Workflow

The proposed framework uses single-cell RNA sequencing (scRNA-seq) data as its real-data input. Each cell is represented by a vector of gene expression values, providing a high-dimensional description of its transcriptomic state. These profiles preserve the biological variation among individual cells and form the basis of the proposed graph-guided generative model. The overall architecture is illustrated in Fig 1, and its main components are described below.

• Data preprocessing: Raw scRNA-seq data are subjected to preprocessing before being provided to the learning framework. Normalization corrects for differences in sequencing depth and library size across cells, bringing expression values onto a comparable scale. Quality control is applied alongside this to filter out cells or genes with poor expression coverage or other problematic characteristics. Together, these steps yield a cleaner transcriptomic representation and limit the extent to which technical noise influences model training downstream.
• Cell neighbourhood graph construction: Once preprocessing is complete, a cell neighbourhood graph is built to capture local relationships among cells based on transcriptomic similarity. Cells are represented as nodes, and edges connect cells whose expression profiles resemble one another. We use a k-nearest neighbours (KNN) approach, following Dong et al. [18], to identify each cell’s k most similar counterparts. The resulting graph captures the local structure of the cell population, grouping transcriptionally similar cells into connected neighbourhoods.
• Graph attention network (GAT): The generated cell graph is subsequently processed by the graph attention network (GAT) (Velivckovic et al. [19]). GAT utilizes an attention mechanism that allows for unequal contributions of the neighboring nodes when aggregating information on graphs. In the proposed framework, the GAT model generates representations of individual cells by aggregating information from their neighboring cells while giving more weight to the more informative interactions. Thus, the generated node representations are able to contain both transcriptomic information of individual cells and their interactions with neighboring cells. The attention mechanism offers additional means for selecting informative cells and interactions for the future generation process.
• Generator: The generator takes as input the embeddings extracted using the GAT along with a noise vector and generates synthetic cell profiles using the aforementioned inputs. The embeddings help bring structural and biological relevance to the data, while the noise helps introduce variability in the synthetic samples generated and also makes them less dependent on the individual training data points. The generator learns, using adversarial training, how to generate gene expression profiles similar to those of the actual scRNA-seq data. The aim is to produce diverse synthetic cells with biological significance contained within the original dataset.
• Discriminator: The discriminator network learns to differentiate between real scRNA-seq profiles and those created by the generator. In each round of training, both real and synthetic expressions are fed to the discriminator network, which determines if a particular sample comes from the actual or generated data. The resultant discriminator loss is then used to update the weights of the network through the process of backpropagation. At the same time, the generator is trained using the feedback received from the discriminator in order to create more realistic samples. This iterative process helps in generating more realistic samples from the generator network. The final generator network can then be used for generating synthetic scRNA-seq profiles for applications such as data augmentation and downstream analysis.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Workflow of the proposed GARAGE model.

A: Preprocessed scRNA-seq data are transformed into a cell–cell neighbourhood graph. A graph attention network (GAT) assigns higher attention weights to rare or distinctive cells, and these prioritized cells are partially leaked into the generator input together with random noise. B: The GAN module (generator + discriminator) uses this graph-attentive input to produce realistic synthetic scRNA-seq data while preserving cell-type proportions and enhancing rare subpopulations.

https://doi.org/10.1371/journal.pcbi.1014601.g001

Underline theory that supports GARAGE

GARAGE model is a combination of two deep learning models: Graph Attention Network (GAT) introduced by Velivckovic et al. [19] and Generative Adversarial Network (GAN) introduced by Goodfellow et al. [12]. Each of them solves a different aspect of the problem of generating synthetic data in the two-stage GARAGE architecture. Specifically, GAT learns the relationships between cells and detects which cells or cellular neighborhoods contain more informative signal. The GAN then uses this information along with the stochastic information to learn the underlying distribution of the data and generate synthetic cells. Thus, GAT provides information based on the learned graph structure, and GAN is responsible for generating synthetic scRNA-seq data. In the following subsections, we describe the theory behind GATs and GANs and later discuss how they are implemented in the GARAGE architecture.

Generative adversarial network (GAN).

Our approach for generating high-fidelity single-cell RNA-seq data is built upon the generative adversarial network (GAN) framework, first proposed by Goodfellow et al. [12]. GANs are a class of deep generative models (Hinton et al. [24]) uniquely suited for learning complex, HDSS data distributions, making them a powerful tool for biological data synthesis. With the help of such an approach, we are able to create realistic synthetic samples, which will allow us to tackle the issues associated with research in the area of bioinformatics, such as the lack of data, expensive experiments, and the issue of patient anonymity.

The architecture of the GAN includes a zero-sum game of min-max between two neural networks: a Generator (G) and a Discriminator (D). The goal of the generator is to learn the distribution of the actual data. To do this, the generator uses a latent variable z and tries to transform it into some synthetic sample G(z) that cannot be differentiated from the real data. At the same time, the discriminator acts like a binary classifier and tries to distinguish the real data from the synthesized fake data. The two networks are trained in an iterative, adversarial manner: the generator tries to maximize the discriminator’s classification error, while the discriminator strives to minimize it. This competitive dynamic continues until an equilibrium is reached, at which point the generator produces samples whose distribution is very close to that of the real data.

From a statistical perspective, generative models (Harshvardhan et al. [25]) aim to learn the underlying probability distribution of a training dataset, denoted as p_data(x), in order to draw new, realistic samples. Traditional approaches, such as the Hidden Markov Model (HMM) (Awad et al. [26]) or Gaussian Mixture Models (Viroli et al. [27]), are often explicit generative models. They define a parametric mathematical form for the probability density function (pdf) and learn its parameters, typically by maximizing the likelihood of the training data. However, for HDSS and complex data like single-cell transcriptomes, defining and optimizing such an explicit density function is often intractable.

GANs offer a fundamentally different and powerful alternative as an implicit generative model. Instead of defining a formula for the pdf, a GAN learns a procedure to sample directly from the target distribution. It achieves this by transforming a simple latent variable z, drawn from a prior distribution p_z(z), through the complex, non-linear function incorporated by the generator network, G(z). This process induces a learned data distribution, p_g. The statistical essence of the adversarial training lies in its mechanism for matching p_g to p_data. The min-max game between the generator and discriminator functions as an effective approach to minimize a statistical divergence between the two distributions. In the original framework, it has been proven that the process is equivalent to the minimization of Jensen-Shannon (JS) divergence (Menendez et al. [28]). Such an indirect approach helps GANs capture all the statistical aspects of the data without having to calculate or even define its likelihood, hence making them extremely suitable for modeling the complicated, sparse, and multi-modal nature of scRNA-seq data. Formally speaking, the training of GAN is modeled after Goodfellow et al. [12]:

(1)

GANs provide several benefits over other generative algorithms such as Variational Autoencoders (VAEs) (Kingma et al. [15]), Restricted Boltzmann Machines (RBMs) (Fischer et al. [29]), Deep Belief Networks (DBNs) (Hinton et al. [30]), and denoising autoencoders (DAEs) (Vincent et al. [31]). These benefits become especially valuable when dealing with high-dimensional and sparse scRNA-seq data. The first one is the fact that the models learn the underlying distribution of the data implicitly. There is no need to explicitly define a probability density function as the generator approximates the distribution based on its adversarial interactions with the discriminator. This approach becomes especially useful for complicated and high-dimensional data sets where formulating a suitable probability model is problematic. In addition, the adversarial objective allows assessing the quality of generated samples based on the resemblance of their distribution to the true distribution. Finally, the generator allows generating samples directly using forward evaluation without iterations used in Markov chain-based methods. More details about GAN formulation are provided in the Subsection (3.2) in S1 File..

Despite these advantages, conventional GANs present several challenges, most notably training instability and mode collapse, in which the generator may produce samples concentrated around only a limited portion of the target distribution. These limitations are particularly important for scRNA-seq data, where rare cellular populations may occupy relatively small regions of the overall data distribution. To address this issue, we introduce GARAGE, an architectural extension that incorporates a graph attention network (GAT) into the generative framework. The GAT provides structure-aware information from the cellular neighbourhood graph and guides the generator toward informative regions of the data space. The proposed integration of graph-based attention with adversarial generation is described in the following section. The complete set of GAN parameters is provided in Subsection (3.2) in S1 File.

The proposed GARAGE framework: A GAT-seeded generative model

The proposed GARAGE framework combines a graph attention network (GAT) and a generative adversarial network (GAN) within a two-stage architecture, as shown in Fig 1. The main motivation behind this design is to address the difficulty of training conventional GANs on high-dimensional, sparse single-cell data. Rather than allowing the generator to learn entirely from randomly sampled inputs, GARAGE first identifies cells that are informative from a biological and graph-structural perspective. These cells are subsequently used to guide the generation process, helping the model remain within a biologically meaningful region of the data space.

The first stage (Fig 1A) is concerned with finding the cells that are most useful as generation seeds. A neighbourhood graph is constructed from the preprocessed scRNA-seq data using the kNN graph approach [18]. This graph is subsequently used to train a GAT for cell-type classification. A notable characteristic of the GAT is that while aggregating information from neighbouring cells, it learns attention weights, which help to estimate the impact of individual cells on the learned representation. Attention weights allow identifying the cells that have the strongest impact on the representation through their attention values. After ranking cells by their aggregate attention scores, the top-k% are chosen as the seed set, which will be used further. The selected cells are representatives of the cellular states and will be sent to the generative phase.

The second stage (Fig 1B) represents the modification of the classic input for the generator of the GAN architecture, which is a novel aspect of the module design. In contrast to the conventional approach where the generator is provided only with a vector of random noise z, GARAGE uses the hybrid input , which is a combination of the seed cells and random noise vectors. Seed cells represent the reference states observed in the real data and random noise enables the generator to generate new synthetic states different from those provided by seed cells. Hybrid inputs are shuffled before being fed into the generator G, which produces the generated sample .

The generator is trained by the adversarial method using the objective function defined by Eq. (1). The discriminator D receives real samples x from the complete dataset as well as synthetic samples generated by , and learns to distinguish between the two. At the same time, the generator attempts to produce samples that are sufficiently realistic to make this distinction difficult for the discriminator. Here, denotes the expectation over the real data distribution, while represents the expectation associated with the hybrid input distribution used by the proposed seeded generator. Through this iterative, adversarial training, the generator learns not only how to transform noise into coherent gene expression profiles but also how to produce plausible variations around the high-quality seed cells. This process continues until an equilibrium is established, at which point the discriminator can no longer reliably distinguish between the real data and the high-fidelity synthetic data generated by GARAGE.

The mathematical formulation of GARAGE

As mentioned earlier, the GARAGE framework comprises two stages: (1) a GAT-based module that identifies a set of attention-prioritized, high-importance “seed” cells; and (2) a GAN whose generator is guided by these seeds through a batch-level hybrid input that mixes sampled GAT-prioritized seed cells with random noise vectors.

Let the preprocessed single-cell dataset be represented by a matrix , where N is the number of cells and M is the number of genes.

Stage 1: GAT-based representative cell selection

• We model cellular relationships as a graph , where indexes cells and edges are built by a k-nearest-neighbor procedure in gene-expression space (KNN graph; Dong et al. [18]). Each node carries an input feature .

• A GAT produces attention coefficients per edge . For a single head,

(2)

where are node features (possibly a projection of ), and are learnable, denotes concatenation, and is the neighborhood of .

• After training the GAT (e.g., for cell-type prediction), we define an importance score that aggregates incoming attention to node :

(3)

Cells are ranked by , and the top-K (or top-q%) form the set of representative priority nodes. We collect their raw or final-layer GAT embeddings into and denote the index set by .

Stage 2: GAT-seeded adversarial generation

The novelty of GARAGE lies in replacing the purely noise-driven generator input with a batch-level hybrid input. In the reported experiments, the GAT-selected priority cells were not collapsed into a single attention-pooled seed vector. Instead, the priority seed set was used directly: at each GAN iteration, a fraction of the generator input batch was sampled from these GAT-prioritized seed cells, while the remaining fraction was sampled from the random noise prior. Thus, the leakage mechanism operates through batch-wise mixing of biologically informative real-cell anchors and stochastic noise inputs.

• Conventional baseline: A standard GAN solves the min–max game

(4)

where p_data is the real data distribution and is a simple prior (e.g., ).

• Attention-ranked seed distribution: After training the GAT, we aggregate the learned incoming attention coefficients to obtain an importance score for each cell:

(5)

Cells are ranked according to , and the top-K cells form the GAT-prioritized seed set

(6)

In the reported experiments, this discrete seed set was used directly. We did not collapse the selected cells into a single attention-pooled vector. Therefore, should be interpreted as an empirical seed distribution from which high-attention real-cell anchors are sampled during GAN training.

• Batch-level hybrid generator input used in the experiments: For a generator input batch of size m, the leakage rate determines how many GAT-selected seed cells are inserted into the batch:

(7)

We sample cells from and random noise vectors from . The hybrid input batch is then

(8)

Here, denotes batch-wise concatenation followed by shuffling, not feature-wise concatenation into a single vector. This batch-level construction is the attention-guided leakage mechanism used in all reported experiments.

• GARAGE objective: Using the batch-level hybrid input distribution, the GARAGE adversarial objective is written as

(9)

Here, denotes the empirical hybrid input law induced by the batch-wise mixture of GAT-selected seed cells and random noise vectors. When , contains only random noise and GARAGE reduces to a conventional GAN. When , a controlled fraction of the generator input batch consists of attention-prioritized real-cell anchors, thereby guiding the generator toward biologically informative and rare-cell-relevant regions of the scRNA-seq manifold.

1. (i) The construction in Eqs. (7)–(8) clarifies that leakage is implemented at the batch level rather than by forming a single pooled seed vector.
2. (ii) The seed component of the batch is obtained from the GAT-prioritized empirical distribution , whereas the remaining component is sampled from the random noise prior .
3. (iii) This description is consistent with Algorithm 1, which was the procedure used in all reported experiments.

For reproducibility, the architecture and training settings used in the reported experiments are summarized in Supplementary Tables (2) and (3) in S1 File. The GAT module was implemented as a two-layer GATConv network on a k-nearest-neighbour cell graph with k = 5. The first GAT layer used 8 attention heads with 32 hidden units per head, followed by a ReLU activation; the second GAT layer used a single attention head to produce the final cell-type prediction. The GAT was trained using cross-entropy loss and the Adam optimizer with learning rate 0.005 for 7501 epochs. The generator was implemented as an MLP with two hidden layers of 512 neurons each and LeakyReLU activation with negative slope 0.2, followed by a linear output layer. The discriminator was implemented as a smaller MLP with hidden layers of 256 and 128 neurons, also using LeakyReLU activation. The generator and discriminator learning rates were 0.0002 and 0.0004, respectively. GAN training was performed for 20001 iterations, with discriminator updates and generator updates per iteration. The hybrid input used in the reported experiments was the batch-level mixture defined in Eqs. (7)–(8), no attention-pooled seed vector or -controlled pooling was used in the reported results.

The detailed algorithm for the entire GARAGE framework can be found in the algorithm 1. The complete architecture and dataset-specific hyperparameters used in all reported experiments, including the GAT configuration, generator and discriminator layer sizes, optimizer settings, leakage rates, and batch sizes, are provided in Supplementary Tables (2) and (3) in S1 File.

Algorithm 1. The GARAGE framework algorithm.

Require: scRNA-seq data matrix ; cell-type labels ; number of nearest neighbours ; GAT architecture: two GATConv layers, first layer with 8 heads and 32 hidden units per head, second layer with 1 head; GAT optimizer: Adam, learning rate 0.005, epochs ; leakage percentage ; batch size m; GAN iterations ; generator architecture [512,512] with LeakyReLU(0.2) and linear output; discriminator architecture [256,128] with LeakyReLU(0.2); generator learning rate 0.0002; discriminator learning rate 0.0004; discriminator/generator steps .

Ensure: Trained generator network .

   Stage 1: GAT-based representative cell selection

1: Construct k-NN graph from data using k_nn.

2: Initialize GAT classifier with parameters .

3: for epoch=1 to do

4:  Train on to predict labels by minimizing cross-entropy loss.

5: end for

6: Extract final-layer attention coefficients from trained .

7: for each cell do

8:  Calculate importance score:

9: end for

10: Select the set of top-K seed cells:

   Stage 2: GAT-seeded adversarial generation

11: Initialize Generator and Discriminator with random weights.

12: for iteration = 1 to T_GAN do

13:  Train Discriminator for step = 1 to

14:  for step =1 to do

15:   Sample a batch of real data from .

16:   ; .

17:   Sample GAT-prioritized seed cells from .

18:   Sample random noise vectors from .

19:   Construct the batch-level hybrid input .

20:   Shuffle before feeding it to the generator.

21:   Generate fake samples .

22:   Update discriminator parameters by ascending the stochastic gradient of:

23:  end for

24:  Train Generator for step = 1 to

25:  for step =1 to do

26:   Sample and construct a new hybrid input batch of size m.

27:   Update generator parameters by descending the stochastic gradient of:

28:  end for

29: end for

30: return Trained Generator

Empirical assessment of training stability

We empirically evaluated whether attention-guided leakage improves the stability of adversarial training. Specifically, we compared the no-leakage baseline with GARAGE under leakage settings using the same generator and discriminator architecture, optimizer settings, preprocessing pipeline, and training schedule. The only difference between the baseline and GARAGE was the construction of the generator input batch: the baseline used random noise only, whereas GARAGE used a batch-level mixture of random noise vectors and GAT-prioritized seed cells.

For each dataset and leakage setting, we monitored the generator loss and discriminator loss across training iterations. We also quantified training stability by measuring loss oscillation over the final training window and by computing the final distributional similarity between real and generated cells. Lower loss oscillation indicates more stable adversarial training, while lower distributional distance indicates better agreement between real and generated transcriptomic distributions.

This analysis evaluates whether attention-guided leakage provides empirical stabilization relative to a conventional noise-driven GAN. The complete loss curves and run-wise stability summaries and detail explanations are provided in subsection “Training-stability ablation,” Supplementary Table (10) and Supplementary Figure (34) in S1 File.

Practical implication of the convergence analysis. The above analysis provides an intuitive explanation for the empirical stability of GARAGE. In a conventional GAN, the generator initially receives only random noise, and many generated samples may lie far from the real single-cell manifold, producing unstable discriminator feedback and increasing the risk of mode dropping. In GARAGE, a controlled fraction of the generator input is anchored by GAT-prioritized real-cell seeds. These seeds provide biologically meaningful reference points that guide the generator toward under-represented but valid regions of the data manifold. Consequently, the generator receives more informative gradients, the discriminator is less likely to saturate early, and adversarial training becomes more stable. Thus, the theoretical analysis supports the practical observation that attention-guided leakage can improve both convergence behavior and rare-cell preservation during synthetic scRNA-seq data generation.

Overview of datasets

To comprehensively evaluate the performance of our proposed method, four public single-cell RNA-seq (scRNA-seq) datasets were utilized: Yan, CBMC, Pollen, and Muraro. These datasets were chosen to span a range of biological systems, cell counts, gene dimensionalities, and class complexities, thereby providing a robust benchmark for generative models. A summary of their properties is provided in Table 1.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Summary of the datasets utilized here.

https://doi.org/10.1371/journal.pcbi.1014601.t001

Data preprocessing

Prior to model training, each scRNA-seq dataset underwent a standardized preprocessing pipeline to mitigate technical noise and ensure data quality (Hwang et al. [36], Tian et al. [37]), a critical step for the stability and accuracy of downstream graph construction and adversarial learning. The raw gene expression count matrices were subjected to a series of quality control (QC) and normalization steps.

First, the QC filtering is applied at the cell level. Cells with an unusually low number of expressed genes or a high percentage of mitochondrial gene counts—often indicative of low-quality captures or apoptotic cells—were removed (Hwang et al. [36]).

Following QC, the filtered count data employed a standard log-normalization approach: the counts for each cell were divided by the total counts for that cell, multiplied by a scale factor (here, 10,000), and then log-transformed with a pseudocount of one (log1p). Let be the raw gene expression count matrix, where is the count of gene j in cell i. The normalized expression value, , is calculated as:

(10)

where is the total number of counts for cell i across all M genes, and S is a scale factor, here 10,000.

For the underlying preprocessing for scRNA-seq data, we used scanpy version 1.9.6 (Py version 3.9.21) (Wolf et al. [38]).

The outcome of this pipeline is a clean, normalized, and quality-controlled gene expression matrix, which serves as the foundational input for constructing the cell neighbourhood graph and training the GARAGE model.

Quantitative benchmarking of GARAGE against leading state-of-the-art models

To rigorously validate the performance of GARAGE, we expanded the benchmarking framework to include three groups of baselines. The first group contains generic generative models, including standard GAN, VAE, f-GAN, and WGAN. The second group contains scRNA-seq-specific generative models, including LSH-GAN, scGAN/cscGAN, scVAE, and a diffusion-based single-cell generative model. The third group contains simple rare-cell balancing controls, including standard GAN and VAE trained on class-balanced or artificially oversampled versions of the original data. These oversampling baselines were included to test whether GARAGE provides benefits beyond simply increasing the representation of rare cell types in the training set.

The oversampling controls were constructed by randomly upsampling minority cell types in the training data until each class reached the size of the largest class or a predefined balancing threshold. Standard GAN and VAE models were then trained on these balanced datasets using the same preprocessing and downstream evaluation pipeline as GARAGE. These baselines are denoted as GAN + ROS and VAE + ROS, respectively. This comparison directly evaluates whether graph-attention-guided seed selection provides an advantage over simple class balancing. The detailed comparison with the additional scRNA-seq-specific baselines and oversampling controls is provided in Supplementary Tables (6) and (7) in S1 File.

The novelty of our GARAGE framework lies in its unique two-stage architecture, which fundamentally rethinks the input to the generative process. While standard GANs and their variants learn to map a random latent space to the data distribution, GARAGE first employs a graph attention network (GAT) as an intelligent selection module. The GAT is trained on a k-NN graph (Dong et al. [18]) of the real data to identify a core subset of samples with the highest attention scores, indicating their importance in defining cell-type identities and the overall data manifold. These attention-driven nodes are then used to “seed” the GAN’s generator by being mixed with random noise in the input batch. This intelligent seeding strategy directs the generative process to known, high-quality cell states that enhance the data generation of synthetic data.

To ensure a fair and robust comparison, all SOTA models were trained on default parameters (stated in the Supplementary Note (3.5) in S1 File) and evaluated under a standardized protocol across four widely used scRNA-seq datasets: Yan, Pollen, CBMC, and Muraro (stated in Table 2). Performance was quantified using three key metrics: the Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI) to assess clustering fidelity, and the macro-F1 score to evaluate the balance and class accuracy after clustering. Also, each model has been assessed under two feature selection methods (Fano index and CV2) and on three data regimes, namely generated, original, and combined (generated + original). This rigorous benchmarking framework allows us to directly compare the capacity of each generative model to preserve the complex biological signals in single-cell data.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Performance comparison across datasets, generative methods, data types, feature selection methods, and metrics. For readability, the best-performing value within each dataset-specific comparison block is highlighted.

https://doi.org/10.1371/journal.pcbi.1014601.t002

Across the benchmark datasets, GARAGE showed stronger performance than the evaluated generic generative models, scRNA-seq-specific generative models, and oversampling-based controls under the same preprocessing and downstream evaluation pipeline (Table 2). In particular, GARAGE achieved the highest cluster fidelity in terms of Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI) across most of the datasets. For example, on the relatively straightforward Yan dataset, GARAGE nearly attained perfect clustering alignment (ARI 0.97, NMI 0.97 using CV2-selected features), substantially surpassing the best baseline (LSH-GAN, ARI 0.87, NMI 0.91). Similarly, in the more heterogeneous Pollen data, GARAGE reached ARI scores up to 0.93 (considering CV2 feature selection method), again exceeding the strongest alternative model by a wide margin (LSH-GAN at ARI 0.83). Even on the most challenging benchmarks, such as CBMC and Muraro datasets, where all methods showed lower absolute performance due to complex, imbalanced cell populations, GARAGE maintained a clear advantage. Notably, on CBMC its synthetic data obtained the highest ARI (0.54 vs. 0.47 for the best baseline) and NMI (0.61 vs. sub-0.54 for others), and on Muraro GARAGE likewise topped the baselines with ARI around 0.53 (compared to 0.48 from the next-best model). These improvements in ARI/NMI indicate that GARAGE-generated cells cluster with greater purity and alignment to ground-truth labels than those produced by any competing method, across datasets ranging from clean and simple to noisy and complex. For GARAGE, the parameters for ARI, NMI, F1, and all additional UMAP results are provided in the Supplementary Note (4) in S1 File.

Also, it is noteworthy that GARAGE showed better balancing when detecting all cell subpopulations, as evidenced by the highest macro-F1 scores at all times. The macro-F1 of GARAGE was always better than that of any other model and proved the ability of our model to accurately reproduce all types of cells, including the minority ones. For example, in the Yan dataset, the GARAGE macro-F1 score was 0.307, while in others, the scores were significantly lower (usually less than 0.15-0.20). This means that our approach does not sacrifice small clusters in favor of the larger ones. The same applies to the most difficult CBMC and Muraro datasets, in which the macro-F1 scores for GARAGE were 0.13 and 0.28, respectively. Again, we see that this score is better than any baseline and shows much more balanced detection of cell types. This is an important distinguishing characteristic because while the standard GAN and VAE generate mostly samples from the most represented cell states (hence poor recall of the minority cells), the attention-guided seeding in GARAGE ensures that even rare subpopulations are learned and reproduced.

To assess the sensitivity of the quantitative results to stochastic variation, we performed an additional repeated-run stability analysis for the main comparison setting. The full pipeline was repeated across independent random seeds for GARAGE and its competing baselines, including LSH-GAN, GAN, FGAN, and WGAN. For each run, model initialization, synthetic-data generation, and clustering initialization were varied, and ARI and NMI were summarized as mean standard deviation. These results are provided in Table 3.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Repeated-seed stability analysis for the main comparison setting. Values are reported as mean standard deviation across independent random seeds.

https://doi.org/10.1371/journal.pcbi.1014601.t003

Overall, these benchmarking results establish GARAGE as a robust, generalizable generative framework that delivers state-of-the-art performance across diverse conditions. Its improvements were consistent under both gene selection strategies tested (Fano index and CV2) and were evident whether we evaluated only generated cells or a combined set of generated+original cells. In all scenarios, GARAGE not only surpassed the other generative models in cluster quality (achieving ARI/NMI values close to the theoretical maximum in easy cases, and the highest scores in difficult cases) but, in some instances, even matched or exceeded the clustering performance of the original dataset itself. These results suggest that GARAGE’s two-stage design, which leverages a araph attention network to seed the generator with high-value cell states, yields synthetic data that preserves the complex structure of single-cell transcriptomes with exceptional fidelity. The ability to maintain high cluster integrity and balanced class representation across multiple datasets and experimental settings underscores GARAGE’s potential for generating realistic single-cell data. The UMAPs with their respective ARI, NMI, and macro-f1 score for all the SOTA models of all four scRNA-seq datasets are provided in the Supplementary Note (5) in S1 File.

Clustering evaluation using fixed biological marker genes

To disentangle the contribution of the generative model from the feature selection step, we performed an additional clustering analysis using a fixed biologically relevant marker-gene set for each dataset. In the previous analysis, feature selection was performed using Fano index and CV2, and therefore the reported ARI, NMI, and macro-F1 scores reflected the combined effect of generation, feature selection, and clustering. In contrast, the present analysis uses the same predefined marker-gene set for all generative methods within each dataset. These marker genes were selected independently of the generated data, based on known cell-type markers and/or marker information reported in the original dataset studies.

Using these fixed marker genes, we repeated the same clustering evaluation for real, generated, and combined datasets. Because the feature set was held constant across methods, differences in ARI and NMI more directly reflect the ability of each generative model to preserve biologically meaningful cell-type structure rather than differences introduced by feature selection. The full marker-gene lists and the corresponding fixed-marker clustering results are provided in Supplementary Tables (8) and (9) in S1 File.

Biological validation of GAT-prioritized high-attention cells

To provide biological context for the cells selected by the GAT module, we performed a post hoc marker-gene validation on the CBMC dataset. This dataset was chosen because it contains diverse immune and progenitor populations, including both common and low-frequency cell states. After GAT training, cells were ranked according to their aggregated attention scores, and the top-20% and bottom-20% of cells were defined as high-attention and low-attention subsets, respectively. We then examined whether known cell-type marker genes showed elevated expression in the high-attention subset.

The analysis included 24 literature-supported marker genes representing eight immune or progenitor populations. For each marker gene, we computed the mean expression in all cells, high-attention cells, and low-attention cells, followed by fold-change and Wilcoxon rank-sum testing between the high- and low-attention subsets. The marker genes were used only for post hoc biological interpretation and were not used to train the GARAGE model.

The high-attention subset showed consistently elevated marker expression relative to the low-attention subset. All 24 examined markers had positive fold-change, and 15 markers reached Wilcoxon significance (p < 0.05). Significant markers included rare or distinctive population markers such as HBA2 and HBB for erythroid cells, PF4 and PPBP for megakaryocytes, LILRA4, CLEC4C, and TCF4 for plasmacytoid dendritic cells, CD34 and KIT for hematopoietic progenitors, and CLEC10A for conventional dendritic cells. Markers of common immune populations, including S100A8, CST3, and CD14 for monocytes, MS4A1 and CD79A for B cells, and KLRF1 for NK cells, also showed increased expression in high-attention cells. These results indicate that GAT-prioritized cells capture biologically interpretable immune and progenitor signatures, supporting their use as biologically informative anchors for attention-guided leakage in GARAGE. The complete marker-expression summary is provided in Supplementary Table 11 in S1 File.

Label-agnostic and label-scarce validation of GARAGE

Because the GAT module uses cell-type labels during training, we further evaluated GARAGE using validation strategies that do not rely solely on the same labels used for attention-guided seed selection. In addition to ARI, NMI, and macro-F1, we computed label-agnostic distributional similarity metrics between real and generated data on the full preprocessed gene-expression matrix, before feature selection. Specifically, we used Maximum Mean Discrepancy (MMD) with an RBF kernel and Sliced Wasserstein Distance (SWD) over random projections of the complete gene-expression space. These metrics assess whether the generated cells match the real transcriptomic distribution without using cell-type labels or clustering assignments. The full label-agnostic distributional similarity results are provided in Supplementary Table (4) in S1 File.

For two empirical distributions of real cells and generated cells , the unbiased squared MMD was computed as

where is an RBF kernel and the kernel bandwidth was selected using the median heuristic. Lower MMD values indicate stronger similarity between the real and generated distributions. To complement MMD, we also computed SWD by projecting real and generated cells onto multiple random one-dimensional directions and averaging the one-dimensional Wasserstein distances across projections. This provides a scalable distributional comparison in the full gene-expression space.

To further assess utility in a realistic label-scarce setting, we performed a held-out rare-cell experiment. For each dataset, a rare or low-frequency cell type was selected, and a subset of its real cells was withheld from GARAGE training and used only for testing. Synthetic cells generated by GARAGE were then used to augment the limited labelled training set, and a downstream classifier was trained on the augmented data. The classifier was evaluated on the held-out real cells. This setup tests whether GARAGE-generated samples improve recognition of rare cell types in unseen real data, rather than merely improving label agreement on the training distribution. The complete held-out rare-cell utility results are reported in Supplementary Table (5) in S1 File.

Analysis of leakage sensitivity of GARAGE across datasets

We systematically assessed the performance of GARAGE model across four single-cell RNA-seq (scRNA-seq) datasets, namely Yan, Pollen, CBMC, and Muraro, under three leakage levels conditions (10%, 20%, 30%). Therefore, at the time of synthetic scRNA-seq data generation, with this innovative approach, we combine the random noise with our “attention-weighted leakage.” To quantitatively observe the performance, we evaluated the fidelity of generated data over original data using the Wasserstein Distance (WD) (Ruschendorf et al. [39]), where the lower values signify a stronger similarity between the two distributions. Please refer to the Supplementary Note (3.4) in S1 File for the WD parameters of different leakage types.

As detailed in Table 4, we compared our model with different percentages of GAT-enabled real data leakage (10%, 20%, 30%) against a baseline with no leakage (0%). The results depict a clear and consistent improvement in WD when employing the seeded generation strategy. Across all diverse scRNA-seq datasets, the seeded generation produced better synthetic data, as a 0% leakage-based architecture, consistently producing the highest, and therefore worst, WD samples. The introduction of any level of data leakage gradually improved the quality of data generation and consequently decreased the WD between real and generated data. This suggests that our GARAGE framework, by guiding the generator with a curated set of real data samples, provides a robust and significant improvement over the traditional methods in synthetic scRNA-seq data generation.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Wasserstein distance for varying leakage percentages across datasets.

https://doi.org/10.1371/journal.pcbi.1014601.t004

For more robustness, at the UMAP-based adjusted rand index(ARI), normalized mutual information(NMI) calculation, we saw the difference in clusters. The data cluster obtained from the generated data not only becomes separated, but also they are significant (in most cases). We performed two feature selection (FS) strategies (Fano index and CV2). The ARI, NMI and macro F1-score were the performance metrics used to facilitate the evaluation on three kinds of dataset samples - generated, real, and combined(generated + real). The methodical structure provides insight into understanding the role played by the intrinsic structure of each dataset and the adopted feature selection method, under controlled leakage levels, in influencing generative fidelity and manifold stability.

GARAGE was evaluated with attention-weighted leakage of 10%, 20%, and 30% over four scRNA-seq datasets (Yan, Pollen, CBMC, Muraro) for both distributional fidelity using Wasserstein Distance (WD) and clustering performance using ARI, NMI, and macro-F1 in combination with Fano and CV2 feature selection applied on generated, original, and concatenated (generated + original) data (Tables 4 and 5). The following observations were made. In comparison with the no-leakage setting, any positive level of leakage caused a decrease in WD, implying better approximation of synthetic by real distributions; the largest effect of WD reduction was observed at an intermediate leakage value. For example, Yan had WD values 4.20 and 1.69 for 0% and 30% leakages, respectively; Pollen did not exhibit a monotonic WD dependence on leakage but had the smallest WD (1.22) at 20%. Finally, in the CBMC dataset where the difference in global distributions was quite small, leakage of 10% additionally decreased WD (to 0.0020), suggesting that even limited seeding sharpens local distributional details.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Clustering performance (ARI, NMI, Macro F1) for different datasets and leakage percentage(%) using Fano and CV2 feature selection.

https://doi.org/10.1371/journal.pcbi.1014601.t005

Clustering analyzes echoed these trends and clarified the role of leakage as a guidance signal that must be balanced against overfitting. On generated data, ARI and NMI typically rose from 0% to a moderate leakage level, often achieving near-perfect alignment on well-structured datasets with CV2 features (e.g., Yan ARI , NMI ). Leakage also positively affected Macro-F1, which depends on the restoration of minority cells, with improvements being prominent while transitioning from 0% leakage to 10%-20%, as shown in high macro-F1 values for Muraro combined data at 20% (). More difficult scenarios such as CBMC still saw the positive effect of leakage on generated-data clustering (e.g., peak ARI value around 0.54 at 30%), meaning that noisy manifolds might need more robust seeds from real signals in order to be separable. On the other hand, heterogeneous datasets, such as Pollen, were sometimes negatively impacted by excessive leakage (30%), especially on combined data—consistent with over-anchoring the generator to training labels and reducing generative flexibility.

To make the choice of leakage rate more interpretable, we consolidated the leakage-sensitivity results across datasets and related the observed trends to dataset-level properties such as class imbalance, rare-cell representation, and manifold complexity. The results suggest that should not be treated as a universal fixed hyperparameter. Moderate leakage values, particularly –0.20, provide a reasonable starting range for datasets with relatively clear subtype structure or moderate imbalance. For datasets with stronger imbalance or weaker cell-state separation, a larger value such as may be beneficial because stronger real-cell anchoring can help the generator recover underrepresented regions of the manifold. However, increasing should be treated cautiously, because excessive leakage may over-anchor the generator to selected seed cells and reduce generative flexibility. The complete leakage-performance trends across datasets are summarized in Supplementary Figure (36) in S1 File.

Sensitivity to leakage and seed-selection hyperparameters. In addition to the leakage-rate analysis, we examined the sensitivity of GARAGE to the top-q% seed-selection threshold. The leakage rate controls the fraction of GAT-prioritized seed cells inserted into the generator input batch, whereas the top-q% threshold controls the size of the high-attention seed set . We varied these parameters while keeping the remaining model settings fixed. The k-NN graph size was fixed at in all reported experiments because GARAGE relies on local neighbourhood structure for rare-cell-aware attention scoring, and varying requires reconstructing the cell graph, retraining the GAT, recomputing the seed set, and rerunning adversarial training. The sensitivity results indicate that GARAGE is stable across moderate leakage and seed-selection ranges. Moderate seed-selection thresholds provide sufficient biological anchoring without over-constraining the generator, whereas overly large seed sets may reduce generative flexibility. The detailed sensitivity results are provided in the subsection “Focused hyperparameter sensitivity analysis” and in Supplementary Table (12) in S1 File.

UMAP-based evaluation of model performance

We further examined how attention-guided leakage shapes the geometry of the learned embeddings using UMAP, reporting ARI, NMI, and macro-F1 for generated, original, and combined (generated + original) data under both Fano Index and coefficient of variance (CV2) feature selection strategies (Figs 2–5). Across datasets, the visual and quantitative trends align with the leakage-sensitivity analysis: small to moderate leakage typically sharpens cluster boundaries and increases alignment with ground-truth labels. In contrast, excessive leakage can compress variability and introduce over-anchoring to the training labels.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Leakage stress test on the Pollen dataset using the Fano index feature selection strategy.

A: UMAP projections of real, generated, and combined data at 10% leakage, with clustering metrics (ARI, NMI, macro-F1) indicated. B: UMAP projections at 20% leakage with corresponding metrics. C: UMAP projections at 30% leakage with corresponding metrics. D: Distributional shift between real and generated data measured by the 1-Wasserstein distance across leakage levels (10%, 20%, 30%).

https://doi.org/10.1371/journal.pcbi.1014601.g002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Leakage stress test on the Yan dataset using the Fano index feature selection strategy.

A: UMAP projections of generated, and combined data at 10% leakage, with clustering metrics (ARI, NMI, macro-F1) indicated. B: UMAP projections at 20% leakage with corresponding metrics. C: UMAP projections at 30% leakage with corresponding metrics. D: Distributional shift between real and generated data measured by the 1-Wasserstein distance across leakage levels (10%, 20%, 30%).

https://doi.org/10.1371/journal.pcbi.1014601.g003

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Leakage stress test on the Pollen dataset using the CV2 feature selection strategy.

A: UMAP projections of generated and combined data at 10% leakage, with clustering metrics (ARI, NMI, macro-F1) indicated. B: UMAP projections at 20% leakage with corresponding metrics. C: UMAP projections at 30% leakage with corresponding metrics. D: Distributional shift between real and generated data measured by the 1-Wasserstein distance across leakage levels (10%, 20%, 30%).

https://doi.org/10.1371/journal.pcbi.1014601.g004

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Leakage stress test on the Yan dataset using the CV2 feature selection strategy.

A: UMAP projections of generated, and combined data at 10% leakage, with clustering metrics (ARI, NMI, macro-F1) indicated. B: UMAP projections at 20% leakage with corresponding metrics. C: UMAP projections at 30% leakage with corresponding metrics. D: Distributional shift between real and generated data measured by the 1-Wasserstein distance across leakage levels (10%, 20%, 30%).

https://doi.org/10.1371/journal.pcbi.1014601.g005

Data and code availability

The benchmark datasets used in this study are publicly available from their original sources. The Yan dataset was obtained from Yan et al. [32]; the raw data are available through the Gene Expression Omnibus (GEO) under accession GSE36552. The CBMC dataset was obtained from the CITE-seq study of Stoeckius et al. [33]; the data are available through GEO under accession GSE100866, with associated BioProject PRJNA393315 and SRA accession SRP111194. The Pollen dataset was obtained from Pollen et al. [34]; the dataset is available through the Human Cell Atlas Data Portal under the project “Low-coverage single-cell mRNA sequencing reveals cellular heterogeneity and activated signaling pathways in developing cerebral cortex.” The Muraro dataset was obtained from Muraro et al. [35]; the raw data are available through GEO under accession GSE85241.

The GARAGE source code, preprocessing scripts, model-training scripts, synthetic-data generation scripts, biological validation scripts, and processed input files used directly by the models, where permitted by file-size constraints, are available in the project repository: https://github.com/RitwikGanguly/GARAGE. For public datasets not redistributed directly in the repository, the repository provides scripts or instructions to download the data from the original sources and reproduce the preprocessing workflow. The dedicated GARAGE documentation is available at https://garage-docs.readthedocs.io/en/latest/.

Computational requirements and reproducibility settings. All experiments were performed using Python-based implementations of the GARAGE pipeline. Two dedicated Conda environments–ritwik_base (Python 3.12.5) and scrna (Python 3.9.25)-were used to isolate the model-training and validation workflows, respectively. The corresponding dependency files are provided in the GitHub repository as requirements_garage.txt (model-training environment, including the core GARAGE pipeline, the state-of-the-art and single-cell-specific benchmarking baselines, the ablation studies, and the biological analysis modules) and requirements_validation.txt (validation environment, including feature selection, Leiden clustering, adjusted Rand index, normalised mutual information, macro-averaged F1-score, UMAP visualisation, Wasserstein distance, and marker-gene enrichment analysis).

The model-training environment (ritwik_base) was configured with Python 3.12.5, PyTorch 2.7.1 (CUDA 11.8), PyTorch Geometric 2.7.0, and NetworkX 3.6.1. The validation environment (scrna) was configured with Python 3.9.25, Scanpy 1.10.3 and other relevant packages present in the corresponding requirements file at GitHub.

All model-training experiments were performed on a single NVIDIA RTX A6000 GPU (48 GB VRAM, CUDA 11.8) with an AMD EPYC 7713P 64-core processor (64 threads) running Ubuntu 24.04 LTS and 314 GB system RAM. All the benchmarking datasets has been run on GPU for efficient training. Wasserstein distance computations and clustering evaluations were performed on CPU.

The repository README and the ReadTheDocs documentation (https://garage-docs.readthedocs.io/en/latest/) provide exact setup commands, per-script usage examples, expected outputs, and a step-by-step end-to-end tutorial for reproducing the core GARAGE results on the four built-in datasets (Yan, Pollen, CBMC, and Muraro). Users who wish to apply GARAGE to custom single-cell RNA-seq data are referred to the Preparing Your Data section of the documentation.