Many recent biological studies integrate transcriptomic data from multiple technological platforms to address different biological questions, to balance cost against resolution, or to leverage existing public datasets. To illustrate the challenges of cross-platform transcriptomics and their impact on transcriptional deconvolution and biological interpretation, we reanalysed a dataset from one of our previous studies in which consecutive tissue sections were profiled using different omic methods. We include the following datasets focusing on immune aggregates in brain tumor tissue (Fig. 1A): Multiplex Immunofluorescence (mIF)¹: 10 glioma FFPE blocks were stained for B cells (CD20) and T cells (CD3). A pathologist annotated aggregates of immune cells and counted the number of B and T cells in each immune aggregate to establish ground truth cell counts from proteomics assay. Nanostring GeoMX DSP¹: Consecutive tissue sections from the FFPE blocks were obtained and 28 of the immune aggregates identified in mIF were further profiled using the Nanostring GeoMX DSP platform (WTA gene panel). Immune aggregates from mIF and Nanostring GeoMX areas can directly be mapped to each other, allowing direct comparison of immune cell frequencies between transcriptomics and proteomics assays. 10x Visium (generated in this study): For a subset of glioma (n = 2) and brain metastasis (n = 2) patients, we performed an additional 10x Genomics Visium v2 FFPE WTA assay on consecutive tissue sections. After sequencing, data corresponding to spots within CD20-enriched regions were extracted (approximately 50–100 spots per patient) following visual inspection of CD20 gene expression using the Loupe Browser. These regions were considered immune aggregates.

We first compared frequencies of B and T cells extracted from mIF with those predicted by deconvolution tools applied to the GeoMx DSP data (Fig. 1B). Deconvolution was performed using quanTIseq²⁰, EPIC²¹, MCPcounter²², and SpatialDecon²³. These methods represent a subset of deconvolution methods that reflect a broad range of design choices and underlying mathematical models (Table 1). We initially used each method’s default signatures to reflect common user practice and provide a pragmatic baseline for cross-platform users. Across 28 ROIs with matched mIF data the overall correlations were moderate and varied across tools. SpatialDecon consistently achieved the highest concordance (Pearson correlation coefficient (PCC) = 0.71) between predicted and observed B and T cell proportions, followed closely by quanTIseq (PCC = 0.69).

Next, we compared B and T cell proportions predicted from GeoMx ROIs and Visium spots (Fig. 1C). In this analysis, we were not able to directly match the regions of interest accross assays due to considerable histomorphological differences between the consecutive tissue sections. However, we were able to compare the combined regions of immune aggregates originating from the same tissue block. For this, deconvolution was performed on the ROI level for GeoMx DSP data and on spot level for the Visium spots (immune aggregates delienated manually using B and T cell marker genes). Although the cellular composition of immune aggregates is expected to be comparable within a given patient¹, predicted proportions differed substantially between the two platforms, with the extent of difference in predictions depending on the deconvolution method used (Fig. 1C). Least-square regression methods (EPIC, quanTIseq) systematically estimated higher B and T cell proportions in GeoMx DSP data compared to 10x Visium data. Among the tested tools, SpatialDecon showed smallest differences between mean predicted proportions but also the small variability in predicitons within one platform.

Given the large variability in deconvolution results across computational methods and technological platforms, we next investigated the impact of individual experimental factors that differ between technologies. We began by examining the effect of the gene panel, i.e. the set of genes measured in a transcriptomic assay. Pre-designed targeted gene panels, which capture only a fraction of the whole transcriptome, are often used to balance cost and technological constraints while capturing the most informative genes. Since gene panels vary widely in size and gene selection, they are expected to influence deconvolution outcomes, with the magnitude of this effect likely depending on the method used.

To comprehensively evaluate the choice of the gene panel on the deconvolution results, a systematic data simulation approach was implemented (Fig. 2A). For this, we used single-cell RNA-seq datasets of PBMCs (Dataset A, B, C and D listed in Table S1) and healthy bone marrow (Dataset E, F and G listed in Table S1). We annotated various immune cell types, including B cells, T cells, monocytes, and NK cells, dendritic cells, basophils, neutrophils and progenitors using the Monaco Immune Data^24,25 as a reference and the SingleR package²⁶ for automated cell-type classification (Table S2). We then created pseudo-bulk samples by aggregating counts from B cells, T cells, monocytes, and NK cells that were randomly sampled from the annotated cell pool of each cell type in a dataset to generate a pseudo-bulk-level count matrix.

The genes in this matrix were filtered to match four transcriptomic panels: BD Rhapsody Immune Response (in-house, 434 genes)²⁷, targeted immunology²⁸ (1,056 genes), targeted pan-cancer²⁹ (1,253 genes), Nanostring GeoMx WTA³⁰ (18,816) and 10x Visium Human Transcriptome³¹ (19,902 genes). In addition, we included a panel containing all genes detected in the dataset with existing HUGO gene symbol that were detected in the dataset (22,649 genes in PBMC and 33,514 genes in Healthy Bone Marrow; see Fig. 2B). A full list of genes included in each gene panel is provided in Table S3. We applied four deconvolution tools (quanTIseq, EPIC, MCPcounter, SpatialDecon), using their default signature matrices, as well as SPOTlight³² with Dataset D (Table S1) as signature reference. Predicted cell proportions were then compared to the ground truth across all simulated datasets.

We first evaluated deconvolution performance using PCC between the predicted and actual cell proportions (Fig. 2C). However, compositional data such as cell-type proportions are susceptible to spurious correlations³³. To address this, we used the residuals from the centered log-ratio (CLR) transformation to quantify the differences between actual and predicted cell-type compositions in a pseudo-bulk sample³⁴ (Fig. 2D). Based on these CLR residuals, we calculated the Aitchison distance (AD) for each sample, which measures the dissimilarity between the compositional vectors of actual and predicted cell type proportion³⁵ (Fig. 2E) (see Materials and Methods for details). Across all gene panels and for both PBMC and bone marrow datasets, EPIC followed by SpatialDecon performed most consistently, showing the highest PCC, the lowest absolute CLR residuals and lowest mean AD (Fig. 2C to E, Fig S1). Of note, the default implementation of EPIC does not support monocyte prediction, so monocyte were excluded from all EPIC-related analyses including the generation of simulated data.

The two constrained least square regression methods, quanTIseq and EPIC performed well on large panels. Interestingly, both tools performed best with the 10x Visium WTA panel, and less well with the Nanostring WTA panel and the complete WTA panel (Fig. 2C, D and Fig. S1A, B). This decreased in performance was mainly due to the prediction of a high proportion of “uncharacterized cells”, which are a result of gene expression that cannot be explained by a function of the signature matrix. This highlights that the prediction of uncharacterized cells by constrained least square regression methods is sensitive to gene panel selection.

The performance of quanTIseq decreased for smaller gene panels (Fig. 2C-E). EPIC failed to predict NK cells for the Targeted Pan-cancer gene panel (Fig. 2C-D) These effects likely stems from the different signature matrices used in both tools. In particular, quanTIseq had minimal overlap with some panels, e.g. only 23 genes are common between the targeted pan-cancer and the default quanTIseq signature matrix.

SPOTlight showed poor performance across all gene panels in PBMCs, with PCC values less than 0.5 (Fig. 1D), and negative values of PCC in bone marrow samples (Fig. S1A). The tool tend to overestimate the proportions of progenitors and dendritic cells in the larger gene panels, at the expense of monocytes. Simiarly in the smaller gene panels, it also misattributed proportions of monocytes and T cells (Fig. 2D and Fig. S1).

The comparisons of MCPcounter results with all other tools are more challenging, as it is based on gene signature quantification. To enable comparison, we scaled both the actual proportions and MCPcounter scores to the interval [0,1] for each cell type, thereby preserving the relative distribution across samples (for details see methods). As a result, the normalized scores did not preserve compositionality within a sample and we could therefore not calculate CLR residuals and AD for MCP-counter. MCPcounter failed to detect of small differences in cell type abundance across samples (Fig. 1C).

We next aimed to validate the differences in deconvolution performance across tools and gene panels using real-world cross-platform transcriptomic data. For this, we analysed publicly available 10x Visium spatial transcriptomic data from the same section of human ovarian cancer tissue, for which three different sequencing libraries covering different gene panels were prepared: whole transcriptome (WTA), targeted immunology (TI; 1,056 genes), and targeted pan-cancer panel (TPC; 1,253 genes) (see Table S1). We applied three deconvolution methods (quanTIseq, EPIC, and SpatialDecon) and focussed on the estimated B and T cell proportions across 500 spatial spots (Fig. 2F). Predicted cell proportions for quanTIseq and EPIC varied notably for the same spot, particularly as gene panel size decreased (Fig. 2F,G). For example, quanTIseq predicted significantly higher B and T cell proportions with the ‘Targeted Immunology’ panel than the other two panels (Fig. 2F, G), which is also reflective of the observations made with simulated data. Despite EPIC’s good performance on simulated data, the tool did not achieve notable correlation between predictions on WTA and Targeted immunology panel in the experimental data (Fig. 2G). SpatialDecon showed the greatest variation in B and T cell predictions across spots with the ‘WTA’ panel, but this variation was nearly absent with the ‘Targeted Immunology’ panel (Fig 2F, G). This decreased variance in predicted values with decreasing gene panel size was also observed in the deconvolution of simulated data. With decreasing gene panel size, least square regression techniques, in particular quanTIseq with default signature matrix, tend to overestimate variability between samples, whereas log-normal regression techniques such as SpatialDecon underestimate this variability.

Normalisation is a standard preprocessing step to account for differences in sequencing depth or sample size. In our previous analyses, we applied TPM normalisation (similar to library-size normalization) prior to all deconvolution approaches, independent of a tool’s documentation that recommended an alternative strategy. To ensure this choice did not bias our results, we tested each deconvolution method in combination with three commonly used normalisation strategies: TPM normalisation, Q3 normalisation, and quantile normalisation (Bolstad et al., 2003) (Fig. S3A). Across both the PBMC and bone marrow datasets, SpatialDecon results produced consistently high PCC results across all normalisation approaches (Fig. S3B, C, E, F). In contrast, quanTIseq and EPIC showed reduced performance with Q3 normalisation, while TPM, raw, and quantile normalisation yield comparable outcomes. Overall, normalisation only had marginal influence on performance within each tool, and importantly, the use of TPM normalisation in our prior analyses did not negatively affect predictions (Fig. S3D, G).