. The weight matrix w_g,m, akin to L1 regularization, is sampled from a Laplace distribution centred at zero with a scale parameter b. We set the scale to one for the intercept and 0.1 for the perturbation weights to impose stronger regularization on the perturbations. The batch effect z_k for each sample k is also sampled from a Laplace distribution (0, 1). The explanatory variable X_r,g is directly sampled from 1 − F_r,g,0, estimated by the perturbation probability model (non-learnable in the GLM).

Gene expression is modelled similarly, with few key differences. As gene expression is recorded as a non-zero integer at the spot level s, we use a Poisson distribution:

$${{Y}}_{{s},{m},{k}} \sim {\rm{Poisson}}({{{\mu }}}_{{s}}\exp ({\sum }_{{g}}{{X}}_{{s},{g}}{{w}}_{{g},{m}}+{{z}}_{k}))$$

In addition to the parameters used in the IHC model, spot sensitivity μ_s is factored in, sampled from the posterior of the perturbation probability model. X_s,g is calculated as ${\sum }_{{r}}{{A}}_{{s},{r}}(1-{{F}}_{{r},{g},0})$. The weight matrix w_g,m for perturbation-related weights is sampled from a Laplace distribution centred at zero with a strongly regularizing scale b = 1 × 10⁻³.

GLM inference

We infer our Bayesian model using a mean field variational posterior approximation. The model and its inference framework are implemented in Pyro (v1.8.6)⁷⁵. The variational approximation is conducted via the stochastic variational inferederive the following relationships:nce method^76,77, employing the Adam optimizer set at a learning rate of 0.01 and using three samples for Kullback–Leibler divergence estimation. At each gradient descent step, the parameters X and μ are sampled from their respective posterior distributions, as estimated by the perturbation probability model. This approach integrates the uncertainties associated with their estimation directly into the GLM framework. We perform inference over 2,000 gradient steps, monitoring the evidence lower bound to assess convergence.

Statistical analysis of the leave-one-out experiment

The leave-one-out experiment (Fig. 7) evaluates whether the removal of the previously identified epistatically interacting perturbation, VEGFA or mtCTNNB1, from the eight-plasmid mix affects the formation of HCC or CCA. Three setups were tested: (1) all eight perturbations, (2) all perturbations minus VEGFA and (3) all perturbations minus mtCTNNB1. For each setup, four mice were used. Tumours were identified and counted on two independent liver sections per mouse, with H&E histopathology and CK19 staining employed to differentiate between HCC and CCA subtypes. To assess group differences using a commonly applied approach, we performed a two-sided Kruskal–Wallis test, followed by Dunn’s post-hoc test with Holm–Bonferroni correction for multiple testing. We further used an alternative approach to model tumour burden across experimental groups while accounting for biological variability and applied a Bayesian hierarchical Poisson regression.

This statistical model aims to estimate the occurrence rate of HCC or CCA types across different experimental conditions, accounting for group-level and animal-specific variability. Tumour counts, y_i, observed for each slide i are modelled as Poisson-distributed outcomes. The rate parameter μ_i for each observation depends on the experimental condition and animal-specific effects. Group-level experiment effects (β_g) are drawn from a normal distribution β_g ~ N(0, σ_β), where variability controlled by the hyperprior σ_β ~ HalfCauchy(1.0). Animal-specific random effects (α_a) account for individual variability and are sampled from N(0, σ_a), with σ_α ~ HalfCauchy(0.05). The latter is assumed to be less strong than the group-level effect, which is reflected in a more regularised hyperprior. The model specifies the expected log-counts as log(μ_i) = β_g + α_a, where g and a index the group and animal associated with observation i. The tumour counts are then sampled from a Poisson distribution, y_i ~ Poisson(μ_i).

The posterior parameter distribution was estimated using 3,000 Markov chain Monte Carlo samples. The significance of differences between experimental effects was assessed by comparing the group-level parameters (β_g) across groups. P values were computed as the proportion of posterior samples where the differences in β_g exceeded or fell below zero, depending on the direction of the expected effect. Model-derived estimates corroborated the findings of the non-parametric Kruskal–Wallis analysis.

Reading Visium space ranger output into data objects

We utilized the anndata package (v0.11) in Python and the SingleCellExperiment package (v1.24.0) in R to generate and manage 10X Visium data objects. To facilitate communication between R and Python, we employed ZellKonverter (v1.12.1). For data processing, we employed Scanpy (v1.9.8), squidpy (v1.4.1) and scater (v1.30.1) in Python and R^76,78,79. We refined our analysis by subsetting all objects to include only features shared across all 11 slides, resulting in a total of 19,464 genes.

Publicly available databases of cell-type markers

We used scLiverDB, PanglaoDB and MSigDB to collect an initial set of marker genes for prevalent cell types and gene sets in normal and tumour liver tissues of mouse and human^80,81,82. This yielded a list of a total 2,323 genes.

Data normalization and preprocessing

After applying filtering criteria to exclude genes with raw counts of <10 or >1 × 10⁶ for any individual slide, as well as barcode genes, the count matrices were normalized to each spot to ensure a total count of 1 × 10⁴. Subsequently, the normalized values were log-transformed (log(x + 1)). This preprocessing was executed in Python using Scanpy (v1.9.8). Utilizing Scanpy (v1.9.8) with the Seurat flavour, we identified 15,000 highly variable genes for each of the 11 Visium and Visium CytAssist slides. The intersection of these sets resulted in 9,205 genes. Subsequently, in the final refinement step, we narrowed the gene set down to 80 core markers, resulting in 7,361 genes. This final gene set was employed for all subsequent phenotype analyses and visualizations.

Gene coexpression networks

Using the spot level-normalized expression values after filtering out uninformative genes, we conducted Gaussian graphical modelling⁸³ to infer a sparse gene coexpression network. We utilized the R package glasso (v1.11) for this purpose. The regularization parameter was optimized through a grid-search approach and set to 0.3. Following the construction of the initial Gaussian graphical modelling, we refined the network by filtering edges to retain only those with a Pearson’s pairwise correlation coefficient of at least 0.25. The isolated genes were, in turn, dropped from the graph. We used the R package igraph (v2.0.1.1) to visualize the graphs.

Nodule-level expression aggregation

We computed two types of aggregates for normalized expression values within each nodule: mean-based and quantile-based. For the mean-based aggregates, we calculated the average normalized expression of each gene across all spots within each nodule. For the quantile-based aggregates, we determined q₉₅ of expression values across all spots per nodule. We used these aggregates in the subsequent nodule-level analyses.

Binarization of nodule phenotypes

After quantile normalization, we computed the average of the scaled expression values for the core markers per phenotype, we then thresholded the values by 0.5, where all nodules with an aggregate value of >0.5 are considered to have the corresponding phenotype signature and otherwise not. The mean-based aggregates for each gene are quantile-normalized further by

$${x}-{{q}}_{25}/{q}_{99}-{{q}}_{25}$$

where q₂₅ and q₉₉ represent the 25th and 99th quantile values of the aggregate gene expression for the corresponding gene across all nodules and all slides. We then binarized the values >0.5 as one (on) and otherwise zero (off).

Binarization of nodule TME signatures

We binarized the TME signatures following the same procedure as for nodule phenotypes but using the initial quantile-based aggregates instead.

Phenotype and TME heatmaps

We generated heatmaps of scaled gene expression using the processed expression values at the nodule level, employing ComplexHeatmap (v2.16.0)⁸⁴. Clustering of both rows and columns was performed based on Spearman’s correlations. In addition, hierarchical clustering was applied to subdivide genes into clusters. The colour bar associated with genes indicates their corresponding phenotypes, with emphasis on the core markers. An attached annotation heatmap illustrates scaled (p¹⁰) estimated plasmid probabilities per nodule.

Spatial integration and nodule unification

To integrate all slides into a unified embedding space and classify spots based on their phenotypic signatures in an unbiased manner, we employed an ensemble spatially-aware classifier implemented in SageNet (v1.1.0)⁸⁵. Data from all slides were trained and fed into this classifier. Subsequently, we clustered the spots within the embedded space using Scanpy’s wrapper of Leiden clustering (with a resolution of one)⁸⁶. We then performed voting classification to classify nodules to the most-dominant class across the spots belonging to the corresponding nodules. We call these classes the ‘unified nodule annotations’. Finally, we extracted spatially informative genes from the SageNet model.

Inter-nodule differential gene expression

To delve deeper into inter-nodule transcriptional differences, we conducted differential gene expression analysis using the FindMarkers method from the R package scran (v1.28.2). This method allowed us to perform a light-weight differential gene expression analysis on the unified nodule annotations.

Cell type-informed factor analysis

We concatenated all raw anndata objects and subsetted them to the set of ‘core markers’ and associated genes (as listed in Fig. 5) as well as 500 highly variable genes across slides. We then used cell2module (github.com/vitkl/cell2module) to perform non-negative matrix factorization. The cell2module model treats raw RNA count data D as negative-binomial (NB) distributed, given transcription rate μ_c,g and a range of variables accounting for technical effects:

$${{D}}_{{c},{g}} \sim {\rm{NB}}({{\mu }}={{{\mu }}}_{{c},{g}},{{\rm{\alpha }}}_{{a},{g}})$$

and

$${{{\mu }}}_{{c},{g}}=({\sum }_{{f}}{{w}}_{{c},\;{f},}{{g}}_{{f},{g}}+{{s}}_{{e},{g}}) {{y}}_{{c}}$$

where μ_c,g denotes expected RNA count g in each cell c, α_a,g denotes the per gene g stochastic/unexplained overdispersion for each covariate α, w_c,f denotes cell loadings of each factor f for each cell c, g_f,g denotes gene loadings of each factor f for each cell c, s_e,g denotes additive background for each gene g and for each experiment c to account for contaminating RNA and y_c denotes normalization for each spot c to account for RNA-detection sensitivity, sequencing depth. We recovered 40 factors representing groups of coexpressing cell-type signatures using the default cell2module parameters. After training, we inferred the posterior of the gene loadings per factor. Subsequently, we extracted genes with the top five posterior median values and compared them with predefined marker gene lists per cell type. Finally, we mapped each factor to the cell type with the highest number of overlapping genes.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability

We used publicly available datasets from scLiverDB (https://guolab.wchscu.cn/liverdb#!/), PanglaoDB (https://panglaodb.se/), MSigDB (https://www.gsea-msigdb.org/gsea/msigdb), GEO (https://www.ncbi.nlm.nih.gov/geo/), MGI (https://www.informatics.jax.org/) and the LiverCellAtlas (https://www.livercellatlas.org/). We deposited data related to this manuscript to https://zenodo.org/records/10986436 (ref. ⁸⁷). In addition, we have launched a web browser for interactive data analyses (https://chocolat-g2p.dkfz.de/). Source data are provided with this paper.

Code availability

Scripts and custom code for data analysis related to this manuscript are available at https://github.com/gerstung-lab/CHOCOLAT-G2P/ (ref. ⁸⁸).

References

Wheeler, D. A. & Roberts, L. R. Comprehensive and integrative genomic characterization of hepatocellular carcinoma. Cell 169, 1327–1341 (2017).
Article CAS PubMed Central Google Scholar
The ICGC/TCGA Pan-Cancer Analysis of Whole Genomes Consortium. Pan-cancer analysis of whole genomes. Nature 578, 82–93 (2020).
Article CAS Google Scholar
Hodis, E. et al. Stepwise-edited, human melanoma models reveal mutations’ effect on tumor and microenvironment. Science 376, eabi8175 (2022).
Article CAS PubMed PubMed Central Google Scholar
van de Haar, J. et al. Identifying epistasis in cancer genomes: a delicate affair. Cell 177, 1375–1383 (2019).
Article PubMed PubMed Central Google Scholar
Lewis, S. M. et al. Spatial omics and multiplexed imaging to explore cancer biology. Nat. Methods 18, 997–1012 (2021).
Article CAS PubMed Google Scholar
Costanzo, M. et al. Global genetic networks and the genotype-to-phenotype relationship. Cell 177, 85–100 (2019).
Article CAS PubMed Google Scholar
Weber, J., Braun, C. J., Saur, D. & Rad, R. In vivo functional screening for systems-level integrative cancer genomics. Nat. Rev. Cancer 20, 573–593 (2020).
Article CAS PubMed Google Scholar
Weber, J. et al. CRISPR/Cas9 somatic multiplex-mutagenesis for high-throughput functional cancer genomics in mice. Proc. Natl Acad. Sci. USA 112, 13982–13987 (2015).
Article CAS PubMed PubMed Central Google Scholar
Rogers, Z. N. et al. A quantitative and multiplexed approach to uncover the fitness landscape of tumor suppression in vivo. Nat. Methods 14, 737–742 (2017).
Article CAS PubMed PubMed Central Google Scholar
Wang, G. et al. Mapping a functional cancer genome atlas of tumor suppressors in mouse liver using AAV-CRISPR-mediated direct in vivo screening. Sci. Adv. 4, eaao5508 (2018).
Article PubMed PubMed Central Google Scholar
Blair, L. M. et al. Oncogenic context shapes the fitness landscape of tumor suppression. Nat. Commun. 14, 6422 (2023).
Article CAS PubMed PubMed Central Google Scholar
Li, S. et al. A combinatorial genetic strategy for exploring complex genotype–phenotype associations in cancer. Nat. Genet. 56, 371–376 (2024).
Article CAS PubMed PubMed Central Google Scholar
de Galarreta, M. R. et al. β-Catenin activation promotes immune escape and resistance to anti-PD-1 therapy in hepatocellular carcinoma. Cancer Discov. 9, 1124–1141 (2019).
Article Google Scholar
Dhainaut, M. et al. Spatial CRISPR genomics identifies regulators of the tumor microenvironment. Cell 185, 1223–1239 (2022).
Article CAS PubMed PubMed Central Google Scholar
Rovira-Clavé, X. et al. Spatial epitope barcoding reveals clonal tumor patch behaviors. Cancer Cell 40, 1423–1439 (2022).
Article PubMed PubMed Central Google Scholar
Yao, D. et al. Scalable genetic screening for regulatory circuits using compressed Perturb-seq. Nat. Biotechnol. 42, 1282–1295 (2024).
Article CAS PubMed Google Scholar
Kudo, T. et al. Multiplexed, image-based pooled screens in primary cells and tissues with PerturbView. Nat. Biotechnol. https://doi.org/10.1038/s41587-024-02391-0 (2024).
Chen, X. & Calvisi, D. F. Hydrodynamic transfection for generation of novel mouse models for liver cancer research. Am. J. Pathol. 184, 912–923 (2014).
Article CAS PubMed PubMed Central Google Scholar
Molina-Sánchez, P. et al. Cooperation between distinct cancer driver genes underlies intertumor heterogeneity in hepatocellular carcinoma. Gastroenterology 159, 2203–2220 (2020).
Article PubMed Google Scholar
Tang, M. et al. Liver cancer heterogeneity modeled by in situ genome editing of hepatocytes. Sci. Adv. 8, eabn5683 (2022).
Article CAS PubMed PubMed Central Google Scholar
Nilsson, M., Barbany, G., Antson, D. O., Gertow, K. & Landegren, U. Enhanced detection and distinction of RNA by enzymatic probe ligation. Nat. Biotechnol. 18, 791–793 (2000).
Article CAS PubMed Google Scholar
Rudalska, R. et al. In vivo RNAi screening identifies a mechanism of sorafenib resistance in liver cancer. Nat. Med. 20, 1138–1146 (2014).
Article CAS PubMed PubMed Central Google Scholar
Wang, C., Lu, T., Emanuel, G., Babcock, H. P. & Zhuang, X. Imaging-based pooled CRISPR screening reveals regulators of lncRNA localization. Proc. Natl Acad. Sci. USA 116, 10842–10851 (2019).
Article CAS PubMed PubMed Central Google Scholar
Guilliams, M. et al. Spatial proteogenomics reveals distinct and evolutionarily conserved hepatic macrophage niches. Cell 185, 379–396 (2022).
Article CAS PubMed PubMed Central Google Scholar
Zucman-Rossi, J. et al. Differential effects of inactivated axin1 and activated β-catenin mutations in human hepatocellular carcinomas. Oncogene 26, 774–780 (2007).
Article CAS PubMed Google Scholar
Cordell, H. J. Epistasis: what it means, what it doesn’t mean, and statistical methods to detect it in humans. Hum. Mol. Genet. 11, 2463–2468 (2002).
Article CAS PubMed Google Scholar
Fan, B. et al. Cholangiocarcinomas can originate from hepatocytes in mice. J. Clin. Invest. 122, 2911–2915 (2012).
Article CAS PubMed PubMed Central Google Scholar
Zender, S. et al. A critical role for notch signaling in the formation of cholangiocellular carcinomas. Cancer Cell 23, 784–795 (2013).
Article CAS PubMed Google Scholar
Lee, Y.-S. et al. SLC15A2 genomic variation is associated with the extraordinary response of sorafenib treatment: whole-genome analysis in patients with hepatocellular carcinoma. Oncotarget 6, 16449–16460 (2015).
Article PubMed PubMed Central Google Scholar
Jendrek, S. T. et al. Anti-GP2 IgA autoantibodies are associated with poor survival and cholangiocarcinoma in primary sclerosing cholangitis. Gut 66, 137–144 (2017).
Article CAS PubMed Google Scholar
Hildebrandt, F. et al. Spatial transcriptomics to define transcriptional patterns of zonation and structural components in the mouse liver. Nat. Commun. 12, 7046 (2021).
Article CAS PubMed PubMed Central Google Scholar
Halpern, K. B. et al. Single-cell spatial reconstruction reveals global division of labor in the mammalian liver. Nature 542, 352–356 (2017).
Article CAS PubMed PubMed Central Google Scholar
Zhang, T. et al. A novel liver zonation phenotype-associated molecular classification of hepatocellular carcinoma. Front. Immunol. 14, 1140201 (2023).
Article CAS PubMed PubMed Central Google Scholar
Ng, C. K. Y., Piscuoglio, S. & Terracciano, L. M. Molecular classification of hepatocellular carcinoma: the view from metabolic zonation. Hepatology 66, 1377–1380 (2017).
Article PubMed Google Scholar
Désert, R. et al. Human hepatocellular carcinomas with a periportal phenotype have the lowest potential for early recurrence after curative resection. Hepatology 66, 1502–1518 (2017).
Article PubMed Google Scholar
Chung, A. et al. Liver cancer initiation is dependent on metabolic zonation but decoupled from premalignant clonal expansion. Preprint at bioRxiv https://doi.org/10.1101/2024.01.10.575013 (2024).
Wei, Y. et al. Liver homeostasis is maintained by midlobular zone 2 hepatocytes. Science 371, eabb1625 (2021).
Article CAS PubMed PubMed Central Google Scholar
Skinner, O. S. et al. Salvage of ribose from uridine or RNA supports glycolysis in nutrient-limited conditions. Nat. Metab. 5, 765–776 (2023).
Article CAS PubMed PubMed Central Google Scholar
Nwosu, Z. C. et al. Uridine-derived ribose fuels glucose-restricted pancreatic cancer. Nature 618, 151–158 (2023).
Article CAS PubMed PubMed Central Google Scholar
Font-Burgada, J. et al. Hybrid periportal hepatocytes regenerate the injured liver without giving rise to cancer. Cell 162, 766–779 (2015).
Article CAS PubMed PubMed Central Google Scholar
Medina, R. et al. Epigenetic control of cell cycle-dependent histone gene expression is a principal component of the abbreviated pluripotent cell cycle. Mol. Cell. Biol. 32, 3860–3871 (2012).
Article CAS PubMed PubMed Central Google Scholar
Foster, D. S. et al. Multiomic analysis reveals conservation of cancer-associated fibroblast phenotypes across species and tissue of origin. Cancer Cell 40, 1392–1406 (2022).
Article CAS PubMed PubMed Central Google Scholar
Montori, M. et al. Cancer-associated fibroblasts in cholangiocarcinoma: current knowledge and possible implications for therapy. J. Clin. Med. 11, 6498 (2022).
Article CAS PubMed PubMed Central Google Scholar
Wang, Y. et al. Intratumoral erythroblastic islands restrain anti-tumor immunity in hepatoblastoma. Cell Rep. Med. 4, 101044 (2023).
Article CAS PubMed PubMed Central Google Scholar
Fang, J. et al. Genome-wide mapping of cancer dependency genes and genetic modifiers of chemotherapy in high-risk hepatoblastoma. Nat. Commun. 14, 4003 (2023).
Article CAS PubMed PubMed Central Google Scholar
Ma, C. et al. Platelets control liver tumor growth through P2Y12-dependent CD40L release in NAFLD. Cancer Cell 40, 986–998 (2022).
Article CAS PubMed PubMed Central Google Scholar
Z, Y. et al. Mast cells mobilize myeloid-derived suppressor cells and T_reg cells in tumor microenvironment via IL-17 pathway in murine hepatocarcinoma model. PLoS ONE 5, e8922 (2010).
Article Google Scholar
Somasundaram, R. et al. Tumor-infiltrating mast cells are associated with resistance to anti-PD-1 therapy. Nat. Commun. 12, 346 (2021).
Article CAS PubMed PubMed Central Google Scholar
Morgan, D. & Tergaonkar, V. Unraveling B cell trajectories at single cell resolution. Trends Immunol. 43, 210–229 (2022).
Article CAS PubMed Google Scholar
Xue, R. et al. Liver tumour immune microenvironment subtypes and neutrophil heterogeneity. Nature 612, 141–147 (2022).
Article CAS PubMed Google Scholar
Xie, X. et al. Single-cell transcriptome profiling reveals neutrophil heterogeneity in homeostasis and infection. Nat. Immunol. 21, 1119–1133 (2020).
Article CAS PubMed PubMed Central Google Scholar
Benhamouche, S. et al. Apc tumor suppressor gene is the ‘zonation-keeper’ of mouse liver. Dev. Cell 10, 759–770 (2006).
Article CAS PubMed Google Scholar
Gallage, S. et al. A researcher’s guide to preclinical mouse NASH models. Nat. Metab. 4, 1632–1649 (2022).
Article CAS PubMed Google Scholar
Imle, R. et al. Somatic gene delivery faithfully recapitulates a molecular spectrum of high-risk sarcomas. Nat. Commun. 16, 5283 (2025).
Article CAS PubMed PubMed Central Google Scholar
Hebert, J. D. et al. Combinatorial in vivo genome editing identifies widespread epistasis and an accessible fitness landscape during lung tumorigenesis. Mol. Biol. Evol. 42, msaf023 (2025).
Article CAS PubMed PubMed Central Google Scholar
Askary, A. et al. In situ readout of DNA barcodes and single base edits facilitated by in vitro transcription. Nat. Biotechnol. 38, 66–75 (2020).
Article CAS PubMed Google Scholar
Replogle, J. M. et al. Combinatorial single-cell CRISPR screens by direct guide RNA capture and targeted sequencing. Nat. Biotechnol. 38, 954–961 (2020).
Article CAS PubMed PubMed Central Google Scholar
Gu, J. et al. Mapping multimodal phenotypes to perturbations in cells and tissue with CRISPRmap. Nat. Biotechnol. https://doi.org/10.1038/s41587-024-02386-x (2024).
Rood, J. E., Hupalowska, A. & Regev, A. Toward a foundation model of causal cell and tissue biology with a Perturbation Cell and Tissue Atlas. Cell 187, 4520–4545 (2024).
Article CAS PubMed Google Scholar
Amen, A. M. et al. Endogenous spacing enables co-processing of microRNAs and efficient combinatorial RNAi. Cell Rep. Methods 2, 100239 (2022).
Article CAS PubMed PubMed Central Google Scholar
DeWeirdt, P. C. et al. Optimization of AsCas12a for combinatorial genetic screens in human cells. Nat. Biotechnol. 39, 94–104 (2021).
Article CAS PubMed Google Scholar
Vickovic, S. et al. High-definition spatial transcriptomics for in situ tissue profiling. Nat. Methods 16, 987–990 (2019).
Article CAS PubMed PubMed Central Google Scholar
Stickels, R. R. et al. Highly sensitive spatial transcriptomics at near-cellular resolution with Slide-seqV2. Nat. Biotechnol. 39, 313–319 (2021).
Article CAS PubMed Google Scholar
Zhang, D. et al. Inferring super-resolution tissue architecture by integrating spatial transcriptomics with histology. Nat. Biotechnol. 42, 1372–1377 (2024).
Article CAS PubMed PubMed Central Google Scholar
Schott, M. et al. Open-ST: high-resolution spatial transcriptomics in 3D. Cell 187, 3953–3972 (2024).
Article CAS PubMed Google Scholar
Khafizov, R. et al. Sub-cellular imaging of the entire protein-coding human transcriptome (18933-plex) on FFPE tissue using spatial molecular imaging. Preprint at bioRxiv https://doi.org/10.1101/2024.11.27.625536 (2024).
Polański, K. et al. Bin2cell reconstructs cells from high resolution Visium HD data. Bioinformatics. 40, btae546 (2024).
Article PubMed PubMed Central Google Scholar
Tschaharganeh, D. F. et al. p53-Dependent nestin regulation links tumor suppression to cellular plasticity in liver cancer. Cell 158, 579–592 (2014).
Article CAS PubMed PubMed Central Google Scholar
Revia, S. et al. Histone H3K27 demethylase KDM6A is an epigenetic gatekeeper of mTORC1 signalling in cancer. Gut 71, 1613–1628 (2022).
CAS PubMed Google Scholar
Chi, Y. et al. Comparative impact of dietary carbohydrates on the liver transcriptome in two strains of mice. Physiol. Genomics 53, 456–472 (2021).
Article CAS PubMed PubMed Central Google Scholar
Ringwald, M. et al. Mouse Genome Informatics (MGI): latest news from MGD and GXD. Mamm. Genome 33, 4–18 (2022).
Article PubMed Google Scholar
Zhu, C. et al. MLL3 regulates the CDKN2A tumor suppressor locus in liver cancer. eLife 12, e80854 (2023).
Article CAS PubMed PubMed Central Google Scholar
Beliveau, B. J. et al. OligoMiner provides a rapid, flexible environment for the design of genome-scale oligonucleotide in situ hybridization probes. Proc. Natl Acad. Sci. USA 115, E2183–E2192 (2018).
Article CAS PubMed PubMed Central Google Scholar
Fellmann, C. et al. An optimized microRNA backbone for effective single-copy RNAi. Cell Rep. 5, 1704–1713 (2013).
Article CAS PubMed Google Scholar
Bingham, E. et al. Pyro: deep universal probabilistic programming. J. Mach. Learn. Res. 20, 973–978 (2019).
Google Scholar
Wolf, F. A., Angerer, P. & Theis, F. J. SCANPY: large-scale single-cell gene expression data analysis. Genome Biol. 19, 15 (2018).
Article PubMed PubMed Central Google Scholar
Hoffman, M. D., Blei, D. M., Wang, C. & Paisley, J. Stochastic variational inference. J. Mach. Learn. Res. 14, 1303–1347 (2013).
Google Scholar
Palla, G. et al. Squidpy: a scalable framework for spatial omics analysis. Nat. Methods 19, 171–178 (2022).
Article CAS PubMed PubMed Central Google Scholar
McCarthy, D. J., Campbell, K. R., Lun, A. T. L. & Wills, Q. F. Scater: pre-processing, quality control, normalization and visualization of single-cell RNA-seq data in R. Bioinformatics 33, 1179–1186 (2017).
Article CAS PubMed PubMed Central Google Scholar
Pan, Q. et al. scLiverDB: a database of human and mouse liver transcriptome landscapes at single-cell resolution. Small Methods 7, e2201421 (2023).
Article PubMed Google Scholar
Franzén, O., Gan, L.-M. & Björkegren, J. L. M. PanglaoDB: a web server for exploration of mouse and human single-cell RNA sequencing data. Database 2019, baz046 (2019).
Article PubMed PubMed Central Google Scholar
Liberzon, A. et al. The Molecular Signatures Database (MSigDB) hallmark gene set collection. Cell Syst. 1, 417–425 (2015).
Article CAS PubMed PubMed Central Google Scholar
Friedman, J., Hastie, T. & Tibshirani, R. Sparse inverse covariance estimation with the graphical lasso. Biostatistics 9, 432–441 (2008).
Article PubMed Google Scholar
Gu, Z., Eils, R. & Schlesner, M. Complex heatmaps reveal patterns and correlations in multidimensional genomic data. Bioinformatics 32, 2847–2849 (2016).
Article CAS PubMed Google Scholar
Heidari, E. et al. Supervised spatial inference of dissociated single-cell data with SageNet. Preprint at bioRxiv https://doi.org/10.1101/2022.04.14.488419 (2022).
Traag, V. A., Waltman, L. & van Eck, N. J. From Louvain to Leiden: guaranteeing well-connected communities. Sci. Rep. 9, 5233 (2019).
Article CAS PubMed PubMed Central Google Scholar
Breinig, M. et al. Integrated in vivo combinatorial functional genomics and spatial transcriptomics of tumors to decode genotype to phenotype relationships. Zenodo https://zenodo.org/records/10986436 (2025).
Breinig, M. et al. Integrated in vivo combinatorial functional genomics and spatial transcriptomics of tumors to decode genotype to phenotype relationships. Github https://github.com/gerstung-lab/CHOCOLAT-G2P/ (2025).

Download references

Acknowledgements

We thank all of the members of the Tschaharganeh, Sahm, Gerstung and Stegle laboratories. We thank the German Cancer Research Center Central Animal Laboratory, the Center for Model Systems and Comparative Pathology of the Institute of Pathology Heidelberg and the Tissue Bank of the National Center for Tumor Diseases for excellent technical support. D.F.T. was supported by the European Research Council (ERC Starting Grant ‘CrispSCNAs’, grant number 948172) and the Deutsche Forschungsgemeinschaft (DFG; grant number TS 293/3-1).

Funding

Open access funding provided by Deutsches Krebsforschungszentrum (DKFZ).

Author information

Artem Lomakin
Present address: Stanford Cancer Institute, School of Medicine, Stanford University, Stanford, CA, USA
These authors contributed equally: Marco Breinig, Artem Lomakin, Elyas Heidari.

Authors and Affiliations

Institute of Pathology, University Hospital Heidelberg, Heidelberg, Germany
Marco Breinig, Lio Böse, Luise Butthof, Lena Wendler-Link, Hendrik Wiethoff, Peter Schirmacher & Darjus F. Tschaharganeh
Artificial Intelligence in Oncology, German Cancer Research Center (DKFZ), Heidelberg, Germany
Artem Lomakin, Elyas Heidari, Gleb Rukhovich & Moritz Gerstung
Computational Genomics and Systems Genetics, German Cancer Research Center (DKFZ), Heidelberg, Germany
Elyas Heidari & Oliver Stegle
Department of Neuropathology, University Hospital Heidelberg, Heidelberg, Germany
Michael Ritter & Felix Sahm
Clinical Cooperation Unit Neuropathology, German Consortium for Translational Cancer Research (DKTK), German Cancer Research Center (DKFZ), Heidelberg, Germany
Michael Ritter & Felix Sahm
Center for Model System and Comparative Pathology, Institute of Pathology, University Hospital Heidelberg, Heidelberg, Germany
Tanja Poth
Robert Bosch Center for Tumor Diseases, Stuttgart, Germany
Moritz Gerstung
Medical Faculty, Eberhard Karls University, Tübingen, Germany
Moritz Gerstung
Universty Hospital Tübingen, Tübingen, Germany
Moritz Gerstung
‘Cell Plasticity and Epigenetic Remodeling’ Helmholtz-University Group, German Cancer Research Center (DKFZ), Heidelberg, Germany
Darjus F. Tschaharganeh

Authors

Marco Breinig
View author publications
Search author on:PubMed Google Scholar
Artem Lomakin
View author publications
Search author on:PubMed Google Scholar
Elyas Heidari
View author publications
Search author on:PubMed Google Scholar
Michael Ritter
View author publications
Search author on:PubMed Google Scholar
Gleb Rukhovich
View author publications
Search author on:PubMed Google Scholar
Lio Böse
View author publications
Search author on:PubMed Google Scholar
Luise Butthof
View author publications
Search author on:PubMed Google Scholar
Lena Wendler-Link
View author publications
Search author on:PubMed Google Scholar
Hendrik Wiethoff
View author publications
Search author on:PubMed Google Scholar
Tanja Poth
View author publications
Search author on:PubMed Google Scholar
Felix Sahm
View author publications
Search author on:PubMed Google Scholar
Peter Schirmacher
View author publications
Search author on:PubMed Google Scholar
Oliver Stegle
View author publications
Search author on:PubMed Google Scholar
Moritz Gerstung
View author publications
Search author on:PubMed Google Scholar
Darjus F. Tschaharganeh
View author publications
Search author on:PubMed Google Scholar

Contributions

M.B. conceptualized the project, and designed the methodology and experiments. L. Böse, L. Butthof and L.W.-L. cloned plasmids, and performed animal experiments and IHC, with help from M.B. M.R. performed 10X Visium spatial transcriptomics and next-generation sequencing with help from M.B. P.S., D.F.T. and F.S. provided reagents. T.P. supervised the technical processing of FFPE liver samples and preparation of IHC stainings. H.W. and D.F.T. performed the histopathological analyses. H.W. and M.B. performed the spatial transcriptomics-guided nodule annotation. M.B. performed the initial bulk RNA-seq data analysis for probe redeployment. E.H. and A.L. performed 10X Visium data preprocessing. A.L. implemented spatial transcriptomics-based genotyping, with input from E.H., M.B., O.S. and M.G. E.H. carried out spatial transcriptomics-based phenotyping, with input from A.L., M.B, O.S. and M.G. A.L. and E.H. performed statistical analyses, with input from O.S. and M.G. M.B., A.L. and E.H. created figures, with input from D.F.T., O.S. and M.G. A.L. and G.R. developed the dedicated webtool interface, with input from E.H. and M.B. M.B., A.L. and E.H. wrote the manuscript, with input from D.F.T. and M.G. All authors read and approved the final paper.

Corresponding authors

Correspondence to Marco Breinig, Moritz Gerstung or Darjus F. Tschaharganeh.

Ethics declarations

Competing interests

All authors declare no competing interests.

Peer review

Peer review information

Nature Biomedical Engineering thanks Benjamin Izar and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data

Extended Data Fig. 1 Overview of perturbation plasmid design and molecular barcoding.

Schematic overview of Sleeping Beauty transposon perturbation plasmids to ectopically overexpress genes-of-interest (oncogenic-driver perturbations) or shRNA to enable gene knockdown (tumor-suppressor perturbations). Functional elements are highlighted. IR: Inverted/direct repeats of sleeping beauty transposon; GFP: green fluorescent protein, mK2: mKate2 red fluorescent protein, EF1: Polymerase II promoter; U6: Polymerase III promoter; pA: polyadenylation signal; sh: short hairpin RNA embedded in mir-E context; BC: a barcode in which 3 redeployed RTL-probe capture sequences (as indicated) are embedded. Note that we used Visium Mouse Transcriptome Probe Set v1 to derive barcodes. Each 50 nt barcode is separated and flanked by ca. 20 nt spacer sequences to avoid potential steric hindrance during hybridization. Spacer sequences used were derived from T7 and T3 promoters and/or AsCas12a-DR sequences and/or 10X Capture sequences cs1 and cs2 (not shown). Functionality of spacer sequences was not tested in this study. Note that plasmids were equipped with multiple orthogonal barcodes at varying positions. See Methods for further information.

Extended Data Fig. 2 Sample processing and ROI selection for spatial transcriptomics.

a, RUBIX with a pool-of-8 plasmid mix results in rapid liver tumour development. Injection of a shRNA targeting Renilla (shRen; matching total plasmid concentration for pool-of-8 mix) served as control, n = 2 each group. Representative H&E-stained samples revealing multiple tumour nodules from the two individual animals are shown. Absence of tumours in the control group (shRen only) indicates that random integration of transposon plasmids itself is unlikely to contribute to tumorigenesis. b, Tissue preprocessing. Following liver tumour development, livers were extracted, divided and processed to FFPE as well as fresh frozen specimens. FFPE samples were initially sectioned to enable sample selection. c, C-G2P liver samples. Overview of three representative FFPE samples used in this study. A total of 513 tumour nodules (red outline) were identified based on histopathological examination (based on H&E). d, Overview of ROIs selected for 10X Visium. 6 segregated regions were selected across 3 FFPE samples. Squares indicate approximate position of ROIs selected for ST. Orange: first 10X Visium run, light orange: first 10X Visium run, replicate ROI; blue: second 10X Visium run; black: 10X Visium CytAssist run. Note overlap between ROIs, where serial sections are used for 10X Visium. e, 10X Visium workflow. Samples for ST are derived directly from FFPE blocks and mounted on 10X Visium slides. f, 10X Visium CytAssist workflow. Samples for ST are derived from sections already mounted on glass slides and transferred to 10X Visium slides using the 10X CytAssist instrument. g, Overview of all samples used for ST. 12 samples from a single RUBIX experiment with two animals were used for 10X Visium in this study. Respective H&E stainings are depicted. Note that the utility of sample ML-II_B_2Cyt is constrained by tissue detachment of the sample during the processing for 10X Visium CytAssist and was not included for further analyses (asterisk). QC summary stats for each sample related to Visium runs performed are provided.

Extended Data Fig. 3 Specific detection of redeployable barcodes by available 10X RTL probes.

a, Expression of all potentially redeployable barcodes derived from sensory receptor transcripts in murine liver. Transcripts associated with three groups of sensory receptors - Tasr (red; n = 26), Vmnr (yellow; n = 194), Olfr (blue; n = 956). Redeployed barcodes (black) are well separated from endogenous sensory receptor transcripts (total n = 1,216). Data for all 11 C-G2P samples generated in this study from a single RUBIX experiment with two animals is depicted. Reference control liver datasets (NAFLD, StSt) are from¹⁷. b, Sensory receptor-associated transcripts expressed in murine liver. Sensory receptor-associated transcripts that reveal average expression between <0.01 and >0.008 in murine liver are depicted. Expression values are aggregated across all 11 C-G2P samples shown in a. c, Expression of all 38 redeployed barcodes used in this study. Grouped according to associated perturbation (see Extended Data Fig. 1, Methods). Expression is averaged across 6 primary C-G2P samples. Note that 5/38 revealed insufficient signal (average expression < 0.05). In all panels, expression values are log1p-transformed. Barcodes expressed from a Pol III (hU6) promotor are highlighted. Barcodes expressed in 5′ mirE position for shRNA constructs are highlighted. Simplified plasmids maps are depicted to illustrate barcode positioning.

Extended Data Fig. 4 Spatially resolved triplet barcode expression enables identification of all 8 perturbations used.

Spatially resolved expression of triplet barcodes for each of the 8 perturbations across all 11 samples in this study from a single RUBIX experiment with two animals. Gene expression is log1p-transformed and quantile-rescaled, as in Fig. 3c. The quantitative barcode expression for all samples can be explored through the interactive web browser (https://chocolat-g2p.dkfz.de/).

Extended Data Fig. 5 Spatially resolved perturbation mapping.

Spatially resolved visualization of the inferred probabilities indicating the presence or absence of each of the 8 perturbations associated with annotated tumour nodules for all 11 samples used in this study from a single RUBIX experiment with two animals (Methods), as in Fig. 3d. Perturbation probabilities for all samples can be explored through the interactive web browser (https://chocolat-g2p.dkfz.de/).

Extended Data Fig. 6 Quantitative reproducibility of spatial perturbation mapping.

a, Quantitative reproducibility. Scatterplots of inferred perturbation probabilities for nodules on the primary section to those on the corresponding replica sections, with Pearson’s correlation values displayed. In total, 136 nodule pairs were analysed. b, Matching nodules across samples. Spatial maps of the inferred probabilities (as in Fig. 3c) indicating the presence or absence of each of the 8 perturbations associated with annotated tumour nodules for all samples that have matching ROIs from a single RUBIX experiment with two animals. Matching nodules were manually annotated (Methods). Addition of “_Cyt” in sample name indicates use of 10X CytAssist. Number of matching nodules is indicated.

Extended Data Fig. 7 Validation of inferred perturbation integration.

a–c, Perturbation–phenotype association. Spatially resolved visualization of the inferred probabilities indicating the presence or absence of each of the 8 perturbations associated with annotated tumour nodules (Methods). a, A representative sample is displayed. A generalized linear model (GLM) predicts phenotype expression signals based on the estimated probabilities of perturbation presence (Methods). b, Phenotypes are defined as direct target transcripts associated with perturbations such as shKmt2c-Kmt2c and shTrp53-Trp53. Expression data are log1p-transformed. Note that shTrp53 is linked to a GFP peptide barcode and shKmt2c is linked to a RFP barcode (Extended Data Fig. 1). c, Hence we infer shTrp53-GFP-positive phenotype and shKmt2c-RFP-positive phenotype. Representative IHCs for a corresponding ROI on a serial section. Baseline depicts background phenotype marker expression. Data are presented as feature coefficients shown as mean and error-bars depict 3σ confidence intervals. As in Fig. 3c, and d. d, H&E and IHC for GFP, RFP, GS. Three representative FFPE samples from a single RUBIX experiment with two animals were sectioned and stained for H&E (see Extended Data Fig. 2). GFP and RFP IHC staining was performed on two individual serial sections. GS IHC staining was performed on serial sections. GFP and RFP were embedded in perturbation plasmids as orthogonal barcodes (see Methods and Extended Data Fig. 1). GS is a well-known marker for liver WNT/mtCtnnb1-signalling activity (see Fig. 3d). e, GFP and RFP detection for the same hepatocellular tumour nodules to spatially map peptide barcode combinations associated with introduced perturbations. Left: H&E. Zoom-in: GFP and RFP respectively. Note that one tumour nodule is positive for RFP alone whereas the other tumour nodule is positive for both GFP as well as RFP peptide barcodes. A representative example is shown. Data is derived from 324 nodules across 6 topographically separated regions used for 10X Visium from a single RUBIX experiment with two animals. Mapping GFP and RFP IHC data is derived from 3 corresponding sections from a single RUBIX experiment with two animals.

Extended Data Fig. 8 C-G2P enables spatial genotype mapping.

a, Converting barcode signals to genotype maps. 2⁸ powerset embedding of spatially mapped perturbations for all 11 samples used in this study from a single RUBIX experiment with two animals, encompassing 622 nodules. Each of the 256 genotypes are colour-coded. As in Fig. 4a. b, Spatially resolved nodule annotation. For all 11 samples used in this study. As in Fig. 3a. c, Tumour genotypes. Scaled (p¹⁰) estimated plasmid probabilities per nodule. As in Fig. 5.

Extended Data Fig. 9 Overview of genotypes and phenotype-binarization for all nodules across 11 samples.

As in Fig. 5. Top panel: Scaled plasmid probabilities and per-nodule phenotype binarization as used in Fig. 6 (Methods). n = 622 total nodules from 11 samples used in this study from a single RUBIX experiment with two animals.

Extended Data Fig. 10 Genotype–phenotype relations as evidenced by GLM.

a, Tumour-intrinsic genotype–phenotype relations. A generalized linear model (GLM) predicts gene expression signals at each 10X Visium spot, using estimated probabilities of perturbation presence (Methods). Data are presented as feature coefficients shown as mean and error bars depict 3σ confidence intervals. Feature coefficients indicate associations between gene expression and perturbations for representative transcripts of four tumour-intrinsic phenotypes. b, TME-related genotype–phenotype relations. As in a for representative transcripts of two exemplary TME phenotypes. c, GLM-inferred genotype–phenotype associations. Top: heatmap of 1,283 genes with at least one significant (3σ) feature weight, ordered by 1D UMAP embedding. Bottom: Detailed views of four representative clusters linked to marker genes of known phenotypic groups. Data is derived from 324 nodules across 6 topographically separated regions used for 10X Visium from a single RUBIX experiment with two animals.

Source data

Supplementary information

Supplementary Information

Supplementary Figs. 1–18.

Reporting Summary

Peer Review File

Source data

Source Data Fig. 7

Statistical source data. Exact numbers (n counts) for HCC and CCA nodules in two liver sections from four animals each from three groups analysed.

Source Data Extended Data Fig. 10

Heatmap source data. Feature coefficients derived from the GLM.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

About this article

Cite this article

Breinig, M., Lomakin, A., Heidari, E. et al. Integrated in vivo combinatorial functional genomics and spatial transcriptomics of tumours to decode genotype-to-phenotype relationships. Nat. Biomed. Eng (2025). https://doi.org/10.1038/s41551-025-01437-1

Download citation

Received: 18 October 2024
Accepted: 22 May 2025
Published: 28 July 2025
DOI: https://doi.org/10.1038/s41551-025-01437-1

This article is cited by

Integrated in vivo combinatorial functional genomics and spatial transcriptomics of tumours to decode genotype-to-phenotype relationships
- Marco Breinig
- Artem Lomakin
- Darjus F. Tschaharganeh
Nature Biomedical Engineering (2025)

Subjects

Abstract

Similar content being viewed by others

Main

Results

A framework to spatially map engineered tumour heterogeneity

RUBIX generates higher-order combinatorial perturbations

PERTURB-CAST hijacks probe-based transcript capture for barcode mapping

Spatial perturbation mapping across tissue

Comparative analyses in hundreds of coexisting cancer clones

Cancer-driving co-dependencies and potential context dependencies

Phenotypic landscapes of engineered tumour heterogeneity

Stratification of coexisting liver tumour subtypes

Tumour–stroma and tumour–immune cell connections

Complex genotype–phenotype relationships

Epistatic regulation of CCA development

Discussion

Methods

Animal experiments

10X Visium for FFPE spatial transcriptomics

Economized molecular barcode selection

Molecular cloning

Myc

mtCtnnb1

NICD

VEGFA

shPTEN

shREN

shTRP53

shKMT2C

Histopathology

IHC

Computational data analysis

Genotyping

Barcode expression pre-preprocessing

Bayesian modelling of perturbation probabilities from barcode counts

Perturbation probability model inference

Occurrence of individual perturbation combinations

Co-occurrence ORs and model of second-order interaction effect

Genotype-to-phenotype GLM

GLM inference

Statistical analysis of the leave-one-out experiment

Reading Visium space ranger output into data objects

Publicly available databases of cell-type markers

Data normalization and preprocessing

Gene coexpression networks

Nodule-level expression aggregation

Binarization of nodule phenotypes

Binarization of nodule TME signatures

Phenotype and TME heatmaps

Spatial integration and nodule unification

Inter-nodule differential gene expression

Cell type-informed factor analysis

Reporting summary

Data availability

Code availability

References

Acknowledgements

Funding

Author information

Authors and Affiliations

Contributions

Corresponding authors

Ethics declarations

Competing interests

Peer review

Peer review information

Additional information

Extended data

Supplementary information

Source data

Rights and permissions

About this article

Cite this article

Share this article

This article is cited by

Search

Quick links