scDIFF: automatic cell type annotation using scATAC-seq data by incorporating bulk-level genomic and epigenomic information in a deep diffusive transformer

Abstract

Single-cell assay for transposase-accessible chromatin sequencing (scATAC-seq) provides an opportunity to look deeply into the gene regulation mechanism at the single-cell resolution. With the rapid accumulation of scATAC-seq data, there is an urgent need for automatic cell type annotations using scATAC-seq data. Most existing methods rely on creating artificial gene activity matrix, due to the extreme sparsity nature of the scATAC-seq data. However, these methods fail to exploit the intrinsic information inherent in the scATAC-seq peaks. We present scDIFF, a diffusive transformer-based method that integrates bulk-level genomic and epigenomic information with scATAC-seq data to annotate cell types without creating artificial gene activity matrix. Our scDIFF performed constantly better than state-of-the-art methods on all 46 benchmarking pairs of reference and query datasets across different sequencing platforms. The complete implementation of scDIFF is well-documented and freely available on GitHub (https://github.com/haoyu-wangg/scDIFF).

Introduction

Chromatin accessibility is a measure of the ability that nuclear macromolecules can physically interact with DNA. Chromatin accessibility is crucial for understanding gene regulation mechanisms [1]. Single-cell chromatin accessibility sequencing (scATAC-seq) can measure chromatin opening states of individual cells [2]. It serves as a way to reveal epigenomic heterogeneity among cells [3], which gives us deeper understanding of biological processes, such as cell development and cell differentiation [4], as well as insights into the mechanisms of complex diseases [5].

Cell type annotations using scATAC-seq profiles are of fundamental value for all downstream analysis [6]. Over the last few years, many efforts have been made in performing cell type annotations with unsupervised clustering algorithms [2, 7–9]. Since clustering based methods rely heavily on the knowledge of cell type-related markers, they represent a tedious and complex process that is difficult to generate objective and consistent results [10]. With the rapid accumulation of publicly available high-quality scATAC-seq datasets, it becomes possible to perform automatic, accurate cell type annotations for newly generated datasets using supervised machine learning models [11, 12].

In scRNA-seq data analysis, cell type annotation methods based on supervised learning have become increasingly popular recently, as it has excellent performance with high-quality reference datasets[13, 14]. However, limited work was published in supervised cell type annotations using scATAC-seq data [6, 13, 15]. This is due to the following facts: (i) scATAC-seq has a very low capture rate. Usually, scATAC-seq captures < 10% of accessible chromatin peaks per cell [16]; (ii) accessible regions in scATAC-seq data vary significantly across different datasets [17]; (iii) an scATAC-seq dataset is a very sparse, very large, nearly binary matrix, which makes it very difficult to be handled by general-purpose machine learning methods [17].

Currently, supervised annotations of scATAC-seq data are generally carried out with two steps: (i) an artificial gene activity matrix is created by aggregating accessibility scores around TSS (transcription start site) of each gene [9, 18]; (ii) scRNA-seq analysis methods, such as Seurat [19], SingleR [20], and other machine learning methods [21, 22], are employed to produce cell type annotations from the artificial gene activity matrix. These methods never directly utilize the intrinsic information in the scATAC-seq peaks.

EpiAnno [6] considers the peak-by-cell matrix directly without creating the artificial gene activity matrix. EpiAnno applies a Bayesian neural network directly on the read counts from peaks to capture latent information. However, the peaks in the reference and query datasets do not align well across batches, which affects its annotation performances [23].

Although DNA sequences may vary among different cells due to somatic mutations or genetic variances, SANGO [15], which is designed for cell-type annotations using scATAC-seq datasets, still successfully integrates bulk-level DNA sequences into its cell representations of scATAC-seq. Similarly, epigenomic markers may vary among different cells. Given that epigenetic modifications and other factors coordinately regulate chromatin accessibility [24], it is possible that the bulk-level epigenomic profiles can help in cell type annotations of scATAC-seq data. However, no existing method incorporates epigenomic profiles in cell representations of scATAC-seq data. Although epigenomic profiles have more variance than DNA sequences at single-cell level, it still worth a try to introduce bulk-level histone modification information into the cell representations of scATAC-seq data.

With the help of diffusive transformers, which is a kind of graph neural network model, we developed scDIFF, to incorporate bulk-level epigenomic profiles in cell representations of scATAC-seq data. scDIFF has two main steps: First, it learns a representation for every cell by integrating bulk-level genomic and epigenomic information within the chromatin accessibility peaks using stacked convolutional neural networks. Second, cell representations from both the reference and query datasets are used to construct a similarity graph. A deep diffusive transformer was trained to transfer cell type labels from the reference dataset to the query dataset. Batch effects are eliminated in the process.

We evaluated scDIFF on 46 paired datasets. scDIFF achieves comparable or better cell-type annotation performances in terms of both accuracy and efficiency than other state-of-the-arts methods with a simpler architecture. Althrough ChIP-seq and scATAC-seq datasets from mismatching experimental conditions may restrict scDIFF's cell-type annotation performances, scDIFF is still an efficient and accurate method for automated scATAC-seq cell-type annotations. In addition, scDIFF provides valuable insights and a common framework for integrating multiple types of bulk-level genomics and epigenomics information within cell representations of single-cell sequencing data. Furthermore, the cell-type annotations from scDIFF can be utilized for downstream analysis, providing functional insights into gene regulatory mechanism.

Materials and methods

Dataset curation

We collected 25 scATAC-seq datasets from various tissues (Table S1 in supplementary). The BoneMarrowA, BoneMarrowB, LungA, LungB, LargeIntestineA, LargeIntestineB, WholeBrainA, WholeBrainB, SmallIntestine, Cerebellum, and PreFrontalCortex datasets are extracted from the adult mouse atlas [2], which can be downloaded from the GEO database (GSE111586). These datasets were generated using the sciATAC-seq technology [25] with reference to the mm9 genome. The MosA1, MosA2, MosM1, MosM2, MosP1, and MosP2 datasets are derived from different regions of the secondary motor cortex in the mouse brain [26], which can also be downloaded from the GEO database (GSE126724). These datasets were generated using the snATAC-seq technology [27] with reference to the GRCm38 genome. The MouseBrain(10x) and NormalCortex datasets were both obtained from the 10x Genomics platform (atac_v1_adult_brain_fresh_5k). These two datasets were generated by the 10x sequencing technology with reference to the mm10 genome. The KidneyA-E datasets are subsets extracted from the human adult kidney datasets (GSE185948 in the GEO database) [28]. These datasets were generated using the snATAC-seq technology with reference to the hg38 genome. The peripheral blood mononuclear cell (PBMC) dataset was obtained from the 10x Genomics platform (pbmc_granulocyte_sorted_10k). It was generated by the 10x sequencing technology with reference to the GRCh38 genome. We calculated an imbalance score to indicate the extent of cell type imbalance in each dataset, as follows:

(1)

where C is the number of cell types in the dataset, n_c the number of cells in the c-th cell type, and n the total number of cells in the dataset. We have I ∈ [0, 1]. A higher imbalance score indicates a more imbalanced dataset.

For bulk-level ChIP-seq data, human and mouse H3K4me1, H3K4me3, and H3K27ac ChIP-seq profiles [29] were collected from the ENCODE database [30] for two human (kidney and PBMC) and seven mouse tissues (bone marrow, forebrain, lung, large intestine, intestine, cerebellum, and cortex) (Table S2 in supplementary).

Overview of scDIFF

The scDIFF pipeline is a deep learning-based method for cell type annotations using scATAC-seq data (Fig. 1). It takes aligned peak-by-cell matrices as input to generate informative cell representations. The cell representations integrate single-cell chromatin accessibility, bulk-level genomic sequence features, and bulk-level epigenomic profiles. Subsequently, these cell representations are further refined by the diffusive transformer module, where batch effects are also removed. Finally, an MLP classifier produces the cell type annotations.

Figure 1.

The schematic overview of scDIFF. (A) Preprocessing: We aligned peak-by-cell binary matrices of reference and query datasets. DNA sequences and histone modifications are used as inputs. (B) Cell representation encoder: Each peak was represented by the one-hot encoding of the surrounding L-bp DNA sequences and epigenomic profiles, resulting in an L×(4 + m) input matrix, where m is the number of different types of modification. The encoder, which is supervised by the peak accessibility across all cells, is composed of a pre-convolution module, a CNN-stack, a post-convolution module and a dense module. The learned weights W₀ directly serve as d-dimensional cell representations. (C) Diffusive transformer: The d-dimensional cell representations of reference and query data are used to construct a cell–cell graph. A diffusive transformer (DIFFormer) removes batch effects and produces refined cell representations. An MLP produces cell-type annotations according to the refined cell representations. (D) Applications: scDIFF enables comprehensive analysis of scATAC-seq data, such as visualization, cell-type annotation, and motif enrichment analysis.

Data preprocessing

For scATAC-seq datasets, we binarize the peak-by-cell count matrices of both reference and query datasets [10, 15]. We use ‘1’ to indicate that one or more reads appear within a peak region, and ‘0’ otherwise. As suggested by the literature [8], we only kept peaks with ‘1’ in at least 1% of all cells. We excluded cells that have ‘0’ in all peaks. As suggested by another literature [31], we performed a cell type selection to ensure a fair quantitative comparison between paired datasets. If a cell type appears in the query dataset but not in the reference dataset, it was removed from the query dataset.

We aligned peaks from reference and query datasets (Fig. 1A). For datasets based on different genome assemblies, we applied LiftOver [32] to map the genome coordinates of the query dataset to that of the reference dataset. Overlapping regions between peaks in the reference and in the query datasets were called by BEDTools [33].

For bulk-level data, corresponding genome sequences were obtained from the UCSC Genome Browser [32]. Tissue-specific histone modifications were mapped to the genomic sequence using LiftOver as binary values. If a nucleotide locates within a ChIP-seq peak, it is marked as ‘1’, otherwise ‘0’. When multiple reference datasets from different tissues were merged, we had two strategies to merge tissue-specific histone modifications: (i) the union strategy: if a nucleotide locates within any tissue's ChIP-seq peak, it is marked as ‘1’, otherwise ‘0’; (ii) the intersection strategy: if a nucleotide locates within all tissues' ChIP-seq peaks, it is marked as ‘1’, otherwise ‘0’. All ChIP-seq peaks were directly downloaded from the ENCODE database [30].

For a pair of reference and query datasets, we noted the name of the pair as the concatenation of their names. The former one is the reference dataset, while the later one query dataset. For example, if the pair name is 'BoneMarrowA_BoneMarrowB', BoneMarrowA is the reference dataset, BoneMarrowB the query dataset.

Cell representations

We used L-bp DNA sequences (L = 1344 [15, 34] across all datasets), which are centered at each peak from the peak-by-cell matrix, along with corresponding epigenomic peaks, as inputs to our cell representation encoder (Fig. 1B). The peak-by-cell matrix is a combination of the reference and query datasets, containing a total of n cells, where n_r cells are from the reference dataset and n_q cells are from the query dataset (n = n_r + n_q).

Every peak is used as an instance for training the encoder. The whole encoding process is supervised by the accessibility of every peak across all cells in the dataset pair. For each peak, we concatenate the one-hot encoded DNA sequence with epigenomic peaks before a pre-convolution module. The pre-convolution module consists of a convolution layer followed by a batch normalization, a max-pooling operation and a ReLU activation. After the pre-convolution module, a CNN-stack composed of four consecutive CNN modules was applied. In each CNN module, the input data was replicated to enter two paths. The first path contains a convolution layer, a batch normalization, a ReLU activation, which is followed by another convolution layer, another batch normalization and an efficient channel attention layer. The other path contains a convolution layer and a batch normalization. The output matrices of two paths are summed before entering a max pooling layer and a ReLU activation. The number of output channels of CNN modules in the CNN-stack was set to 64, 128, 256, and 512. The kernel size of all CNN modules is 5. After the CNN-stack, a post-convolution module containing a convolution layer, a batch normalization and a ReLU activation was applied before a global flatten layer and a dense module. The dense module contains a fully-connected layer, a batch normalization, a ReLU activation and a dropout operation, which eventually produces a 64-D compact representation of the peak. The output of the dense module was used to reconstruct peak accessibilities across all cells by a fully-connected layer W₀ (Fig. 1B). We optimize all parameters in this encoder to minimize cross-entropy loss, as follows:

(2)

where n represents the total number of cells, y_i represents the observed accessibility in the i-th cell at that peak, and p_i represents the recovered accessibility at the given peak. We took the weights in W₀ as a 64-D representation for the corresponding cell.

Diffusive transformer

A diffusive transformer (DIFFormer) stack containing 4 consecutive DIFFormer modules [35] was trained on the paired dataset for cell type annotations of the query dataset (Fig. 1C). The DIFFormer can propagate shared information between similar cell representations, which effectively mitigates batch effects across datasets. For readers' convenience, we briefly introduce the DIFFormer in the context of our scDIFF here.

Let x₁, x₂, …, x_n be the cell representations that are generated as in Fig. 1B. We have the initial cell representation matrix X ∈R^64 × n as follows:

(3)

We use an iterative process to transform cell representations to their cell type annotations. Let S(t) be a diffusion matrix that describes the dynamic relationship between every two cells, and Z(t) dynamic cell states in the transformation process at time t. We can use a differential equation to describe cell state dynamics in the transformation process, as follows [36, 37]:

(4)

where z_i(t) is the dynamic state of the i-th cell. We have the cell state matrix Z(t), as follows:

(5)

To solve Eq(4), we use the Euler method to establish the discretization form of Eq(4) with timestep τ, as follows:

(6)

where the upper script k indicates the k-th step. The initial states are the cell representations. We have Z⁽⁰⁾ = X. When τ∈(0, 1), Eq(6) converges to a stable state [35], which can be further utilized by simple classifiers to produce cell type annotations. The optimal S^(k) can be approached by Ŝ^(k), where Ŝ^(k) is a learnable parameter matrix that is optimized during training (see Supplementary Materials for details).

Integrating prior cell–cell relationship network

The DIFFormer model focuses on pairwise relationship between cells, while prior cell–cell relationships provide significant value for cell type annotations [38, 39]. A cell–cell relationship network was established using k-nearest neighbour algorithm [40]. Let A be the adjacency matrix of this cell–cell network, we change the updating rule in Eq(6) according to the literature [35], as follows:

(7)

where G = D^−1/2 AD^−1/2, and D is the degree matrix of A.

Loss function in the classifier

In Fig. 1C, we set the DIFFormer depth to 4. A Multi-Layer Perceptron (MLP) was applied on the output of the DIFFormer to produce cell type annotations. The MLP contains 64 neurons in the input layer. The number of neurons in the output layer equals the number of cell types C. The input layer is fully connected to the output layer directly. For the i-th cell, the output of the MLP is treated as the probability that the i-th cell belongs to the corresponding cell type. We applied cross-entropy loss in training the MLP, as follows:

(8)

where y_i is the true cell type of the i-th cell, I_yi = c a binary value indicator function, p_i,c the probability that the i-th cell belongs to the c-th cell type. It should be noted that the loss was calculated only for cells in the reference dataset. Cells in the query dataset enter MLP to get their cell type annotations.

Performance evaluation

As suggested by the literature [6, 15, 41], we measured the annotation performance of scDIFF using four statistics, including accuracy (Acc), Cohen's kappa coefficient (κ), macro-F1 score (mF1) and weighted-F1 score(wF1). These statistics are defined as follows:

(9)

(10)

(11)

(12)

where q is the number of cells with correct cell type annotations, F1(c) the F1-score for the c-th cell type, p_e the theoretical probability of correct cell type annotation from random guess:

(13)

q_c the number of cells that are annotated by scDIFF as the c-th cell type.

Downstream task implementation

Coverage plot

The coverage plot is produced by using Signac [42]. Fragment indices, which were generated from query dataset, were used with Signac to produce visualization of chromatin accessibility using the CoveragePlot function.

Cell-type-specific accessibility peaks

We identify cell-type-specific peaks using Signac and Seurat pipelines for scATAC-seq analysis. Cell-type-specific accessibility peaks were identified by differential accessibility analysis using likelihood ratio tests within a generalized linear model. We applied stringent filtering criteria, requiring peaks to show accessibility in at least 5% of cells in the respective cell type. We focused exclusively on genomic regions showing higher accessibility in the target cell type. For each cell type, we ranked differentially accessible regions by P-value to select the top 1000 regions as specific markers. Additionally, to support a broader downstream analysis, we constructed extended sets containing up to 10 000 peaks per cell type. We retained all peaks for cell types with <1000 significant ones to ensure complete representation across the cellular landscape.

SNPsea analysis

SNPsea [43] is a computational tool that identifies signature cell types of human diseases. It works by testing whether disease-associated genes from Genome Wide Association Studies are preferentially expressed in the given cell types. The method maps genetic variants to genes using linkage patterns and regulatory data. It finds differential gene expression patterns across different cell types, which are associated with diseases. By pinpointing the cellular contexts, where diseases seem to originate, SNPsea provides clearer targets for treatment development. In this study, SNPsea serves as an auxiliary validation tool to test the accuracy of our cell type annotations.

Considering the substantial conservation of chromatin accessibility patterns observed across mouse and human species [44], we utilized LiftOver to convert genomic coordinates of identified cell-type-specific accessibility peaks from mouse to human genome assembly (GRCh37/hg19). Subsequently, we conducted SNPsea analysis using standard parameters on single nucleotide polymorphisms (SNPs) within each set of cell-type-specific peak regions. We then quantify the enrichments for tissue-specific expression patterns using expression profiles of 17 581 genes across 79 distinct human tissue types from the Gene Atlas database [45].

Motif enrichment analysis

We conducted motif enrichment analysis on cell-type-specific peaks using the Signac package. For each set of differentially accessible peaks, we selected a matched background set of 40 000 peaks with comparable GC content, accessibility levels, and peak widths using Signac's FindMotifs function to ensure statistically rigorous comparisons.

Results and discussions

Cell type annotation performances

We benchmarked scDIFF on 14 pairs of reference and query datasets against eight competing supervised cell type annotation methods, including SANGO, EpiAnno, Cellcano, Seurat, SingleR, and three statistical machine learning methods. The three machine learning methods are support vector machine, random forest (RF), and k-nearest neighbours.

scDIFF demonstrated strong performances across 14 dataset pairs, with an average κ of 0.94 (Fig. 2A and Fig. S1 in supplementary). scDIFF outperformed competing methods on 10 datasets, while ranking second only to SANGO on the remaining 4. All methods using peak-by-cell matrix performed better than those using artificial gene activity matrix.

Figure 2.

Performance analysis on dataset pairs. (A) Boxplots summarizing Cohen's kappa for each method. Colored points represent 14 biologically independent dataset pairs. (B) Row-normalized confusion matrix of scDIFF and other methods in comparison on the BoneMarrowA_BoneMarrowB pair. (C) Row-normalized confusion matrix of scDIFF and other methods in comparison on the WholeBrainA_WholeBrainB pair. (D) UMAP visualization of query cell representations in the BoneMarrowA_BoneMarrowB pair. Left panel shows true cell types, right panel scDIFF annotations. The circle highlights the dense mixture of NK cells, regulatory T cells and other T cells. (E) UMAP visualization of query cell representations in the WholeBrainA_WholeBrainB pair. Left panel shows true cell types, right panel scDIFF annotations. The circle highlights the dense mixture of CPN, CThPN, and SCPN cells.

Additionally, we compared the computational efficiency between scDIFF and SANGO (Fig. S2 in supplementary). The computational efficiency is compared by the wall-clock time of executing same cell-type annotation tasks after extracting cell representations on the same hardware. Benefiting from the more streamlined architecture of DIFFormer, scDIFF achieved faster annotation speeds than SANGO.

Rare cell type annotation

We took two pairs of datasets, BoneMarrowA_BoneMarrowB and WholeBrainA_WholeBrainB, to see if scDIFF can identify rare cell types in the datasets. Row-normalized confusion matrices (Fig. 2B and C) were created for this analysis.

For the BoneMarrowA_BoneMarrowB pair, the number of NK cells and regulatory T cells is <1% of all cells. For these two rare cell types, scDIFF correctly annotated 73% of all NK cells and 44% of all regulatory T cells, which are comparable or better than all other methods in comparison. For the WholeBrainA_WholeBranB pair, the number of purkinje cells and SOM+ interneurons are <2% of all cells. For these two rare cell types, scDIFF correctly annotated 92% of purkinje cells and 62% of SOM+ interneurons, which are still comparable or better than all other methods in comparison. Therefore, scDIFF has the state-of-the-art performance in annotating rare cell types.

Separating similar cell sub-types

We still took the BoneMarrowA_BoneMarrowB and WholeBrainA_WholeBrainB pairs for this analysis. For the BoneMarrowA_BoneMarrowB pair, the NK cells and regulatory T cells are generally sub-types of T cells. Separating these cell sub-types from the other T cells is difficult. As shown in Fig. 2B, EpiAnno and Cellcano were both unable to separate NK cells and regulatory T cells. However, as we have mentioned, scDIFF can separate NK cells and regulatory T cells from other T cells with much better performances. Specifically, scDIFF achieved accuracy of 73% for NK cells, 44% for regulatory T cells, and 36% for T cells, while SANGO showed performance with 54% for NK cells, 50% for regulatory T cells, and 20% for T cells. In contrast, EpiAnno and Cellcano almost entirely misclassified these cell subtypes into other cell types, demonstrating their inability to distinguish these closely related T cell populations. We further illustrate the scatter plot using UMAP on the cell representations entering the diffusive transformers (Fig. 2D). Intuitively, the NK cells, regulatory T cells and other T cells form a quite dense mixture, which seems to be difficult to separate. However, the result of scDIFF presents separable yet tiny margins between them. Considering that the original cell type annotations may contain errors, scDIFF can be used to refine cell type annotations for similar cell sub-types.

For the WholeBrainA_WholeBrainB pair, the CPN, CThPN, and SCPN are all sub-types of the cortical projection neurons [46]. Neither EpiAnno nor Cellcano is able to separate them well (Fig. 2C). On the contrary, scDIFF presents the best performance in this task. scDIFF correctly annotated 92% of CPN cells, 86% of CThPN cells, and 78% of SCPN cells. On the UMAP scatter plot (Fig. 2E), CPN, CThPN, and SCPN cells are closely located, scDIFF correctly annotates most of these cell sub-types.

The impact of dataset imbalance

As we mentioned, all datasets in this study are imbalanced datasets. We investigated the relationship between the imbalance score and cell type annotation performances. scDIFF provides a cell type annotation performance that is almost constant on varying dataset imbalance scores. However, most of the other methods in comparison tend to drop their annotation performance when the dataset imbalance degree increases (Fig. 3A).

Figure 3.

Performance comparison across datasets with varying complexity. Cohen's kappa for different methods on 14 dataset pairs, ordered by reference dataset characteristics: (A) imbalance degree and (B) number of cell types. scDIFF demonstrates consistently superior performance regardless of reference dataset complexity measures.

We also validated the relationship between annotation performances and the number of cell types in the datasets. The results are similar to the imbalance degrees. Most methods have a decreasing performance value when the number of cell types increases. scDIFF still presents robust performance against the increasing number of cell types, showing consistent performance alongside SANGO (Fig. 3B).

Performance comparison across different sequencing platforms

As we have mentioned, platform difference between reference and query datasets is a challenge in generalizing cell type annotation models [23]. Currently, scATAC-seq data are obtained by different sequencing technologies, including 10x Genomics, snATAC-seq, and sciATAC-seq. It is interesting to see if scDIFF can be used across different sequencing technologies. We took 19 dataset pairs (Table S3 in supplementary) for this analysis. We achieved an average κ of 0.70 using scDIFF, which is higher than other state-of-art methods (Fig. 4A and Fig. S3 in supplementary). These results support that scDIFF is capable of performing cell type annotations across different sequencing platforms.

Figure 4.

Performance analysis on cross-platform dataset pairs. (A) Boxplots summarizing Cohen's kappa for each method. Colored points represent 19 biologically independent dataset pairs. (B) UMAP visualization of query cell representations in the MosP1_MouseBrain(10x) pair. Left panel shows true cell types, right panel scDIFF annotations. (C) River plots showing the relationship between predicted and actual cell types in the MosP1_MouseBrain(10x) pair.

The performances of scDIFF on cross-platform pairs are lower than those on the same platform. This may be due to the incompatible tissues of ChIP-seq profiles. It should be noted that scDIFF does not have ChIP-seq profiles from matching tissues in this analysis. We only have ChIP-seq profiles from cortical tissues. This severely damages the performance of scDIFF. The only matching pair is the MosP1_Mouse(10x) pair. For this cross-platform tissue-matching pair, scDIFF achieved a κ of 0.96, which is much higher than other pairs. Figure 4B and C present UMAP visualizations and river plots, respectively, illustrating scDIFF's superior ability to accurately annotate cell types across platforms.

Merging multiple reference datasets

As we have mentioned, scDIFF presents good ability in reducing batch effects between datasets from different sequencing platforms. It is possible that scDIFF can further expand this ability to merge datasets from different batches of sample and from different tissues as a reference dataset. We took two steps to analyze scDIFF's performance in this context. The first step is to merge reference datasets from different biological samples and different batches of experiments, but from the same tissue. The second step is to merge reference datasets from different tissues of different biological samples.

In the first step, we took the five human kidney datasets. We iteratively used one dataset as the query dataset and merged the remaining four datasets as a reference dataset, which creates five different pairs of reference and query datasets. When multiple reference datasets from the same tissue are merged, scDIFF outperformed other state-of-the-art methods in four of five pairs of datasets (Fig. 5A and Fig. S4 in supplementary). We did not perform comparison against EpiAnno, as it reports error in handling the merged reference dataset. In addition, when we tried to use only one kidney dataset as the reference dataset, the performances drop constantly on all query cases. Therefore, merging reference datasets from the same tissue may improve cell type annotation performances of scDIFF.

Figure 5.

Performance on merging reference datasets from the same tissue. (A) Cohen's kappa comparison. scDIFF-SR is for scDIFF with single reference dataset. The five single reference pairs are KidneyA_KidneyE, KidneyB_KidneyA, KidneyC_KidneyB, KidneyD_KidneyC, KidneyE_KidneyD. scDIFF-MR is for scDIFF with a merged reference dataset. (B) River plots when KidneyA is the query dataset; (C) river plots when KidneyB is the query dataset; (D) river plots when KidneyC is the query dataset; (E) river plots when KidneyD is the query dataset; (F) river plots when KidneyE is the query dataset.

We produced the river plot of all five pairs with merged reference dataset (Fig. 5B-F) to see the details of scDIFF in this analysis. Overall, all experiments supported scDIFF's good performances in annotating major kidney cell types. Core functional units, including proximal tubules (PT and its subtype PT_VCAM1), thick ascending limb of Henle's loop (TAL), distal convoluted tubule, and connecting tubule are accurately identified. Annotation performances varied across different samples, which may be due to different sample quality, cellular composition complexity, or technical issues in the sequencing processes. In general, scDIFF provides an excellent performance when merging multiple reference datasets.

In the second step, we took datasets from mouse brain (composed of four datasets) and intestine (composed of three datasets). We iteratively used one dataset as query data and merged all remaining datasets as the reference dataset, which creates seven different pairs of reference and query datasets. As we mentioned in the data preprocessing methods, we took three approaches to merge epigenomic modification information from different tissues: taking the intersection, taking the union, and not using modification information. In all cases, scDIFF performed better than other state-of-the-art methods (Fig. 6A and Fig. S5 in supplementary).

Figure 6.

Performance on merging reference datasets from different tissues. (A) Cohen's kappa of scDIFF with different ways of handling merged epigenomic profiles across different brain and intestine tissues (cerebellum, LargeIntestineA, LargeIntestineB, PreFrontalCortex, SmallIntestine, WholeBrainA, WholeBrainB). scDIFF-union is for using the union strategy. scDIFF-intersect is for using the intersection strategy. scDIFF-noepi is for using no epigenomic profiles. Each group of comparison uses a single query dataset, as marked on the axis. All remaining datasets are merged as the reference dataset. The reference dataset contains different number of cell types in different dataset pairs. (B) River plots when the cerebellum dataset is the query dataset. (C) River plots when the LargeIntestineA dataset is the query dataset. (D) River plots when the LargeIntestineB dataset is the query dataset. (E) River plots when the PreFrontalCortex dataset is the query dataset. (F) River plots when the SmallIntestine dataset is the query dataset. (G) River plots when the WholeBrainA dataset is the query dataset. (H) River plots when the WholeBrainB dataset is the query dataset.

The river plots in Fig. 6B-H provide detailed views of the annotation results for each cell type, revealing consistently high annotation accuracy with major cell populations, supporting advantages of scDIFF in handling merged reference dataset. For example, cerebellar granule cells may maintain perfect annotations (Fig. 6G). Complex cell types, such as multiple excitatory neuron subtypes (SCPN, CThPN, and CPN) can be accurately distinguished from each other (Fig. 6E). Although incorrect annotations occur between transcriptionally similar cell types, like endothelial subtypes in intestinal datasets (Fig. 6C, D, F), the overall performance of scDIFF is still superior after merging reference datasets from different tissues. Notably, when merging references from different tissues, a small portion of cells are incorrectly labeled with cell types from other tissues, which slightly affects the performance.

When multiple reference datasets are used, scDIFF exhibits a more pronounced performance advantage over SANGO (Fig. 5A and Fig. 6A). This improvement can be explained by several factors. First, incorporating multiple reference datasets provides scDIFF with a richer cell-type representation space, thereby enhancing its ability to capture cellular heterogeneity. Second, the Diffusive Transformer in scDIFF can more effectively integrate information from different batches and experimental conditions, thereby improving cross-dataset consistency. In contrast, when handling multi-source heterogeneous data, the graph transformer in SANGO may have some limitations in its generalization ability, which may worth further investigations in future works.

scDIFF with different epigenomic profiles

Within the same tissue, we explored the impact of using different epigenomic profiles in scDIFF using five human kidney datasets (Table S1 in supplementary). Each dataset represents a different individual (Table S4 in supplementary). The setup of the experiments is the same as we investigated the effect of merging different reference datasets. When we remove epigenomic profiles from scDIFF, its performance varies across different pairs of datasets. On the KidneyB dataset, the scDIFF without epigenomic profiles performs the best, while on the KidneyC dataset, scDIFF with all epigenomic profiles performs best. This observation implied that the effect of using bulk-level epigenomic profiles in scDIFF is not consistent across different datasets (Fig. 7A and Fig. S6 in supplementary).

Figure 7.

The impact of using different epigenomic profiles. (A) Cohen's kappa of scDIFF variants with different epigenomic profiles are compared on five human kidney datasets. scDIFF-noH3k4me1 is for scDIFF variant without using H3k4me1 profiles. scDIFF-noepi is for scDIFF without using any epigenomic profiles. (B) Cohen's kappa of different scDIFF variants and SANGO on human PBMC dataset. scDIFF-H3k4me1 represents scDIFF variants using individual epigenomic profiles. scDIFF-H3k4me3 + H3k4me1 represents scDIFF variants using two epigenomic profiles. scDIFF represents the complete model using all three epigenomic profiles.

To explore this further, we conducted additional experiments using human PBMC datasets, where cells were randomly split into training and testing sets in a 1:1 ratio. We compared scDIFF variants and SANGO (Fig. 7B). Interestingly, the results demonstrated that individual epigenomic modifications could outperform the whole combination of all modifications. Specifically, scDIFF-H3k4me1 achieved the highest κ of 0.893, followed by scDIFF-H3k4me3 (0.887), surpassing the complete scDIFF model (0.879). All scDIFF variants significantly outperformed SANGO (κ = 0.851), confirming the superiority of our approach. This suggests that selective incorporation of epigenomic features may be more effective than all available modifications. Therefore, we provide users of scDIFF an option to include epigenomic modification information by their own choices.

scDIFF enables interpretable annotation

We demonstrated our method's biological relevance by annotating cell types in 10X scATAC-seq data NormalCortex using PrefrontalCortex as reference. Since the query data lacked ground truth labels, we compiled marker genes for scDIFF-annotated cell types. We included Ndrg2 for Astrocytes, Itm2a for Endothelial II cells, Neurod6 for Excitatory neurons and several marker genes for other cell types (Fig. 8A and Fig. S8 in supplementary). Coverage plots revealed distinct peak enrichment in corresponding marker gene regions for each cell type. Overlapped peaks appear only between Inhibitory neurons and SOM+ Interneurons.

Figure 8.

scDIFF enables interpretable annotations. (A) Coverage plots of chromatin accessibility within genomic regions spanning ±3 kb around marker genes for scDIFF-annotated cell types: Ndrg2 (astrocytes), Itm2a (endothelial II cells), and Neurod6 (excitatory neurons). (B) SNPsea analysis showing the top 30 tissues with the highest excitatory neuron content based on tissue-specific expression enrichment of SNPs within cell type-specific peaks. Heatmap displays Pearson correlation coefficients organized through hierarchical clustering. (C) UpSet diagram displaying cell-specific motifs and shared motifs among different cell types. Horizontal bars represent the top 40 significant motifs for each cell type, while vertical bars indicate cell-type-specific and shared motifs. Inhibitory neurons and SOM+ interneurons show the highest specificity (82.5%), while oligodendrocytes had the lowest specificity (32.5%). (D) Overrepresented DNA motifs identified by cell type-specific peaks for astrocytes, endothelial II cells, and excitatory neurons.

We further validated annotations using Signac to identify cell type-specific peaks, followed by SNPsea analysis of SNPs within these regions to evaluate tissue-specific expression enrichment. Quantifying enrichment across 79 tissues, we identified the top 30 tissues with the highest excitatory neuron content (Fig. 8B). Results demonstrated significant enrichment of scDIFF-annotated cell types in brain-related tissues, providing additional validation for our approach.

Finally, motif enrichment analysis through Signac revealed that most motifs were specific to the annotated cell types (Fig. 8C). Inhibitory neurons and SOM+ interneurons have the highest specificity, with 82.5% (33 of the top 40 motifs) being cell-type-specific, while oligodendrocytes the lowest specificity at 32.5%. The motifs for each cell type (Fig. 8D and Fig. S9 in supplementary) were extensively verified by the literature [47, 48].

Conclusion

We proposed scDIFF, which integrates bulk-level genomic and epigenomic information in a diffusive transformer to annotate cell types from scATAC-seq data. The performance of scDIFF is comparable or better than state-of-the-art methods in all comparisons. Our scDIFF can eliminate batch effects between different datasets, which enables it to integrate datasets from different sequencing platforms together. With optimized epigenomic profiles, scDIFF presents superior annotation performances, indicating it has great potential to facilitate analysis of scATAC-seq data. Our scDIFF provides both accurate cell type annotations and meaningful insights into the gene regulatory mechanisms at single cell level. The superior performance of scDIFF result from the unique multi-omics integration strategy and the distinct architecture of the Diffusive Transformer. In particular, the incorporation of bulk-level epigenomic data may provide a broader regulatory landscape that helps stabilize feature learning and improve annotation accuracy across datasets. Future work will further clarify these factors and explore more efficient architectures.

Key Points

• The scDIFF pipeline incorporates bulk-level genomic and epigenomic information in scATAC-seq based cell-type annotations.

• Deep diffusive transformers are applied in refining cell representations.

• The scDIFF pipeline can remove batch effects efficiently to merge multiple reference datasets from different sources.

Author contributions

H.Y.W. collected the data, constructed the model, implemented the algorithm, performed experiments, analyzed results, and partially wrote the manuscript. H.D. helped in implementing the algorithm. P.F.D. supervised the whole study, conceptualized the algorithm, analysed the results, and partially wrote the manuscript.

Conflict of interest: None declared.

Funding

This work is partially supported by the National Natural Science Foundation of China [NSFC 62372320 and NSFC 61872268].

Data availability

The code and data for reproducing the results of this paper are available in a GitHub repository (https://github.com/haoyu-wangg/scDIFF).