Benchmark of cellular deconvolution methods using a multi-assay dataset from postmortem human prefrontal cortex

Abstract

Cellular deconvolution of bulk RNA-sequencing data using single cell/nuclei RNA-seq reference data is an important strategy for estimating cell type composition in heterogeneous tissues, such as the human brain. Here, we generate a multi-assay dataset in postmortem human dorsolateral prefrontal cortex from 22 tissue blocks, including bulk RNA-seq, reference snRNA-seq, and orthogonal measurement of cell type proportions with RNAScope/ImmunoFluorescence. We use this dataset to evaluate six deconvolution algorithms. Bisque and hspe were the most accurate methods. The dataset, as well as the Mean Ratio gene marker finding method, is made available in the DeconvoBuddies R/Bioconductor package.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13059-025-03552-3.

Background

Increasing numbers of bulk RNA-sequencing (RNA-seq) and single cell or nucleus RNA-seq (sc/snRNA-seq) datasets have been generated, sometimes uniformly processed, and publicly shared [1–4]. RNA-seq data historically has been cheaper to generate than sc/snRNA-seq data, leading to a surge in methods that perform cellular deconvolution and estimation of cell type proportions using reference sc/snRNA-seq data [5–11]. Some methods can use these estimated cell type proportions to deconvolve cell type-specific gene expression [12–14] to overcome cellular heterogeneity and identify nuanced gene expression signals that would otherwise be masked in bulk RNA-seq data. Some downstream applications include differential expression analysis adjusting for cell type composition confounders [15], cell type-specific eQTL discovery [16, 17], and quality control assessment of dissections in heterogeneous tissues.

While each new computational deconvolution method typically compares itself against other leading methods, estimated cell type proportions can be widely variable across methods making it difficult for users to select the appropriate algorithm for a given analysis. Several comprehensive benchmarking efforts have been performed by independent groups [17–20] which evaluated deconvolution methods across a wide range of tissues, simulation scenarios, and normalization methods [17–20]. However, the performance rankings of different deconvolution methods have mostly been inconsistent due to several plausible reasons, including (1) the tissue for which they were initially developed did not match the tissue used for evaluation, (2) biases and variability in reference sc/snRNA-seq datasets, (3) variability in the selection of cell type marker genes, (4) differences in cell type heterogeneity in the tissue under study, (5) choice of cell type resolution (fine or broad), (6) factors regarding how the tissue samples were extracted or preserved, (7) differences between the target RNA-seq and reference sc/snRNA-seq data regarding cell fractions profiled and library preparation strategies, and (8) differences in data normalization and processing [17–22].

One main challenge has been the limited availability of “ground truth” or “gold/silver standard” cell type proportions against which deconvolution methods can be benchmarked. In the absence of cell type proportion standards, it has been common to pseudobulk sc/snRNA-seq data or use mixture simulations to generate pseudobulk RNA-seq data, use the same sc/snRNA-seq data as the reference, and compare the deconvolution results against the cell type proportions observed in the sc/snRNA-seq data [18–20, 23, 24]. However, sc/snRNA-seq library preparation protocols have filtering steps that can introduce biases in the estimated cell type proportions [21], limiting their use as “ground truth” references. Immunohistochemistry in tissue sections can generate orthogonal measurements of cell type proportions and has been previously used for benchmarking deconvolution algorithms in brain tissue [17, 25]. Flow cytometry has also served as a gold standard for deconvolution in blood [26]. However, there is a need for more datasets with orthogonal measurements of cell type proportions from the same tissue blocks used for RNA-seq and sc/snRNA-seq [21]. In particular, the human brain is a complex heterogeneous tissue which is typically studied from fresh frozen postmortem samples where it can be difficult to obtain gold standard measurements using immunohistochemistry and flow cytometry [21]. Additional orthogonal cell composition reference data will be useful to detect shortcomings and areas of improvement for computational deconvolution methods across a range of complex tissues.

One common assumption in deconvolution methods is that bulk RNA-seq is uniform; however, there are different types of RNA library preparation types and RNA extraction protocols. Some protocols enrich the cytosolic or nuclear cell fractions [27, 28], while most capture RNAs from the whole cell. In addition to different RNA extraction methodologies, there are several options for RNA-seq library preparation, with two common approaches including poly(A) + enrichment for mRNA profiling [29] and ribosomal RNA depletion via the Ribo-Zero Gold kit [30] for total RNA profiling. These two library preparations show differences such as polyA having a higher exonic mapping rate, RiboZero/RiboZeroGold having a higher intronic mapping rate, and RiboZero/RiboZeroGold capturing a larger diversity of gene biotypes [31–34]. As RNA-seq and sc/snRNA-seq quantify different RNA populations and sc/snRNA-seq unique molecule identifier (UMI) counts are enriched for zeros [35, 36], these assays present different gene count statistical properties and gene biotype quantification differences. Some deconvolution methods already model statistical differences between sequencing assays [6], while others employ normalization methods. Ultimately, the RNA extraction method and library preparation protocol used for bulk RNA-seq can impact the benchmarking results of deconvolution methods [22].

Computational deconvolution did not originate with gene expression data as cellular deconvolution algorithms were originally developed for DNA methylation (DNAm) data, where it is possible to identify CpG sites that are binary markers for a given cell type [37]. In contrast, cell types in sc/snRNA-seq data are defined through cell type marker genes, which can be identified through many different methods [38]. Some methods identify genes with high expression in a target cell type, but these genes may also be expressed in other cell types, making cell type markers identified through sn/scRNA-seq [38–41] noisier than cell type marker CpGs in DNAm data [42, 43]. Improvements in cell type marker gene selection for deconvolution is an active area of development for sn/scRNA-seq data [21]. In contrast to DNAm data, reference feature selection is more challenging in the deconvolution of RNA-seq data and benchmark data can help drive the development of this crucial analytical step.

This study presents a rich multi-assay dataset in postmortem human brain tissue that can be used to rigorously benchmark computational methods for deconvolution of bulk RNA-seq data in heterogeneous tissues. Data from the dorsolateral prefrontal cortex (DLPFC) was generated across multiple donors [39, 40], tissue blocks, and modalities. Single-molecule fluorescent in situ hybridization (smFISH)/immunofluorescence (IF) for six broad cell types and bulk RNA-seq from three RNA extraction protocols and two library types were generated from adjacent tissue sections (Fig. 1A), in addition to previously described snRNA-seq and spatial transcriptomics data from the same tissue blocks [40]. This multimodal dataset from matched tissue blocks is a comprehensive resource for evaluating existing and future deconvolution algorithms in complex tissue with highly organized laminar structure.

Fig. 1.

Experimental design overview and exploration of gene detection in different assays. A Human postmortem brain dorsolateral prefrontal cortex (DLPFC) tissue blocks across the anterior to posterior axis from 10 donors were dissected for a total of 19 tissue blocks, these tissue blocks are a subset of the 30 tissue blocks that were used in a previous spatial transcriptomic study [40]. For each block, sequential slides were cut for different assays while maintaining the same white matter vs gray matter orientation. B snRNA-seq data, generated as part of the same spatial transcriptomic study was collected for 19 tissue blocks [40], from which bulk RNA-seq data was also generated across two library preparations (polyA in purple or RiboZeroGold in gold) and three different RNA extractions targeting different cell fractions: cytosolic (Cyto, light color), whole cell (Total, intermediate color), or nuclear (Nuc, dark color) in this study. C tSNE plot of the reference snRNA-seq data at the broad cell type resolution. D Scatter plot of bulk RNA-seq principal components (PCs) 1 and 2. PC1 is associated with library type and PC2 with RNA extraction method. Colors are the same as groups in B. E Volcano plots for the differential expression analysis between polyA and RiboZeroGold, faceted by RNA extraction method. The colors of the points are the same as B. Horizontal dotted line denotes FDR < 0.05 cutoff, vertical dotted lines are logFC = − 1 and 1. F Volcano plot for the differential expression analysis between Total bulk RNA-seq (point colors same as E) and snRNA-seq (blue points). Annotations are the same as E

In this study, six leading deconvolution algorithms were benchmarked on this multi-assay DLPFC dataset. The methods evaluated were DWLS [5], Bisque [6], MuSiC [7], BayesPrism [8], CIBERSORTx [11], and hspe [9] previously known as dtangle [44]. These algorithms are among the best performers in recent benchmarking studies [17–20, 22] and optimize predictive performance by extension of weighted least squares [5], assay bias correction [6], source bias correction [7], Bayesian methods [8], high collinearity adjustment [9], and machine learning [11]. Additionally, we evaluated several marker gene selection methods, including a new method for cell type marker gene selection called Mean Ratio. Mean Ratio identifies cell type marker genes that are expressed in the target cell type with minimal gene expression in non-target cell types. Accuracy of cell type proportion predictions was evaluated against the orthogonal RNAScope/IF cell type proportion measurements from the same tissue block.

Results

Multi-modal dataset from human DLPFC

To create a multimodal dataset that can assess the performance of cellular deconvolution methods on a variety of RNA-seq conditions, RNA-seq was performed on 19 tissue blocks across the anterior–posterior axis of the dorsolateral prefrontal cortex (DLPFC) from 10 adult neurotypical control donors with different RNA extraction protocols and RNA-seq library preparations types (Fig. 1A, Additional file 1: Fig. S1, Additional file 2: Table S2). Additionally, combined single molecule fluorescent in situ hybridization (smFISH) and immunofluorescence (IF) data using RNAScope/IF technology was generated on consecutive tissue sections to estimate the proportions of six broad cell types as well as the nuclear size and total RNA content for individual cells. Single-cell and spatial transcriptomics data were also collected on these same tissue blocks in a previous study by Huuki-Myers et al. [40] (Fig. 1A, Additional file 1: Fig. S1B).

For bulk RNA-seq data, total and fractionated RNA was extracted from 19 DLPFC tissue blocks using adjacent tissue cryosections. Fractionated RNA contained the nuclear (Nuc, n = 38) and cytoplasmic cellular fractions (Cyto, n = 37 as one sample failed). Total RNA included RNA from whole cells (Total, n = 38). For each aliquot of RNA, two libraries were prepared using either polyA (n = 56) or RiboZeroGold library preparation techniques (n = 57) [29, 30]. After sequencing, alignment, and quality control data processing, 110 RNA-seq samples were included in the study (Fig. 1B, Additional file 1: Fig. S2, Additional file 1: Fig. S3). The sequencing round did not drive large changes in the gene expression, observed in the top 6 PCs (Additional file 1: Fig. S4).

For benchmarking deconvolution algorithms, previously analyzed snRNA-seq data from the same tissue blocks were leveraged, which served as a paired snRNA-seq reference dataset [40] (n = 19, Fig. 1A, Additional file 1: Fig. S1). This snRNA-seq dataset contained gene expression profiles from 56 k nuclei representing seven broad cell type populations in the DLPFC including, astrocytes (Astro), endothelial/mural cells (EndoMural), microglia (Micro), oligodendrocytes (Oligo), oligodendrocyte precursor cells (OPC), excitatory neurons (Excit), and inhibitory (Inhib) neurons (Fig. 1C).

Principal Component Analysis (PCA) of Total and fractionated (Cyto and Nuc) RNA-seq samples showed that there were large differences in gene quantification between the library preparation types and RNA extraction methods. PC1, which explained the largest percent of variation (62.7%) showed a clear split between polyA and RiboZeroGold library types. PC2, the next largest percent of variation (8.9%), showed separation between Total, Cyto, and Nuc RNA extractions (Fig. 1D).

Differential gene quantification between RNA-seq library preparations

The adjacent cryostat sections from each tissue block should have nearly identical gene expression profiles as they can be considered technical replicates. However, they are not pure technical replicates as the RNA library type or RNA extraction changes. DGE analysis between library types or RNA extractions identified many significantly differentially expressed genes (DEGs). To recognize that these DEGs should have similar biological expression levels, we refer to them instead as “differentially quantified genes” (DQGs). When comparing two RNA library combinations, we refer to DQGs that had significantly higher RNA-seq counts in one library combination over the other one as over-quantified genes.

DQGs were identified when comparing polyA against RiboZeroGold RNA library types, and also across different cell fractions (Cyto, Nuc, or Total) for the same RNA library type (Fig. 1E, Additional file 1: Fig. S5, Additional file 2: Table S3, Additional file 2: Table S4). Between library types in the total RNA extraction, out of 21,745 genes, 996 (4.58%) were over-quantified in RiboZeroGold against polyA, and 1005 (4.62%) in polyA against RiboZeroGold (FDR < 0.05, Fig. 1E). Many of the DQGs between library types in the different RNA extractions were shared: 4060 genes were over-quantified in RiboZeroGold and 2776 in polyA for both Nuc and Cyto, possibly due to these fractions originating from the same RNA extraction kit, while overall 751 and 390 DQGs were respectively shared across all RiboZero or polyA libraries (Additional file 1: Fig. S6A). RiboZeroGold and polyA libraries detected similar levels of expressed protein-coding genes and other major gene biotypes with RiboZeroGold detecting more genes from uncommon biotypes (Additional file 1: Fig. S7A). Cyto fractions showed the largest difference between RiboZeroGold and polyA libraries: 1081 more genes were detected, mostly from noncoding gene biotypes such as lncRNAs (Additional file 1: Fig. S7A). Functional enrichment analysis at the cellular component level for the DQGs revealed that ribosomes are more likely to be captured with polyA libraries compared to RiboZeroGold libraries for both Cyto and Nuc RNA fractions (Additional file 1: Fig. S8A). In contrast, the RiboZeroGold libraries were enriched for nucleosome and synaptic membrane cellular components among DQGs (Additional file 1: Fig. S8A). These results support the direct depletion of rRNAs with RiboZeroGold and its expected capability to capture RNA species that are not polyadenylated.

There were also notable differences in gene quantification between RNA extractions in the same library preparation. For RiboZeroGold libraries, the largest difference was between the Cyto and Nuc extractions (1887 DQGs (8.68%), and 2775 DQGs (12.8%), respectively); for polyA libraries, the largest difference was between Total and Cyto extractions (3,269 DQGs (15.0%), and 2639 DQGs (12.1%), respectively, Additional file 1: Fig. S5). Across RNA extractions, under-quantified DQGs in Cyto against Total highly overlapped with the under-quantified DQGs in Cyto against Nuc in both RiboZeroGold and polyA (Additional file 1: Fig. S6B). Functional enrichment analysis of cellular component gene sets for DQGs across pairwise RNA fraction comparisons, particularly Cyto versus Nuc, revealed genes encoding for products working in ribosomes and mitochondria, and neuronal mRNA being more recently transcribed in the nucleus, respectively (Additional file 1: Fig. S8B–C).

Some deconvolution algorithms, such as Bisque [6], take into account the different statistical properties of measuring gene expression with bulk RNA-seq and snRNA-seq assays. We observed large differences in gene expression quantification between the snRNA-seq data (pseudobulked by tissue block) compared to the corresponding Total extraction RNA-seq data (Fig. 1F, Additional file 2: Table S5). Considering gene biotypes, snRNA-seq detected more expressed genes and a larger fraction of lncRNAs than bulk RNA-seq data (Additional file 1: Fig. S7B). Functional enrichment analysis of DQGs identified over-quantification of cytoplasmic-related RNAs in both polyA and RiboZeroGold Total RNA-seq samples against snRNA-seq (Additional file 1: Fig. S8D). This showed that RNA library preparation and assay type have a significant impact on gene expression quantification.

Orthogonal cell type proportion measurement with RNAScope/IF imaging

To estimate cell type proportions of six major cell types in the DLPFC using an orthogonal assay, multiplex single molecule fluorescent in situ hybridization (smFISH) combined with immunofluorescence (IF) was employed using RNAScope/IF technology. Across 21 tissue blocks, two probe combinations were designed to include three cell type markers, a total RNA expression marker gene, AKT3 [45], and a nuclear marker, DAPI. One RNAScope/IF marker combination “Star” included SLC17A7 (marking Excit), TMEM119 (Micro), OLIG2 (OligoOPC, marking both Oligo and OPC), and AKT3. The second combination “Circle” included GFAP (Astro), CLDN5 (Endo), GAD1 (Inhib), and AKT3 (Fig. 2A, Additional file 1: Fig. S1B, Additional file 2: Table S6). Following staining and imaging, cells were segmented and classified with HALO (Indica Labs). Quality control analysis showed “Star” tissue sections (n = 12) had a median of 37,553 cells (range 28,093:53,709), whereas “Circle” (n = 13) tissue sections had a median of 44,835 cells (range 32,425:57,674, “Methods”). Cells were phenotyped based on the expression of cell type marker probes/antibodies using HALO (Indica Labs, Fig. 2B–C, “Methods”, Additional file 1: Fig. S9, Additional file 1: Fig. S10). In a given tissue section, proportions were calculated for each labeled cell type. Across tissue sections, the median proportions for each cell type were Astro = 0.09, Endo = 0.04, Inhib = 0.11, Excit = 0.23, Micro = 0.03, and OligoOPC = 0.12 (Fig. 2D, Additional file 1: Fig. S11, Additional file 2: Table S7). In each combination, a large proportion of “Other” cells were observed. This is expected as we only probe for 3 of the 6 broad cell types in each assay due to multiplexing limitations with RNAscope/IF.

Fig. 2.

Estimated cell type proportions in tissue sections labeled with RNAScope/Immunofluorescence. A Schematic of experimental design for combined single molecule fluorescent in situ hybridization (smFISH) with immunolabeling using RNAScope/immunofluorescence (IF). The “Star” combination of probes/antibodies marked Excitatory Neurons [Excit], Microglia [Micro], and Oligodendrocyte/OPC [OligoOPC] cell types in 13 tissue blocks. The “Circle” combination marked Astrocytes [Astro], Endothelial cells [Endo], and Inhibitory Neuron [Inhib] cell types in 12 tissue blocks. The total RNA expression gene (TREG) [45], AKT3, was also included in each combination to estimate RNA abundance. B Representative fluorescent image of a single field for the “Star” combination showing expression of SLC17A7, TMEM119, OLIG2, AKT3, and the nuclear marker DAPI. C Representative fluorescent image of a single field for the “Circle” combination showing expression of GFAP, CLDN5, and GAD1. D Barplots of estimated cell type proportions from RNAScope/IF data for “Circle” and “Star” experiments. E Scatter plots of cell type proportions estimated from snRNA-seq data (x-axis) vs. RNAScope/IF, annotated with the Pearson correlation (cor), root mean squared error (rmse), and relative rmse (rrmse) against the mean RNAScope/IF proportions

Cell composition derived from snRNA-seq data can be biased by nuclei sorting steps and some cell types can more frequently be filtered out during data quality control. Given the orthogonal measurements provided by RNAScope/IF, we compared cell type proportions measured by RNAScope/IF and snRNA-seq for the same sample (star n = 12, circle n = 11) (Additional file 1: Fig. S1B, Fig. 2E). The two proportion estimates were compared with Pearson correlation (cor) and relative root mean squared error (rrmse) to the mean of the RNAScope/IF-derived proportions. Inhib had the closest proportion values to the snRNA-seq with the highest correlation (0.813) and lowest rrmse (0.242). Other cell types were inconsistent between assays with Endo showing the lowest cor (− 0.414) and Oligo showing the highest rrmse (1.73) (Fig. 2E).

Selection of deconvolution marker genes: mean ratio method

To eliminate noise in deconvolution analyses, cell type marker genes should have cell type-specific expression characterized by high expression in the target cell type and low expression in other cell types (Fig. 3A). There are several currently available methods to identify cell type marker genes [38], such as “1 vs. All” differential expression analysis (1vALL) [41]. However, cell type marker genes identified by this approach may also be expressed in other cell types. Therefore, we propose a new method, called Mean Ratio, to identify marker genes with more robust cell type-specific expression that are better suited for deconvolution analyses. Mean Ratio is defined by calculating the ratio of the mean expression of a gene in a target cell type over the highest mean expression of the non-target cell types. Genes with the highest Mean Ratio above 1 are the best maker genes for the target cell type (Fig. 3B). In this dataset, genes with the highest Mean Ratios often had the highest log fold changes compared to the 1vALL method, whereas the opposite was not true as 1vALL was more permissive of expression in non-target cell types (Fig. 3C). In the DLPFC snRNA-seq data with seven broad cell types [40], the Mean Ratio method identified cell type marker genes with less noisy signal among non-target cell types compared to 1vALL. In contrast, the top marker genes selected by 1vALL showed some expression in non-target cell types (i.e., PLP1, an Oligo marker gene, had some expression in several non-Oligo cell types) (Fig. 3D). Heatmaps of the top sets of marker genes from the two methods showed more cell type-specific expression for marker genes selected by Mean Ratio (Fig. 3E, F). For deconvolution benchmark analyses, a set of 151 marker genes was selected; genes in this set were both in the top 25 genes ranked by the Mean Ratio value for each of the seven cell types and also present in the filtered RNA-seq data (Additional file 2: Table S8, Additional file 1: Fig. S12).

Fig. 3.

Establishing cell type marker genes in snRNA-seq reference data. A Schematic of an ideal deconvolution cell type marker gene heatmap, where the marker genes for a target cell type (rows) only have high expression in the target cell type (columns), and low expression for all other cell types. B Illustration of marker gene selection strategies for an example target cell type, Oligo. (i) 1vALL combines the non-target cell types into one group and identifies differentially expressed genes between the two groups. (ii) Mean Ratio maintains all the cell type groups, finds the ratio between the mean expression of the target cell type, and the highest mean expression from a non-target cell type. C Scatter plots of the Mean Ratio value vs. 1vALL standard log fold change, for all genes by cell type. The top 25 ranked by Mean Ratio are indicated by point color, and the top gene from both methods is annotated with the gene symbol. D Violin plots of the gene expression over cell types for the top gene from 1vALL (top row) and Mean Ratio (bottom row) methods. E Heat map of the top 5 marker genes from 1vALL logFC. Rows are genes, columns are pseudobulked samples by cell type and tissue block. Color is the Z-score of logcount expression scaled by gene. F Heat map (similar to E) of the top 5 Mean Ratio marker genes

As deconvolution methods have mostly assumed that different bulk RNA-seq library preparations and RNA extraction methods are similar, we investigated whether cell type marker genes were among the DQGs in these RNA extraction and library preparation combinations. We found that any of these marker genes were differentially quantified between library preparations (Additional file 1: Fig. S13), or RNA extractions (Additional file 1: Fig. S14). An enrichment analysis for the 151 Mean Ratio top25 cell type marker genes among library type and RNA extraction DQGs, exposed significant over-quantification of Oligo markers in polyA against RiboZeroGold for Total and Nuc RNA extractions. For the rest of the cell types, markers were over-quantified in RiboZeroGold against polyA for Cyto and Nuc extractions (Additional file 1: Fig. S15A). For RNA extraction DQGs, significant enrichment was found just in polyA. Oligo, Excit, and Inhib cell type markers were significantly over-quantified in Total against Cyto extractions in polyA libraries. Inhib and Micro cell type markers were also over-quantified in Nuc against Cyto, and Cyto against Nuc, respectively in polyA libraries (Additional file 1: Fig. S15B).

Benchmark of selected deconvolution methods with mean ratio top25 marker genes

To test the performance of different approaches to computational deconvolution of RNA-seq data (Table 1), six deconvolution methods (DWLS, Bisque, MuSiC, BayesPrism, CIBERSORTx, and hspe) were run on this DLPFC dataset of 110 bulk/fractionated RNA-seq samples [5–9, 11] to estimate proportions of seven broad cell types. The reference snRNA-seq dataset was subset to the Mean Ratio top25 cell type marker genes for deconvolution. Cell type proportion predictions for each sample varied for each method, and over RNA extraction and RNA-seq library type (Additional file 1: Fig. S16). For example in Br2720_mid library combination PolyA_Cyto, the estimated proportion of Excit ranged from a min of 0.163 from DWLS to a max of 0.688 predicted by CIBERSORTx (Additional file 2: Table S9). Cell type proportion predictions between Bisque and hspe were most similar (cor = 0.938, Additional file 1: Fig. S17). Similarly, the correlation was high between MuSiC and DWLS (cor = 0.886, Additional file 1: Fig. S17), likely related to their adjustment on multi-reference bias sources. Bisque and hspe had the highest correlations with the snRNA-seq proportions (cor = 0.743, 0.696, respectively, Additional file 1: Fig. S17). Overall Bisque and hspe produced similar cell type proportion predictions for this DLPFC dataset.

Table 1.

Selected deconvolution methods. The six reference-based deconvolution methods selected for the benchmark analysis rely on different mathematical approaches and marker gene selection strategies. Also noted is the software availability and other benchmark studies where these methods were noted as top performers

Method	Citation	Approach	Marker gene selection	Availability	Top benchmark performance
DWLS (Dampened weighted least-squares)	Tsoucas et al., Nature Comm, 2019 [5]	Weighted least squares	-	R package on CRAN	Cobos et al. [18]
Bisque	Jew et al., Nature Comm, 2020 [6]	Bias correction: Assay	-	R package on GitHub	Dai et al. [17]
MuSiC (Multi-subject single-cell)	Wang et al., Nature Communications, 2019 [7]	Bias correction: Source	Weights Genes	R package GitHub	Jin et al. [20]
BayesPrism	Chu et al., Nature Cancer, 2022 [8]	Bayesian	Pairwise t-test	Webtool R package on GitHub	Hippen et al. [22]
hspe (dtangle) (hybrid-scale proportion estimation)	Hunt and Gagnon-Bartsch, Ann. Appl. Stat. 202 [9, 44]	High collinearity adjustment	Multiple options- default “ratio” 1vALL mean expression ratio	R package on GitHub	Dai et al. [17]
CIBERSORTx	Newman et al., Nat Biotech, 2019 [11]	Machine Learning	Differential Gene expression	Webtool, Docker Image	Jin et al. [20]

To evaluate the accuracy of these deconvolution methods, predicted cell type proportions were compared to measured proportions for six broad cell types from RNAScope/IF data and evaluated by Pearson’s correlation (cor) and root mean squared error (rmse) (Fig. 2D). To match the cell types in the RNAScope/IF data, the predicted proportions for Oligo and OPC were summed to the OligoOPC combined cell type for these calculations. Across all six RNA extraction and library preparation combinations for each tissue block, Bisque had the highest correlation with RNAScope/IF proportions (cor = 0.538, rmse = 0.141), followed closely by hspe (cor = 0.532, rmse = 0.143), and then CIBERSORTx (cor = 0.49, rmse = 0.154, Fig. 4A, Additional file 2: Table S9). Other methods had less accurate performance, including MuSiC (cor = 0.051, rmse = 0.2), BayesPrism (cor = 0.009, rmse = 0.159), and DWLS (cor = 0, rmse = 0.228) (Fig. 4A).

Fig. 4.

Deconvolution methods performance with Mean Ratio marker genes. A Scatter plot of cell type proportions estimated by RNAScope/IF (x-axis) vs. the predicted cell type proportions by the deconvolution methods. Points are colored by the cell type and shaped by the combination of bulk RNA-seq RNA extraction method and library type. The annotation lists the overall Pearson’s correlation (cor) and root mean squared error (rmse). B Correlation (cor) between the predicted proportions by deconvolution methods and the estimated RNAScope/IF proportions across RNA extraction method and library type combinations, point size reflects the rmse value

Over the six RNA extraction and library type combinations, the performance of the six deconvolution algorithms varied when evaluated against RNAScope/IF cell composition data (Fig. 4B), for example, Bisque had correlations against RNAScope/IF ranging from 0.479 to 0.683 (Additional file 1: Fig. S18). While Bisque had the highest correlation with RNAScope/IF proportions across polyA library types for all RNA extraction methods (max cor = 0.683 for Cyto), hspe was the top performer for RiboZeroGold (max cor = 0.549 for Total), although Bisque and hspe had marginal correlation mean differences: 0.051 in polyA and 0.029 in RiboZeroGold (Fig. 4B, Additional file 1: Fig. S18). CIBERSORTx performed similarly well to Bisque and hspe in polyA samples (Fig. 4B, Additional file 1: Fig. S18), with correlation mean differences of 0.0256 against both Bisque and hspe. MuSiC, BayesPrism, and DWLS were the poorest performers across the data types (Fig. 4B, Additional file 1: Fig. S18). Bisque and hspe had the highest accuracy measured by correlation against cell type proportions derived from the RNAScope/IF data. Alternatively, evaluating the proportion predictions with Spearman’s correlation ranked Bisque as the top method across all library types and RNA extractions (Additional file 1: Fig. S19A). Overall the ranking of deconvolution methods was similar across the two correlation measures. Given the discrepancy between RNAScope/IF and snRNA-seq proportions for astrocytes (Fig. 2E), likely due to the difficulty of segmenting astrocytes due to their complex morphology, we evaluated deconvolution results without astrocytes. We found that the correlation of all methods improves against RNAScope/IF and the overall ranking of methods is mostly invariant, with the exception of hspe slightly outperforming Bisque (Additional file 1: Fig. S20).

Each cell fraction across the six RNA library combinations from the same tissue blocks is expected to be consistent. As an example to evaluate the cell fraction variation, we combined the Excit and Inhib neuron proportions and observed that some methods predict more variance between the neuronal fraction of samples prepared with different library types and RNA extractions from the same tissue block (Additional file 1: Fig. S21A). The relative standard deviation (RSD or coefficient of variation [CV]) of the neuronal proportion per tissue block samples was computed for each method, with CIBERSORTx showing the highest median RSD (Additional file 1: Fig. S21B). Globally, more stable neuronal predictions per sample were observed within each tissue block with Bisque and hspe (Additional file 1: Fig. S21B).

Given the observed overlap between polyA against RiboZeroGold DQGs and oligodendrocyte marker genes (Additional file 1: Fig. S15A), we evaluated the consistency of deconvolution results across RNA library types. This inspection of the estimated oligodendrocyte proportions between polyA and RiboZeroGold RNA library types across deconvolution methods and RNA extractions revealed that DWLS and MuSiC estimated a higher proportion of oligodendrocytes in polyA and had the largest mean differences for the total RNA extraction, 0.181 and 0.167 respectively (Additional file 1: Fig. S22), whereas CIBERTSORTx had the opposite result, and the lowest mean difference − 0.076 for Cyto. For DWLS and MuSiC, inconsistencies could be due to the over-quantification of a significant number (up to 15) of Mean Ratio top25 oligodendrocyte marker genes in polyA (Additional file 1: Fig. S13, Additional file 1: Fig. S15).

Benchmark of deconvolution methods with different gene sets

The selection of cell type marker genes can have important effects on deconvolution results and we thus evaluated the performance of deconvolution methods with other marker gene sets beyond the Mean Ratio top25 described so far. Deconvolution methods were also run with the “Full” set of common genes between the snRNA-seq and bulk RNAseq data (defaulting to each method built-in marker selection if present) as well as three other marker gene selection variations: the top 25 genes ranked by 1vALL, Mean Ratio sets “over 2” (for each cell type, all genes with Mean Ratio > 2), and Mean Ratio “MAD3” (all genes three median absolute deviations larger than median for genes with a MeanRatio > 1 for each cell type) (Methods: Marker Genes, Additional file 2: Table S8). We also tested sets of 10% to 100% of Highly Variable Genes (HVGs) in the snRNA-seq data (Methods: Highly Variable Genes).

Deconvolution methods had varying and unpredictable performance when using different sets of genes: marker genes (Fig. 5) and increasing HVGs sets (Additional file 1: Fig. S23). Broadly, Bisque, hspe, and CIBERSORTx had the top correlation values compared to the RNAScope/IF proportions across the different marker sets. Bisque was the most stable across the different marker gene sets (Fig. 5). For Bisque the top correlation was with MeanRatio top25 (cor = 0.538) and the lowest marker gene set was MeanRatio over2 (cor = 0.504, Fig. 5A, Additional file 1: Fig. S24, Additional file 1: Fig. S25) with the lowest overall input set being HVG70 (cor = 0.504, Additional file 1: Fig. S23). hspe was more sensitive to marker selection as evident by the range of correlation when compared to RNAScope/IF proportions with different gene set inputs (max cor: 0.596 with MeanRatio over2, min cor: − 0.11 with HVG100, Fig. 5A, Additional file 1: Fig. S23, Additional file 1: Fig. S24, Additional file 1: Fig. S26). The performance of BayesPrism substantially improved with the Full marker set (Fig. 5A, Additional file 1: Fig. S23, Additional file 1: Fig. S24). The best choice of marker gene set is dependent on the selected deconvolution method, for example, MuSiC as evaluated by correlation is sensitive to the set of marker genes with 1vAll top25 producing the most accurate results. For most methods, the relationship between the number of marker genes and correlation or rmse is not linear (Additional file 1: Fig. S27), and accuracy cannot be simply predicted from the number of genes. Overall, Bisque was the most stable when correlation and rmse were jointly considered across library types and RNA extractions (Fig. 5B). These results are overall robust to the choice of correlation estimator (Additional file 1: Fig. S19B).

Fig. 5.

Deconvolution methods performance with various marker sets. A Scatter plot of cell type proportions estimated by RNAScope/IF (x-axis) vs. the predicted cell type proportions by the deconvolution methods, for 1vALL top25 marker genes and the full set of common genes. See Additional file 1: Fig. S24 for the Mean Ratio MAD3 or MeanRatio over2 results. Points are colored by the cell type and shaped by the combination of bulk RNA-seq RNA extraction method and library type. The annotation lists the overall Pearson’s correlation (cor) and root mean squared error (rmse). B Scatter Plot between the cor and rmse values for cell type proportion predictions evaluated by bulk RNA-seq RNA extraction method and library type (shape), for all five gene sets evaluated (point color). For BayesPrism, all three MeanRatio marker gene sets resulted in identical cor and rmse values. Related to Additional file 1: Fig. S23

In terms of computation efficiency for deconvoluting this data, Bisque was the fastest method running in only a few minutes, and hspe was the least memory intensive, requiring less than 5 GB (Additional file 1: Fig. S28).

Changing cell type proportions in reference dataset

In heterogeneous tissues, cell types are not present in equal proportions and this is reflected in snRNA-seq data, despite biases this assay introduces in cell composition estimation. Within the DLPFC snRNA-seq dataset, different cell types have different proportions (most common cell type: Excit 0.44, least common: Micro 0.03; Methods: Cell Type Proportion Calculation). These cell type proportions can be different in other snRNA-seq datasets for the same tissue type, as shown in two DLPFC datasets among many other examples [46, 47]. For a deconvolution method to be reliable, it should be robust to variability in cell composition in the reference dataset. To test the sensitivity of the top two deconvolution methods, Bisque and hspe, to different cell type proportions in the reference snRNA-seq dataset, simulations were run downsampling snRNA-seq data to an equal number of nuclei across the seven cell types and observing how it impacted the deconvolution results (Methods: Equal Proportion Reference Simulation). Across one thousand simulations in which different random subsets of the snRNA-seq nuclei were used as a reference to predict cell type proportions, hspe produced more variation in estimated proportions than Bisque (Additional file 1: Fig. S29). The variability across random subsets can be compressed by computing the mean estimated proportions across the 1000 random subsets. Comparing the results with the full data (Additional file 1: Fig. S16) against the mean estimated proportions across the random subsets (Additional file 1: Fig. S30A) showed that hspe was less influenced than Bisque by changes in cell composition on the input snRNA-seq data. hspe had a correlation of 0.936 with the full data estimated proportions, compared to a correlation of 0.343 for Bisque (Additional file 1: Fig. S30B). Moreover, the resulting mean estimated proportions across the random subsets had a very low correlation with the RNAScope/IF data for Bisque (cor = − 0.018), while hspe maintained a similar correlation to prior results with the non-downsampled snRNA-seq dataset (cor = 0.465) (Additional file 1: Fig. S30C). This suggests that while each random subset was more variable in hspe than Bisque, when summarizing the results by taking the mean across the 1000 random subsets, Bisque was more sensitive than hspe to changes in cell type proportions in the snRNA-seq input.

Considering cell size in deconvolution

Differences in cell sizes across cell types may be an important factor in accurate deconvolution [21, 48]. MuSiC has an option to supply a cell size value for each cell type [7]. To test the impact on MuSiC’s performance with this feature, three cell size metrics from the RNAScope/IF data were utilized: nuclear area, the number of copies of AKT3 (a relative measure of total RNA expression) [45], and the product of the two values (Additional file 2: Table S10). These metrics highlight the larger size and increased RNA content of neuronal vs. glial cell types (Additional file 1: Fig. S31A–C). MuSiC results were the most accurate when the nuclear area (Additional file 1: Fig. S31A) was supplied as the cell size metric with an overall correlation of 0.36 with the RNAScope/IF proportions (Additional file 1: Fig. S31D). This improvement in correlation to RNAScope/IF data was consistent across the different library types and RNA extractions (Additional file 1: Fig. S31E, Additional file 2: Table S11).

Benchmark of deconvolution methods with other datasets

Non-paired snRNA-seq reference data

Having paired reference snRNA-seq and bulk RNA-seq data is ideal for deconvolution analysis, but not the typical scenario. To test how the top performing methods performed with non-paired snRNA-seq reference datasets, Bisque and hspe were run with two additional snRNA-seq DLPFC datasets: one smaller dataset with less donor diversity (Tran et al.), including 11,183 DLPFC nuclei from 3 neurotypical control donors; and one larger dataset from a study of Alzheimer’s disease (Mathys et al.) including 70,634 DLPFC nuclei from 48 donors, 24 donors diagnosed with Alzheimer’s disease (Fig. 6A–B) [46, 47]. The Mathys et al. snRNA-seq dataset had a similar proportion of broad cell types to the paired snRNA-seq dataset, whereas the Tran et al. snRNA-seq dataset contained more Oligo nuclei and fewer neurons (Fig. 6C). With the Tran et al. snRNA-seq dataset as the reference, Bisque had a much lower overall correlation to RNAScope/IF data (cor = 0.179) than with the paired snRNA-seq dataset (cor = 0.538, both using MeanRatio top25 marker genes). This is because Bisque over-estimated the fraction of Oligos with the Tran et al. reference data (Fig. 6D, E, Additional file 2: Table S12). In contrast, hspe only showed a small decrease in correlation to RNAscope/IF data when Tran et al. was used as the reference snRNA-seq dataset (Fig. 6D–E). With the Mathys et al. dataset as a reference, hspe produced a high correlation (0.626), but also a high rmse (0.214) with a high estimation of Excit neurons (Fig. 6D, E). Bisque had similar correlation and rmse against the RNAScope/IF data using the paired dataset (Fig. 4A) or the Mathys et al. data as input (Fig. 6D), suggesting that deconvolution results with Bisque are more stable for larger reference snRNA-seq datasets (Fig. 6D). This pattern (a) of high correlation in both methods with the Mathys et al. data, (b) slight decrease in correlation for hspe with Tran, and (c) poor correlation with Bisque with Tran et al. input, was consistent across library types and RNA extractions (Fig. 6E, Additional file 1: Fig. S32).

Fig. 6.

Deconvolution methods performance on different reference and bulk datasets. Bar plots comparing the features of the three snRNA-seq datasets used as references in deconvolution, noting: A number of donors and diagnosis, B total number of nuclei, and C overall cell type composition. D Scatter plot of cell type proportions estimated by RNAScope/IF (x-axis) vs. the predicted cell type proportions by hspe and Bisque (columns) and two external reference snRNA-seq datasets (rows). E Correlation (cor) between the predicted proportions by deconvolution methods (color) and the estimated RNAScope/IF proportions across RNA extraction method and library type combinations. Point size reflects the rmse value, line type shows the reference. F Cell type composition barplot showing the mean estimated proportion for the GTEx v8 brain bulk RNA-seq dataset over thirteen brain regions

To expand on the performance of these methods on increasingly large datasets, we utilized the PsychENCODE Common Mind Consortium DLPFC snRNA-seq dataset to test the accuracy of these methods from 3 to all 52 donors [49, 50]. With this input dataset, Bisque’s mean proportions across iterations have a higher correlation with the RNAScope/IF data than hspe, across all tested numbers of donors (Additional file 1: Fig. S33A). Observing performance in individual subset runs, Bisque shows large improvements in consistency with an increasing number of donors, and hspe shows some improvements with more donors and has much more variable performance. For both methods, the median cor and rmse values are relatively flat across increasing donors (Additional file 1: Fig. S33B, C).

Cross-region bulk RNA-seq

To observe trends in deconvolution in more diverse bulk RNA-seq datasets we used the GTEx v8 Brain dataset from recount3 [4, 51], which includes 2,670 samples across 13 brain regions. Deconvolution was performed with Bisque and hspe using Huuki-Myers et al. DLPFC snRNA-seq dataset as the reference. The mean predicted cell type proportions from hspe were more varied than those from Bisque across the different brain regions (Fig. 6F). Brain regions are expected to have a large variability in cell composition, such as the cerebellum having a large proportion of inhibitory neurons [52], with both hspe and Bisque results matching the expected higher proportion of inhibitory neurons. Other expected differences are more subtle, such as the caudate having an increased proportion of inhibitory neurons compared to frontal cortex [53], and both hspe and Bisque capture this expected increase. In the absence of detailed cell composition reference data for all human brain regions, it is challenging to untangle whether hspe or Bisque results are more accurate across expected biological differences. In the meantime, it is a positive result to observe variation in deconvolution results across brain regions, which showcases the sensitivity to variation in the target bulk RNA-seq data and how this is balanced against variation in the reference snRNA-seq data.

Discussion

This study presents a comprehensive reference dataset across bulk RNA-seq, snRNA-seq, and RNAScope/IF data from the DLPFC from postmortem human brain samples that can be used for multi-assay data integration (Fig. 1A). In particular, this dataset was used to benchmark the performance of RNA-seq deconvolution algorithms based on reference snRNA-seq data and address challenges when computationally deconvolving heterogeneous tissue [21]. A unique aspect of this study is that RNAScope/IF is used to label the six broad cell types of the DLPFC providing estimates of cell type proportions that circumvent some of the pitfalls of measuring cell type proportions with snRNA-seq, which are driven by flow sorting and other quality control steps [21]. Of the six methods tested, the cell type proportions estimated by Bisque and hspe were the most correlated with the RNAScope/IF proportions. Marker genes selected by the new method, Mean Ratio, led to more accurate deconvolution. These results were consistent across different types of bulk RNA-seq samples. Additional testing revealed Bisque’s accuracy is negatively impacted when using small reference datasets or biased cell type proportion inputs.

Most RNA-seq deconvolution methods and benchmark studies have overlooked the fact that not all bulk RNA-seq datasets are generated using the same technologies. While most publicly available RNA-seq datasets have been generated with Illumina sequencers, different RNA extraction kits, and RNA-seq library types can be utilized. Comparison of data from different RNA extraction methods and library types showed large differences in gene quantification of ribosomal genes across polyA and RiboZeroGold libraries, as well as more subtle differences such as genes encoding synapse cellular components being depleted in Cyto against both Nuc and Total in polyA. For a deconvolution algorithm to be applicable across diverse RNA-seq datasets, it should perform well across different data types. In this regard, this multi-assay dataset is a useful benchmarking tool offering orthogonal bulk RNA-seq, snRNA-seq, and RNAscope/IF datasets from the same tissue block. Of the deconvolution algorithms that were evaluated, hspe [9] and Bisque [6] performed similarly and were the top two performers, followed by CIBERSORTx [11] (Fig. 4). Across bulk RNA-seq library types, between Bisque and hspe, Bisque was marginally the best for polyA RNA-seq data and conversely, hspe was the best for RiboZeroGold libraries (Fig. 4B). Both hspe and Bisque produced similarly accurate results in polyA and RiboZeroGold libraries, unlike BayesPrism and CIBERSORTx which were more variable across RNA library types (Fig. 4B). While BayesPrism, MuSiC, and DWLS results had a low correlation against RNAScope/IF and snRNA-seq, they have a higher correlation among themselves, particularly among Excit and OligoOPC (Additional file 1: Fig. S17). MuSiC and DWLS are both using weighted least squares, which could explain the similarity in their results (Additional file 1: Fig. S17). Consistent with other independent benchmark findings using human brain data (Table 1) [17], we identified Bisque and hspe (an update on dtangle [44]) as the most accurate deconvolution algorithms and showed that they are robust to RNA-seq library types, RNA extractions, or choice of correlation method for benchmarking the results (Additional file 1: Fig. S19). Fast runtimes and relatively low memory demands should make these methods both accessible choices for researchers (Additional file 1: Fig. S28).

As has been noted previously [17], immunohistochemistry and sc/snRNA-seq proportions do not always match, suggesting that orthogonal datasets often show variation. As expected, differences in cell type proportions were observed between RNAScope/IF and snRNA-seq assays across all cell types. Astrocytes were significantly undercounted while oligodendrocytes were overcounted in snRNA-seq compared to RNAScope/IF (Fig. 2E). As some bulk RNA-seq deconvolution methods have a tendency to infer cell type proportions similar to those observed in the reference snRNA-seq data [6], this discrepancy in cell type proportions between assays likely drove some inaccuracies observed when benchmarking RNA-seq deconvolution methods against RNAScope/IF-derived cell type proportions. Most evaluated methods underestimated the proportion of astrocytes and frequently overestimated the proportion of oligodendrocytes/OPCs (Fig. 4A), matching the discrepancies observed between the snRNA-seq and RNAScope/IF proportions. This is important to keep in mind when choosing a deconvolution computational algorithm as global performance across all cell types could be misaligned with performance for an individual cell type of interest.

It should be acknowledged that there are also some limitations with using RNAScope/IF data as a reference for cell type proportions given the challenges associated with image segmentation and cell type classification in complex tissues. To mitigate these challenges, we performed image segmentation and cell phenotyping with a widely used image analysis software, HALO, and provided our settings to support the reproducibility of our cell type quantifications. As cell type proportions can also be skewed by imperfections in tissue slices arising from technical variability due to cryopreservation, sectioning, slide placement, and fluorescent staining, three experts in human brain microscopy assigned a confidence level to each image based on tissue section morphology and fluorescence background. Sixty percent of images passed these rigorous quality control checks (Additional file 1: Fig. S1B, Methods: RNAScope/IF Data Generation and HALO analysis) were used in comparative analyses.

We note that cell types under investigation were restricted to broad cell types given the limitations of multiplexing in the RNAScope/IF assay. This cell type resolution is similar to those evaluated in prior deconvolution benchmarks (5–10 cell types) [17, 18]. Future studies could investigate rarer cell types and finer cell type resolutions using spatial transcriptomics technologies that support higher multiplexing, such as MERFISH and Xenium [54, 55].

Despite limitations noted for RNAScope/IF-based cell type labeling, sc/snRNA-seq assays have their own challenges, such as biases that occur during the dissociation of tissue and only a partial capture/sampling of all cells affecting the accuracy of cell type proportion estimates [21, 56, 57]. Ignoring the limitations of sc/snRNA-seq protocols and using pseudobulked versions of the data to benchmark deconvolution algorithms can potentially lead to misleading conclusions on the performance of deconvolution computational algorithms. Thus, it is valuable to generate orthogonal measurements across technologies, despite the associated limitations.

An often overlooked challenge for applying computational deconvolution algorithms is the selection of cell type marker genes. There are many statistical methods for finding sc/snRNA-seq cell type marker genes that have different properties [38], with findMarkers() from scran [41] implementing many options which are commonly used in Bioconductor-based analysis workflows [58]. Some of them, such as the 1vAll selection method, do not penalize genes that have high expression in outlier cells among the non-target cell type. The Mean Ratio method, developed here, was designed to select cell type marker genes that are not only more highly expressed in the target cell type but also have the cleanest signal compared to the second highest cell type (Fig. 3). Mean Ratio cell type marker genes can provide more specific inputs to cell type deconvolution algorithms to ultimately improve the accuracy of results. In general, using a small set of marker genes per cell type (one gene per cell type would be the extreme case) is prone to overfitting given the variability in gene expression measurements for the same gene across bulk RNA-seq and sc/snRNA-seq assays. Evaluating the six deconvolution methods with the full set of common genes, and four sets of marker genes with diverse numbers of marker genes per cell type showed that most methods are sensitive to marker selection. Overall Bisque was the most robust to different marker sets, and the MeanRatio top25 gene set best-balanced correlation and rmse in the hspe and Bisque results (Fig. 5), outperforming the version of MeanRatio that had more marker genes per cell type. In addition to choosing the number of marker genes, it is important to note that cell type marker genes can be differentially quantified across RNA-seq library types or RNA extractions (Additional file 1: Fig. S15); Bisque and hspe were more robust than other methods to these effects (Additional file 1: Fig. S22). When using hspe or Bisque, we recommend using MeanRatio top25 for selecting marker genes.

As sc/snRNA-seq assays have matured, the number and scale of publicly available datasets have increased in recent years [1, 2]. Specifically for the human brain, large efforts such as the BRIAN Initiative Cell Census Network (BICCN) and the PsychENCODE Consortium (PEC) have expanded the understanding of cell clusters, states, and types in neurotypical donors as well as those affected by different psychiatric disorders [50, 59, 60]. In some scenarios, it may be more important to have access to a large and diverse sc/snRNA-seq reference dataset for deconvolution. For instance, the leave-one-out cross-validation performance across 8 donors in Jew et al. revealed large performance gains for Bisque when increasing the reference size from 2 to 4 donors [6]. In testing hspe and Bisque on two additional reference snRNA-seq datasets [46, 47], Bisque had poor performance on the 3 donor Tran et al. dataset, while hspe maintained similar metrics, which supports that Bisque performs best with 4 or more donors [6]. Both methods performed well with the Mathys et al. dataset which included Alzheimer’s case donors (Fig. 6D, E). This suggests that it is advantageous to select larger, more diverse snRNA-seq reference datasets for deconvolution.

Another factor that may impact the performance of Bisque is that it seems to be biased to the cell type proportions of the snRNA-seq reference dataset. Downsampling the paired snRNA-seq reference to even cell type proportions was detrimental to the performance of Bisque (Additional file 1: Fig. S30). This suggests that Bisque may not be an appropriate method to choose if the cell type proportions in the reference snRNA-seq data deviate significantly from the expected makeup of the tissue, or are technically biased due to the enrichment of different cell populations during flow sorting steps.

RNAScope/IF can be used to estimate cell size and total RNA expression [45] (Additional file 1: Fig. S31A–C). This is important as cell size has different associations with total RNA expression across cell types [45], which can potentially impact how cell size and total RNA scaling factors are implemented in deconvolution methods. For example, due to differences in cell sizes, it is likely that current cellular deconvolution algorithms recover the fraction of RNA in different cells rather than estimating cell composition directly [21, 48]. For instance, adjusting for mean nuclear area by cell type improved the performance of MuSiC (Additional file 1: Fig. S31D–E). Integrating cell size information into other deconvolution methods may also improve the accuracy of cell type estimations [48, 61]. The RNAScope/IF-derived cell size and RNA total expression data provided here will enable comparisons of the performance of methods that can adjust for or incorporate these variables given that cell size and total RNA present heterogenous relationships across cell types [45].

Benchmarking, in general, is a challenging endeavor as new technological improvements and software implementations can impact accuracy in different ways [21]. Benchmarking studies face the complication that a “ground truth” is most commonly not available or might be biased. This is why many sc/snRNA-seq studies initially pseudobulked sc/snRNA-seq data, deconvolved the pseudobulk data with the same reference data, and evaluated the results against the cell type proportions on the same sc/snRNA-seq reference data. Variability across tissues, or brain regions (Fig. 6F) [10], can provide a qualitative guide on the accuracy of the results when cross-referenced with experts in the expected cell type proportions on a given tissue. Removing a cell type as well as using a thousand pseudobulk mixtures of reference datasets [18] are complimentary evaluation strategies to the ones presented here that can be used to further evaluate the effects of different normalization strategies [18, 19]. Accurate cell fraction estimates are needed [17], as they are the input for downstream analyses such as cell type-specific expression quantitative trait loci (eQTL) analyses and estimating cell type-specific gene expression [12–14]. The dataset from this study enriches the options available for benchmarking computational deconvolution algorithms by providing multi-assay data from adjacent tissue slices on a heterogenous tissue and by investigating the effects from RNA-seq extraction and library type preparations.

Conclusions

Estimation of cell type proportions in bulk RNA-seq data using snRNA-seq reference-based deconvolution methods presents many challenges. Here we provide a resource for addressing these challenges by generating a multi-assay dataset from adjacent tissue sections across a set of tissue blocks. Different bulk RNA-seq library types and RNA extraction kits were surveyed. Broad cell type proportions using RNAScope/IF were generated, and other data types such as cell sizes, total RNA, and spatially-resolved transcriptomics data are available as well. This data-rich study was used to benchmark the performance of six leading deconvolution algorithms in a complex heterogeneous tissue, postmortem human brain. Of the deconvolution algorithms that were evaluated, hspe and Bisque were the top performers across different RNA extraction kits and RNA-seq library preparation types. Different cell type marker gene identification methods were evaluated, including the newly proposed Mean Ratio method that maximizes the difference between the target cell type and the second transcriptionally closest cell population. Interactions between input snRNA-seq dataset properties and sensitivities of deconvolution methods can majorly impact cell composition results. The highly integrated orthogonal datasets generated here are a novel resource for further benchmarking and developing computational deconvolution methods for RNA-seq data.

Methods

Cryosectioning and tissue sample collection of orthogonal datasets

Assays were completed using 22 individual blocks of postmortem human DLPFC tissue collected across anterior (Ant), middle (Mid), and posterior (Post) positions (Additional file 1: Fig. S1). These were a subset of the same tissue blocks used for Visium (10 × Genomics) and 3′ gene expression snRNA-seq assays described in Huuki-Myers et al. [40]. Tissue sections for the majority of assays were cryosectioned on the same day for each run (i.e., a round of 3–4 tissue blocks balanced across anterior–posterior DLPFC axis) to minimize tissue loss that occurs when obtaining a flat cutting face on the tissue block (Fig. 1A, Additional file 1: Fig. S1A). In a Leica 3050 cryostat, blocks were allowed to equilibrate for 30 min prior to mounting with an optimal cutting temperature (OCT) medium. Excess OCT was scored from each side of the block using a chilled razor blade to minimize interference of OCT in RNA extraction. Each block was trimmed to achieve a flat cutting face then several 10-μm serial sections were collected for RNAScope/immunofluorescence (IF) assays across the 3–4 tissue blocks in that round. In particular, eight slides containing 4 tissue sections each were collected per round, with one tissue section from each block on a given slide. Following collection of serial sections for RNAScope/IF, the cutting thickness was adjusted to 100 μm, and ten serial sections (~ 1 mm of tissue) were collected for single nucleus RNA sequencing (snRNA-seq). Processing and analysis of snRNA-seq data was reported in Huuki-Myers et al. [40]. Immediately following tissue collection for snRNA-seq, six 100 μm serial sections (~ 600 μm tissue) were collected for bulk RNA extraction. Six additional 100 μm serial sections were collected for fractionated (nuclear/cytosolic) RNA extraction (Additional file 1: Fig. S1B). Collected tissue was stored in Eppendorf tubes at − 80 °C until use.

During data generation, we took precautions to minimize batch effects. All assays were performed by the same experimenter. Furthermore, tissue for all assays was cut in the same cryosectioning session. For data collection, donors were divided into 5 sex balanced groups (4 batches of 4, 1 batch of 3, n = 19). These batches were used during cryosectioning, snRNAseq, and RNAscope/IF experiments (sections for each batch were mounted on the same slide).

For bulk experiments, these groups were merged and run in 2 batches, which were inherently balanced due to the experimental design described above. It should also be noted that nuclear and cytoplasmic bulk RNA fractions are run on the same column within the same extraction. For the bulk RNA-seq, the sequencing round did not drive large changes in the gene expression (Additional file 1: Fig. S4).

Total RNA extraction and sequencing

Total RNA was extracted from tissue aliquots (2 extractions per block for 19 tissue blocks, n = 38, Additional file 1: Fig. S1B) using the Qiagen RNeasy mini kit (RNeasy Mini Kit, Cat No. 74104, Qiagen, Hilden, Germany). A modified version of Qiagen’s “Purification of Total RNA from Animal Tissues” protocol from the 2020 version of the RNeasy Mini Handbook was used. Briefly, cryosections were homogenized via a wide-bore pipette in 0.7 mL of TRIzol. Next, 0.14 mL of chloroform was added, and the aqueous phase of the gradient was removed and transferred to a new tube. An equal volume of 70% ethanol was added, and then the mixture was put onto an RNeasy mini-column. At this point, RNA was extracted according to the manufacturer’s instructions with DNAse digestion treatment (RNAse-Free DNase Set, Cat No. 79254, Qiagent, Hilden, Germany). RNA quantity was measured using a Qubit 4 fluorometer (Qubit 4 fluorometer; Qubit dnDNA HS Assay Kit, Cat No. Q32854 Invitrogen, Eugene, OR, USA). RNA quality was assessed using an Agilent RNA Nano kit on a BioAnalyzer instrument (RNA 6000 Nano Kit, Agilent, Santa Clara, CA, USA). Libraries were subsequently prepared and sequenced at Psomagen. For each sample, 100–500 ng of RNA from the same tube was used to prepare a “RiboZeroGold” library with the TruSeq Stranded Total RNA with Ribo-Zero Gold Library Prep kit (Illumina) and “PolyA” library with TruSeq Stranded mRNA Library Prep Kit (Illumina) according to manufacturer’s instructions. Libraries were sequenced on an Illumina Novaseq 6000 targeting 80 million reads per sample. ERCC spike-in sequences were included in all samples except the initial pilot round (n = 24).

Cytoplasmic/nuclear RNA extraction

Fractionated RNA extraction was performed on tissue aliquots using the Cytoplasmic and Nuclear RNA Purification kit (Norgen Biotek, Cat. No. 21000, ON, Canada) (Cyto n = 38, Nuc n = 37, Additional file 1: Fig. S1B) according to the “Animal Tissues” protocol in the manufacturer’s manual (PI21000-19, section B). Briefly, reagent J was added to the tissue, which was homogenized via a wide-bore pipette. Lysate was spun resulting in a supernatant and a pellet. The supernatant was removed and used for the cytoplasmic fraction, and the pellet was retained for the nuclear fraction. For cytoplasmic RNA purification, buffer SK and 100% ethanol were added to the supernatant, which was then transferred to a spin column, and centrifuged. Flow through was discarded and on-column DNA removal was completed (RNase-Free DNase I Kit, Cat No. 25710, Norgen Biotek, ON, Canada). For nuclear RNA purification, the pellet was resuspended in buffer SK and 100% ethanol was added. Lysate was then passed through a 25 gauge needle five times and added to a different spin column. Following centrifugation, the flow through was discarded. For both fractions, the spin columns were washed twice with Wash Solution A before drying the membranes. Cytoplasmic and nuclear RNA were eluted from each column in Elution Buffer E. Both fractions of RNA were stored at − 80 °C. RNA quality and quantity were measured as described above for total RNA extraction. RiboZeroGold and PolyA libraries were subsequently prepared and sequenced at Psomagen as described above. In all, 113 RNA-seq samples were generated (19 * 2 * 3 = 114 minus one) as a sample failed during library preparation due to an insufficient amount of starting material (Additional file 1: Fig. S1B).

Bulk RNA-seq data processing and quality control

FASTQ files were aligned to Gencode v40 using SPEAQeasy (version 712ad37) [62]. The settings were: --sample "paired",--reference "hg38", --strand "reverse", --strand_mode "accept", and --ercc. This resulted in 4.6 to 155.9 million reads mapped per RNA-seq sample (median 84.2, mean 89.9) with an overall mapping rate (overallMapRate) of 0.2148 to 0.9910 per sample (median 0.9737, mean 0.9586). The dataset contained 61,544 genes.

Sample quality was evaluated to exclude samples with low concordMapRate, numMapped, numReads, overallMapRate, totalAssignedGene, and totalMapped or high mitoRate. See https://research.libd.org/SPEAQeasy/outputs.html#quality-metrics for the definition of these QC metrics. Samples were classified as “drop” (n = 2), “warn”(n = 10), or “pass” (n = 101) based on their relationship to cutoffs for each metric determined by 3 median absolute deviations (MADs) from the median value for either polyA or RiboZeroGold library type samples as calculated using isOutlier() from scran [41]. All SPEAQeasy metrics and QC classifications are available (Additional file 2: Table S2).

Two samples were flagged to drop: 2107UNHS-0293_Br2720_Mid_Nuc (low concordMapRate, numMapped, overallMapRate, totalAssignedGene), and AN00000904_Br2743_Ant_Cyto (low numMapped, numReads, and totalMapped) and excluded from downstream analysis (Additional file 1: Fig. S1B, Additional file 1: Fig. S2).

Bulk RNA-seq dimension reduction and expression filtering

Principal component analysis (PCA) was performed on the n = 111 samples that passed sequencing quality control checks. PCs were computed using prcomp()on log2(RPKM + 1) gene expression values, filtered for mean RPKM > 0.1. One sample AN00000906_Br8492_Mid_Nuc, previously classified as “warn” based on QC metrics, was identified as an outlier for PC2 and PC5 (Additional file 1: Fig. S3). This sample was excluded from downstream analysis, bringing the final number of bulk RNA-seq samples to n = 110 (Fig. 1B, Additional file 1: Fig. S1B). Genes with low expression were excluded from the dataset with expression_cutoff(log2(RPKM + 1))from jaffelab v0.99.32 (https://github.com/LieberInstitute/jaffelab); 21,745 genes remained after filtering. The PCA analysis was repeated on the filtered data set (Fig. 1D).

Single nucleus RNA-seq data

Single nucleus RNA-seq data collection and analysis from these same tissue blocks (n = 19) is described in Huuki-Myers et al. [40] (Fig. 1B, Additional file 1: Fig. S1B). Only “broad” cell type resolution was considered in this study (Fig. 1C).

Biotypes of expressed genes in bulk libraries and snRNA-seq data

The expressed genes in bulk samples for the different library combinations were obtained from the bulk RNA-seq dataset already filtered for lowly-expressed genes (21,745 kept genes in total) by recovering genes with non-zero counts across all the samples for each of the six library preparations. For snRNA-seq, all genes in the dataset were included (29,962 total genes) as all passed previous filtering steps [40]. The gene biotypes in bulk data were obtained from Gencode v40 annotation and those for snRNA-seq from the 10 × Genomics annotation for the human genome reference GRCh38 (Gencode v32/Ensembl 98) version 2020-A (Additional file 1: Fig. S7).

Bulk RNA-seq differential expression to identify DQGs

Differential gene expression (DGE) was performed between the two library types (PolyA vs. RiboZeroGold) for the three RNA extraction protocols (Cyto, Total, and Nuc) (Fig. 1E), as well as between the RNA extractions (ex. Total vs. Nuc) for both library types (Additional file 1: Fig. S5). DREAM was applied to leverage the increased power and reduction of false positives in DGE in experiments with multiple samples per donor, as in this experiment’s design [63]. DGE was performed with calcNormFactors() from edgeR v3.42.4 [64], voomWithDreamWeights() and dream() from variancePartition v1.30.2 [63], and eBayes() and topTable() from limma v3.56.2 [65]. To test between library types, samples were separated by RNA extraction, the model used was ~ library_type + (1|BrNum) + mitoRate + rRNA_rate + totalAssignedGene, where BrNum is the donor identifier. To test between RNA extractions samples were separated by RNA library type and compared in a pairwise fashion between the three extractions (Total vs. Cyto, Total vs. Nuc, and Cyto vs. Nuc), the model used was ~ library_prep + (1|BrNum) + mitoRate + rRNA_rate + totalAssignedGene. Genes with an FDR < 0.05 from these analyses are the differentially quantified genes (DQGs).

Bulk vs. snRNA-seq differential expression

To explore gene expression quantification differences between snRNA-seq and bulk RNA-seq data, DGE with DREAM was performed (same methodology above) between the bulk RNA-seq samples (polyA and RiboZeroGold, Total RNA extraction) [63], and the corresponding pseudobulked snRNA-seq samples (Fig. 1F). The snRNA-seq samples were pseudobulked with registration_pseudobulk(var_registration = "BrNum", var_sample_id = "Sample") from spatialLIBD v1.12.0 [66]. The model used was ~ data_type + (1|BrNum)where data_type was bulk or snRNA-seq.QC metrics could not be accurately computed for the pseudobulked snRNA-seq samples and were excluded from this DGE analysis.

Cellular component GO enrichment analysis among DQGs

To assess the significant enrichment of cellular component (CC) gene sets annotated in the Gene Ontology (GO) database within the groups of DQGs (Additional file 1: Fig. S8), over-representation analyses (ORA) were implemented with enrichGO() from clusterProfiler v4.10.0 [67], which applies a one-sided version Fisher’s exact test. The gene universes considered for the analysis on clusters of DQGs between library type and RNA extraction, and between Total bulk and snRNA-seq, corresponded to the total expressed genes used as input for DGE analysis in the bulk RNA-seq dataset (21,745 genes), and in the bulk vs snRNA-seq data (17,660 common genes), respectively. The resulting p-values were adjusted with the Benjamini and Hochberg’s method to control the false discovery rate (FDR) [68].

RNAScope/immunofluorescence data generation and HALO analysis

To quantify six broad cell types across tissue sections (n = 16 sections), multiplex single molecule fluorescent in situ hybridization (smFISH) with RNAScope technology (Advanced Cell Diagnostics) was performed in combination with immunofluorescence (IF) using the RNAScope Fluorescent Multiplex Kit v.2, 4-plex Ancillary Kit, and RNA–Protein co-detection ancillary kit (Advanced Cell Diagnostics ACD, Cat No. 323100, 323,120, and 323,180, Newark, CA, USA) as previously described [69] (Additional file 1: Fig. S1). Briefly, the combined RNAScope/IF protocol involved fixing tissue sections in chilled 10% neutral buffered formalin (NBF), dehydrating in a series of graded alcohols, treating with hydrogen peroxide, and incubating overnight with primary antibodies for GFAP (Thermofisher, Cat No.13–0300, Waltham, MA, USA), Claudin 5 (CLDN5) (Thermofisher, Cat No. 35–2500, Waltham, MA, USA), TMEM119 (Sigma0Aldrich, Cat No. HPA051870-100UL, St. Louis, MO, USA), and OLIG2 (R&D systems, Cat No. AF2418-SP, Minneapolis, MN, USA) (Additional file 2: Table S6). Sections were fixed again in 10% NBF, permeabilized with protease IV, and hybridized with probes for AKT3, GAD1, and SLC17A7 (Additional file 2: Table S6; ACD, Cat No. 434211, 404,031-C2, and 415,611-C3, Newark, CA, USA). According to the manufacturer’s instructions, probes were amplified using AMPs 1–3 and labeled with Opal dye 520, 570, 620, or 690, respectively (Additional file 2: Table S6, Akoya Biosciences, Cat No. FP1487001KT, FP1488001KT, FP1495001KT, and FP1497001KT, Marlborough, MA, USA). Antibodies were labeled with appropriate host species secondary antibodies (Additional file 2: Table S6): donkey anti-mouse IgG conjugated to Alexa 488 (Thermofisher, Cat No. A-21202, Waltham, MA, USA), donkey anti-rabbit IgG conjugated to Alexa 555 (Thermofisher, Cat No. A-31572, Waltham, MA, USA), donkey anti-rat IgG conjugated to Alexa 594 (Thermofisher, Cat No. A-21209, Waltham, MA, USA), or donkey anti-goat IgG conjugated to Alexa 647 (Thermofisher, Cat No. A-21447, Waltham, MA, USA). Finally, sections were stained with DAPI and mounted with FluromountG (Southern Biotechnology, Cat No. 0100–01, Birmingham, AL, USA). Slides were imaged using a Polaris slide scanner (Akoya Biosciences, Marlborough, MA, USA) (Fig. 2B). The final QPTIFF files were further pre-processed to generate spectrally unmixed slide image dataset, using Phenochart (Akoya Biosciences, Marlborough, MA, USA), inForm (Akoya Biosciences, Marlborough, MA, USA), and HALO® image analysis platform (Indica labs, Albuquerque, NM, USA), respectively.

As previously described [69], images were analyzed with HALO software (Indica Labs) using the FISH-IF module to quantify (1) the number of cells for each broad cell type in a tissue section, (2) cell size measured by nuclear area, and (3) number of AKT3 puncta per nucleus, with reference to the manufacturer’s guidelines: HALO 3.3 FISH-IF Step-by-Step guide (Indica labs, Version 2.1.4 July 2021) and Digital Quantitative RNAScope Image Analysis Guide (Indica labs). Segmentation was optimized for accurate identification of punctate AKT3 RNAScope/IF signals as well as each cellular object for excitatory neurons (SLC17A7), inhibitory neurons (GAD1), astrocytes (GFAP), microglia (TMEM119), a combined class of oligodendrocytes and oligodendrocyte precursor cells [OligoOPC] (OLIG2), and endothelial/mural cells (Claudin 5: CLDN5), using user-defined size and intensity thresholds. Two RNAScope/IF probe combinations were used: Star (SLC17A7, TMEM119, OLIG2, AKT3, DAPI) and Circle (GFAP, CLDN5, GAD1, AKT3, DAPI, Fig. 2A–C, Additional file 2: Table S6). To estimate RNA abundance in each cell type, AKT3 was included in each combination as a representative total RNA expression gene (TREG) [45].

A thorough visual inspection was performed to verify the quality of the segmentation outputs regarding object size and shape. The same thresholds were applied to a given cell type across all tissue samples, regardless of the donors, including the size thresholds for the nucleus and cytoplasm. Only the copy intensity parameter was adjusted by individual tissue sections to address staining variations of AKT3 transcripts between samples, using their average cell intensity of RNAScope/IF signals. Based on our previous experience with nuclear segmentation in DLPFC [70], nuclear size was set in a range from 5 μm² to 200 μm² to avoid the inclusion of nuclear fragments and debris. Cytoplasmic size was set a dilation of 4.2 μm around the segmented nucleus to include some cytoplasmic transcripts with high confidence of belonging to that cell (i.e., to include an axon hillock for neurons). We aimed to be conservative so as not to capture transcripts from adjacent cells or passing fibers. Nuclear segmentation parameters were defined to capture single cells and split cell aggregates. Beyond these user-defined parameters, which are required to run HALO, phenotype algorithms were optimized based on staining intensity to avoid overcalling and undercalling cell types. Each phenotype algorithm was optimized across several randomly selected areas within multiple donors before finalizing the threshold for that cell type. These phenotypes were utilized in the majority of donors. In some cases, phenotype algorithms had to be adjusted to accommodate variability across donors, such as higher background, intense staining, or changes in cell density. For example, astrocytes were often overcalled in white matter regions, which interfered with the detection of oligodendrocytes, microglia, and endothelial/mural cell signals, leading to false positives. In tissue sections where the default algorithms did not perform optimally, we optimized the algorithms to represent the cell type distribution as accurately as possible.

In the Circle sections, GAD1 and DAPI were classified as IF probes, while GFAP, AKT3, and CLDN5 were classified as FISH probes. In the Star sections, SLC17A7, OLIG2, and DAPI were classified as IF dyes, while AKT3 and TMEM119 were classified as FISH probes. For Circle sections, 7 phenotypes were used: DAPI/AKT3, GAD1/AKT3, GFAP/AKT3, CLDN5/AKT3, GAD1, GFAP, and CLDN5. For STAR sections, 7 phenotypes were used: DAPI/ATK3, OLIG2/AKT3, SLC17A7/AKT3, TMEM119/AKT3, OLIG2, SLC17A7, and TMEM119.

Detailed information on the parameters used for segmentation, phenotyping, and quantification by the HALO algorithms can be accessed through the HALO settings files, which are available on GitHub at https://github.com/LieberInstitute/Human_DLPFC_Deconvolution/tree/main/raw-data/HALO/settings_files [71]. Output files with additional cell measurements are also located on GitHub https://github.com/LieberInstitute/Human_DLPFC_Deconvolution/tree/main/raw-data/HALO [71]. These files can be used to reprocess or reproduce our segmentation results.

Due to the fragility of tissue sections following the RNAScope/IF procedure, the quality of tissue morphology and IF staining for each tissue section was evaluated by three microscopy experts using a scale of Low, Okay, or High. Low samples contained tears, folds, or missing segments of tissue. A subset of these samples also showed poor staining quality and overexposure during imaging leading to inaccurate segmentation. Okay samples contained some imperfections in morphology or staining quality but were overall intact tissue sections. High samples displayed superior morphology and staining quality. For the Star combination, this resulted in 8 Low, 9 Okay, and 3 High; and for Circle combination 9 Low, 3 Okay, and 9 High. Samples with Low overall quality were excluded from the analysis. The final number of tissue sections that passed morphological and staining quality control was n = 12 for Circle and n = 13 for Star combinations (Additional file 1: Fig. S1B, Additional file 1: Fig. S9, Additional file 1: Fig. S10). In the High quality, a total of 1077,136 cells were segmented. Based on the average expected size of cell, cells were excluded if their radius exceeded 5 µm. This excluded 31,539 cells (2.3%). In the filtered dataset, the median number of cells in a tissue section was 40,035 (range 28,093:57,674).

Cell type proportion calculation

RNAScope/IF cell type proportions were calculated by dividing the number of nuclei for a given cell type by the total number of nuclei segmented in that tissue section (circle = 32,425:57,674, star 28,093:53,709) (Fig. 2D, Additional file 2: Table S7). Cell type proportions were calculated the same way for snRNA-seq data (Fig. 6C, Additional file 1: Fig. S11). RNAScope/IF proportions were compared to snRNA-seq proportions when a sample had data for both assays (star n = 12, circle n = 11) with Pearson correlation and root mean squared error (rmse) from Metrics v0.1.4 [72], as well as relative-rmse to the RNAScope/IF proportions (rrmse: rmse/mean( RNAScope/IF prop)) (Fig. 2E).

Marker genes

Mean Ratio statistics were calculated with getMeanRatio2() and genes were selected with the rank_ratio metric. 1vALL statistics were calculated by findMarkers_1vAll(), both functions are from DeconvoBuddies v0.99.0 [73]. findMarkers_1vAll() is a wrapper function for findMarkers() from scran v1.26.2 [41] that for each cell type performs t-Student tests between a target cell type and all other cell types (Fig. 3, Additional file 1: Fig. S12, Additional file 2: Table S8).

Five sets of genes were selected to benchmark in the deconvolution methods:

1. Full: set of genes common between the bulk and snRNA-seq datasets (17,804 genes). Equivalent to the default marker selection for each method.

2. 1vALL top25: top 25 genes ranked by fold change for each cell type, then filtered to common genes (145 genes)

3. MeanRatio top25: top 25 genes ranked by MeanRatio for each cell type, then filtered to common genes (151 genes)

4. MeanRatio over2: All genes for each cell type with MeanRatio > 2 (557 genes)

5. MeanRatio MAD3: All genes for each cell type with MeanRatio > 3 median absolute deviations (MADs) greater than the median of all MeanRatios > 1 (520 genes)

Gene expression visualization

Violin plots were created with plot_gene_express() from DeconvoBuddies v0.99.0 (Fig. 3D). Heatmaps were plotted with ComplexHeatmap v2.18.0 (Fig. 3E–F, Additional file 1: Fig. S12) [74].

Cell type marker gene enrichment in DQG sets

The positive association between the cell-type specific expression of genes (i.e., if they were cell type markers) and their over-quantification with polyA/RiboZeroGold in Total/Cyto/Nuc RNA fractions (i.e., if they were differentially quantified), was assessed with one-sided Fisher’s exact tests. Specifically, enrichment of the 151 Mean Ratio top25 marker genes per cell type among the DQG sets between library type (Additional file 1: Fig. S15A) and RNA extraction (Additional file 1: Fig. S15B) was tested. The gene universe was defined as all expressed genes in the bulk RNA-seq dataset (21,745 genes). Heatmaps were made with ComplexHeatmap v2.18.0 [74].

Highly variable gene sets

Gene variation in the snRNA-seq data was modeled with scran v1.34.0 [41] modelGeneVar, using the subset of genes common to both the snRNA-seq and bulk data (17 k genes). We selected sets of highly variable genes (HVGs) using getTopHVGs(prop), with a range of prop from 0.1 to 1.0 with 0.1 increments. The resulting HVG sets were 811 to 8113 genes, making up 5 to 46% of the total common genes.

Deconvolution

Deconvolution was completed on the 110 bulk RNA-seq samples, using the snRNA-seq data as the reference at the broad cell type level. Each method was run with the full set of common genes and four sets of selected marker genes (see the “Methods: Marker genes” section) and sets of HVGs (see the “Methods: Highly variable gene sets” section). The following deconvolution methods were applied to the data using the following software packages and functions. Unless noted, default parameters were used (Additional file 1: Fig. S16, Additional file 2: Table S9).

• • DWLS v0.1.0, [5]

  • ◦ Functions: buildSignatureMatrixMAST(), trimData(), solveDampenedWLS
  • ◦ Marker gene and HVGs handling: subset datasets
• • BisqueRNA v1.0.5, [6]

  • ◦ Functions: ReferenceBasedDecomposition(use.overlap = FALSE)
  • ◦ Marker gene & HVGs handling: subset datasets
• • MuSiC v1.0.0, [7]

  • ◦ Functions: music_prop()
  • ◦ Marker genes and HVGs handling: supplied as list to music_prop (marker)
• • BayesPrism v2.1.1, [8]

  • ◦ Functions: cleanup.genes(), select.gene.type(), get.exp.stat(), select.marker(), new.prism(), run.prism()
  • ◦ Marker gene and HVGs handling: subset datasets prior to all functions
• • hspe v0.1, [9] 

  • ◦ Function: hspe()
  • ◦ Marker genes handling: supplied as named list to hspe(marker)
  • HVGs handling: subset dataset hspe()
• • CIBERSORTx [11]

  • ◦ CIBERSORTxFractions docker image (sha256: 9dc06b0a3f58)
  • ◦ Non-default parameters/arguments: --single_cell TRUE and --rmbatchSmode TRUE, which are recommended when deriving a signature matrix from 10X Chromium snRNA-seq reference data; --fraction 0
  • ◦ Marker gene and HVGs handling: subset datasets
  • ◦ Details of data preparation are below

To prepare the snRNA-seq reference data for CIBERSORTx, its raw gene counts were filtered to include just the intersection of the genes measured in the bulk RNA-seq data, snRNA-seq data, and each marker gene set of interest. Next, cells where less than 5% of markers had nonzero expression were dropped to ensure sufficient within-cell variance needed for singular-value decomposition as computed internally.

To test the inclusion of cell sizes in MuSiC, music_prop(cell_size), three different metrics for each cell type were tested from the RNAScope/IF data: the median of the nuclear area, the median number of copies of the total RNA expression gene AKT3, and the median of the product between the nuclear area and the AKT3 copies (Additional file 2: Table S10). Deconvolution was run with the Mean Ratio top25 marker genes (Additional file 1: Fig. S31, Additional file 2: Table S11).

Evaluation of deconvolution results

The accuracy of the deconvolution runs was evaluated against RNAScope/IF quantification for the tissue blocks with available data (n = 12 or 13, see the Methods: “RNAScope/immunofluorescence data generation and HALO analysis” section). To match the cell type level in the RNAScope/IF data, proportions of Oligo and OPC were added to create an OligoOPC combined cell type proportion. Pearson’s correlation (cor) calculated with cor from stats [75], and root mean squared error (rmse), calculated by rmse() from Metrics v0.1.4 [72], were computed for all RNA-seq samples for each method and marker gene set (Figs. 4A and 5A, Additional file 1: Fig. S17, Additional file 1: Fig. S24), and for samples grouped by RNA extraction and library preparation (Additional file 1: Fig. S18, Figs. 4B and 5B, Additional file 1: Fig. S18).

To compare the output from all six deconvolution methods to each other, and to RNAScope/IF plus snRNA-seq proportions, a pairwise scatter plot matrix was created using ggpairs() from GGally v2.2.0 [76] (Additional file 1: Fig. S17). Pairwise scatter plots were also used to compare the different marker sets used in each method (Additional file 1: Fig. S25, Additional file 1: Fig. S26).

Neuronal proportion consistency within tissue blocks

The proportion of inhibitory and excitatory neurons estimated by each deconvolution method in each sample was compared against the estimations for the rest of the samples from the same tissue block to explore how consistent these methods are over different RNA-seq library combinations (Additional file 1: Fig. S21A). The relative standard deviation (RSD or coefficient of variation CV) of these neuron proportions was computed for the samples of each tissue block and for the implementation of the six deconvolution methods. RSD is the ratio of the standard deviation (σ) to the mean (μ); higher values of RSD indicate greater deviation of the samples’ estimated proportions from the mean proportion within a tissue block (Additional file 1: Fig. S21B).

Equal proportion reference simulation

Deconvolution was performed for Bisque and hspe as described in the Methods: “Deconvolution” section, but the paired reference snRNA-seq dataset was subset so there was the same number of nuclei from each cell type (downsampled to match the number of Micro nuclei, the rarest cell type). For Bisque, which involved dropping input cells with no expressed genes, each simulated run involved randomly selecting 1599 cells of each type; this was the largest number possible while preserving an equal count of each cell type. Similarly, for hspe, which involved no dropping of input cells, 1601 cells of each type were randomly selected for each simulated run before performing deconvolution. Deconvolution was performed using the Mean Ratio top25 marker genes and the full bulk RNA-seq data. For both Bisque and hspe, this randomized downsampling and deconvolution was performed 1000 times with different seeds, and the estimated bulk proportions were compared across runs and methods (Additional file 1: Fig. S29). The mean cell type proportion from the 1000 replicates was evaluated against the RNAScope/IF proportion in the same way as the standard deconvolution runs (see the Methods: “Evaluation of deconvolution results” section, Additional file 1: Fig. S30).

Donor subsampling simulation

To evaluate the impact of donor quantity on deconvolution results, deconvolution was performed using Bisque and hspe (Methods: Deconvolution) on various subsets of donors. The 52 neurotypical control donors from the CMC data were used [49, 50], with the.var[‘subclass’] column of the AnnData object matched to broad cell-type resolution, which involved dropping Immune, PC, and SMC cells. Mean Ratio top25 markers were computed and filtered to ensure ratios exceeded 1; this resulted in 25 markers for each cell type with the exception of EndoMural cells, which had 19 markers. Next, data was downsampled to include various numbers of donors between three and the full set of 52. For each of these ten fixed numbers of donors, 100 iterations of deconvolution with both Bisque and hspe were performed, each selecting a different random combination of donors to control for possible selection bias.

External datasets

Two other DLPFC snRNA-seq datasets were used to benchmark methods: Tran et al. with 3 control donors [46] and Mathys et al. with 48 donors, half control and half diagnosed with Alzheimer's disease (Fig. 6A) [47]. Both datasets had cell type annotations similar to the cell type broad level in the paired snRNA-seq dataset, for Tran et al. the rare population of T-cells was excluded. For Mathys et al., pericytes and endothelial cells were combined to match the EndoMural cell type (Fig. 6B, C). Marker genes for both datasets were computed as described in Methods: Marker Genes, the MeanRatio top25 sets were used for deconvolution. Deconvolution was performed with hspe and Bisque as described in the Methods: “Deconvolution” section, with the Tran et al. and Mathys et al. datasets as the snRNA-seq reference (Additional file 2: Table S12). The accuracy was assessed against the RNAScope/IF proportions in the same fashion as the paired dataset (see the Methods: “Evaluation of deconvolution results” section, Fig. 6D–E).

The GTEx v8 brain bulk RNA-seq dataset, 2670 samples across 13 brain regions [51], was accessed with recount3 v1.12.0 [4]. Deconvolution was performed with hspe and Bisque using the paired DLPFC snRNA-seq dataset as the reference, with the MeanRatio top25 marker genes, but instead subset to the 34,057 common genes between the snRNA-seq and GTEx data (see the Methods: “Deconvolution” section, Fig. 6F, Additional file 1: Fig. S32).

Software

Ggplot2 v3.4.3 and earlier versions [77], R versions 4.2, 4.3, and 4.4 [75], and Bioconductor versions 3.14, 3.16, 3.18, and 3.20 were used for the analyses [78].

Authors’ contributions

Conceptualization: LAHM, KDM, SHK, SKM, SCH, KRM, LCT. Methodology: LAHM, SKM, SCH, LCT, KRM. Software: LAHM, NJE, LCT. Formal Analysis: LAHM, NJE, DGP, SKM. Investigation: KDM, SHK, SC, KRM. Resources: JEK, TMH, KRM, LCT. Data Curation: LAHM, KDM, NJE. Writing-original draft: LAHM, KDM, SC, NJE, DGP, KRM, LCT. Writing-review and editing: LAHM, SKM, KRM, LCT. Visualization: LAHM, KDM, DGP. Supervision: LCT, KRM. Project administration: SCH, KRM, LCT. Funding acquisition: SCH, KRM, LCT. All authors read and approved the final manuscript.

Authors’ X handles

X handles: @lahuuki (Louise A. Huuki-Myers); @sanghokwon17 (Sang Ho Kwon); @daianna_glez (Daianna Gonzalez-Padilla); @MadenSean (Sean K. Maden); @stephaniehicks (Stephanie C. Hicks); @kr_maynard (Kristen R. Maynard); @lcolladotor (Leonardo Collado-Torres).

Funding

This project was supported by the Lieber Institute for Brain Development, and National Institute of Mental Health (NIMH USA) grant R01 MH123183. LAHM and LCT were partially supported by NIMH USA grant R01 MH111721. All funding bodies had no role in the design of the study and collection, analysis, and interpretation of data and in writing the manuscript.

Data availability

As documented previously [40], snRNA-seq FASTQ files are available via Globus endpoint 'jhpce#DLPFC_snRNAseq’ endpoint listed at http://research.libd.org/globus as well as the PsychENCODE Knowledge Portal (https://PsychENCODE.synapse.org/) through https://www.synapse.org/#!Synapse:syn51032055/datasets/ [79]. Bulk RNA-seq FASTQ files are available via the Globus endpoint ‘jhpce#humanDeconvolutionBulkRNAseq’ listed at http://research.libd.org/globus as well as via the NIH BioProject under accession PRJNA1086804 and Sequence Read Archive study SRP494701 [80]. HALO raw images are available via the Globus endpoint ‘jhpce#humanDeconvolutionRNAScope’ at http://research.libd.org/globus and Zenodo [81]. The HALO exported setting files and data CSV files are available at https://github.com/LieberInstitute/Human_DLPFC_Deconvolution/tree/main/raw-data/HALO and at Zenodo [71]. The combined HALO output data is available into an R object is available at https://github.com/LieberInstitute/Human_DLPFC_Deconvolution/blob/main/processed-data/03_HALO/halo_all.Rdata and at Zenodo [71]. Source code for this project is available at https://github.com/LieberInstitute/Human_DLPFC_Deconvolution and at Zenodo [71] under the MIT License. The Mean Ratio method is implemented in the R package DeconvoBuddies and is available at https://github.com/LieberInstitute/DeconvoBuddies and at Zenodo [73] under the Artistic-2.0 license.

Declarations

Ethics approval and consent to participate

As documented previously [40], post-mortem human brain tissue was obtained at the time of autopsy with informed consent from the legal next of kin, through the Maryland Department of Health IRB protocol #12–24, and from the Western Michigan University Homer Stryker MD School of Medicine, Department of Pathology, the Department of Pathology, University of North Dakota School of Medicine and Health Sciences, and the County of Santa Clara Medical Examiner-Coroner Office in San Jose, CA, all under the WCG protocol #20111080.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.