The potential of deep learning on the discovery of new genes implicated in differences of sex development

Abstract

Despite advances in understanding genetic causes of DSD (differences of sex development), the molecular cause remains unknown for over half of affected individuals. Next-generation sequencing (NGS) has improved diagnosis, but interpreting results can be challenging, especially when no known DSD gene mutations are found, or only variants of unknown significance appear. Identifying new genes involved in sex development from whole exome sequencing (WES) alone is difficult. To overcome this, we introduce “GONAD-ResNet,” a residual convolutional neural network designed to predict novel DSD-associated genes by learning complex patterns in time-dependent single-cell gene expression data. When applied to WES data from six patients (three XX, three XY) with DSD, GONAD-ResNet prioritized genes with expression profiles similar to known DSD genes while disregarding ubiquitous or irrelevant genes. This narrowed the list of potential candidates from around 1000 to a few promising novel genes per patient. This innovative approach accelerates the discovery of new DSD-related genes, opening new research avenues and potentially improving patient outcomes.

Graphical Abstract

1. Introduction

Sex development is a complex process orchestrated by numerous genes, hormones, and cellular interactions, meticulously coordinated in both time and space [2]. Disruptions in these mechanisms give rise to differences of sex development (DSD) [3], a diverse group of conditions affecting chromosomal, gonadal, and/or phenotypical sex [4]. Clinically, DSD often presents with ambiguous genitalia, infertility, and increased susceptibility to testicular or ovarian cancer [5]. Despite significant research progress, our understanding of the genetic networks governing gonad morphogenesis and sexual differentiation remains incomplete [6]. Recent advancements in next-generation sequencing (NGS) have enabled the identification of mutations in both known and novel genes among individuals with DSD. However, about half of all patients display no variants in known DSD genes, highlighting substantial gaps in our understanding of the full genetic landscape of sex development [7], [8], [9], [10], [11], [12], [13], [14], [15]. This persistent high rate of undiagnosed cases emphasizes the urgency for innovative approaches to analyze and interpret large-scale NGS datasets to improve patient care and support.

At the same time, the emergence of new technologies has led to an exponential growth of biological and medical “omics” data [16]. Identifying the most promising candidate genes within this data has become increasingly challenging. For example, when patients with DSD undergo Whole Exome Sequencing (WES), roughly 1000 variants per patient typically remain as potential contributors to the phenotype, especially if no variant are found in known DSD genes. Evaluating the possible disease relevance of each of these approximately 1000 candidates extremely demanding and time-consuming for both researchers and clinicians [17].

Artificial Intelligence (AI) has transformed biomedical research by enabling automated data analysis, pattern recognition, and predictive modeling [18]. Deep learning (DL), an advanced branch of AI, employs multi-layered neural networks to extract complex information from large datasets without explicit programming. This ability makes DL particularly valuable in biological and medical contexts, where high-dimensional data- such as genomic or transcriptomic profiles- require efficient computational analysis [19].

A major application of DL in genomics the identification of meaningful links between genetic variants and observable traits or diseases. By learning intricate gene expression patterns, DL methods support variant prioritization and improve the identification of potential disease-causing genes. In the context of rare genetic disorders such as differences of sex development (DSD), AI-driven DL models can facilitate the interpretation of high-throughput sequencing data and reveal novel candidates for experimental validation [19], [20].

This is where DL approaches such as Convolutional Neural Networks (CNN) become particularly important. DL enables computer algorithms to perform automated learning from data and assist humans in managing large volumes of multidimensional information [21]. CNN have been very successful in automated pattern recognition tasks including facial recognition [22], [23]. Applying CNN to predict gene-disease connections allows researchers to simplify the analysis of large datasets, reduce workload, and increase accuracy in identifying disease-causing variants [20]. However, as CNN become deeper, they tend to make more mistakes during training because the learning process becomes less efficient, and models may begin to memorize the training data rather than generalize from it. Residual Network (ResNet), created by Kaiming He and his team in 2015, addresses these issues by incorporating shortcut connections that bypass layers, simplifying learning and increasing training efficient [24], [25]. These shortcuts help the network focus on relevant changes instead of relearning information from scratch [26] and they preserve important signals across layers, enabling much deeper networks without problems such as vanishing gradients (signals or errors). This results in improved performance and accuracy in image recognition tasks, making ResNet very useful for complex predictions such as gene-disease studies [27], [28], [29].

Although ResNets are extremely powerful, their performance depends on access to large amount of input data. This presents challenges in areas with limited data, including rare diseases. To mitigate this limitation, we applied transfer learning (TL) by using a pre-trained model and fine-tuning it with our own gonadal dataset, thereby improving performance despite input constrains [30]. This strategy opens new possibilities for using CNN-ResNet in the context of single-cell RNA sequencing and genomics.

With our new GONAD-ResNet tool, we can reduce the list of potentially disease-related genetic variants by approximately 91 % compared to previous approaches. By prioritizing the top 12 genes, GONAD-ResNet has emerged as a powerful tool with substantial potential for discovering the genetic causes of DSD. This method not only accelerates the discovery new genes for downstream functional studies but also holds promise for improving patient outcomes in the future (Fig. 1).

Fig. 1.

This graphical abstract summarizes the methodology in 11 steps, where we start with single-cell RNA sequencing data from the human developing gonads, transform the data into images, and then predict relationships of genes with differences in sex development (single-cell RNA sequencing (scRNAseq), Cell Types and Time Points (CTTP), differences in sex development (DSD), GONAD- Residual Neural network (GOnAD-ResNet), and Residual Neural network (ResNet)).

2. Methods

In this section, we describe the complete workflow of GONAD-ResNet step by step, from data acquisition and preprocessing through image construction, model training/validation, patient-level prediction, and benchmarking. To make the pipeline easier to follow, Table 1 provides a concise summary of each major step, its aim, and the corresponding output used in subsequent analyses.

Table 1.

Overview of the GONAD-ResNet workflow from input data to prediction and benchmarking.

Step	Aim	Output
Data acquisition & preprocessing	Obtain sex-specific developmental scRNA-seq reference data and prepare it for modeling	Clean male and female expression matrices (log10-transformed; selected gonadal cell types)
CTTP matrix construction	Encode per-gene expression across cell types and developmental time points	One Cell Types and Time Points (CTTP) matrix per gene per sex
Linear interpolation	Harmonize CTTP matrices across time points and avoid missing columns due to uneven sampling	Interpolated CTTP matrices with uniform time-axis structure
RGB image generation	Convert CTTP matrices into CNN-compatible representations	Per-gene RGB heatmap images capturing temporal–cellular expression patterns
Image resizing/standardization	Ensure equal input size and efficient computation	Uniform 100 × 50 CTTP images per gene per sex
Model architecture definition	Adapt a deep CNN for binary gene classification	GONAD-ResNet (ResNet-50 + added FC layers + binary output)
Transfer learning & fine-tuning	Improve learning with limited labeled DSD genes	Trained sex-specific GONAD-ResNet models initialized from ImageNet
Cross-validation evaluation	Assess leakage-free performance on known DSD vs. non-DSD genes	Performance metrics (ROC-AUC, accuracy, stability)
Patient prediction	Prioritize candidate genes in genetically unsolved DSD cases	Ranked variant gene lists with GONAD-ResNet probability scores
Benchmarking vs. prior CNN	Compare performance to existing scRNA-seq CNN approach	Head-to-head metrics against Yuan et al. model under matched evaluation

2.1. Data source

For this study, we analysed publicly available single-cell RNA-sequencing (scRNA-seq) data from the Single-cell Roadmap of Human Gonadal Development [1], which can be directly downloaded from the Reproductive Cell Atlas (https://www.reproductivecellatlas.org/) at the raspatory “Reproductive Development, [1]”. The dataset comprises cells from human gonadal and adjacent extragonadal tissues collected between 6 and 21 gestation week (gw), covering the first and second trimesters. In this dataset, male and female samples were analysed separately, and cell types were annotated based on the expression of known markers. In total, the dataset comprised 213,898 cells for female gonadal tissue and 133,811 cells for male gonadal tissue. To optimize analysis and computational efficiency, we focused on cell populations known to be involved or affected in human sex-organ development. Consequently, we excluded erythroid, immune, neuronal, and other cell types not implicated in these processes. For our analysis, we specifically selected the following cell types (for detailed information on cell type markers and identity, see [1]): in females, coelomic epithelial cells, endothelial cells, pre-granulosa cells, germ cells, mesenchymal GATA2-positive cells, mesenchymal LHX9-positive cells, and supporting cells; in males, coelomic epithelial cells, endothelial cells, fetal Leydig cells, germ cells, mesenchymal GATA2-positive cells, mesenchymal LHX9-positive cells, Sertoli cells, and supporting cells. Raw expression values were logarithmically transformed using base 10 (log10) prior to downstream analyses.

The six Patients are part of a large international DSD cohort that was collected and analyzed as part of Sinergia grant CRSII_171007. The authors confirm that all methods were carried out in accordance with relevant guidelines and regulations. All experimental protocols were approved by a approved by the Cantonal (Vaud, Switzerland) Ethics Commission of the human research (CER-VD) (Protocol no. 42/15, 2015). The case probands were sourced from multiple clinics in Switzerland, Ukraine, Poland and Egypt. Informed consent was obtained from all subjects and/or their legal guardian(s). Individuals were considered eligible for enrolment if they had received a DSD diagnosis based on their phenotypic presentation but lacked genetic etiology.

2.2. Constructing input 2D images

To convert scRNA-seq data into red-green-blue (RGB) images, we followed several distinct steps. Initially, we generated a separate image for the expression data of each gene over time of development within either the male or female dataset. The data were organized into a two-dimensional numerical matrix refered to asCellTypes and Time Points (CTTP matrix) with rows corresponding to cell types and columns to time points (e.g., gestational week). For the female CTTPs, data were available for gestational weeks (gw) 6, 6.5, 7.5, 8, 8.6, 8.8, 9, 10, 11, 12, 13, 14, 16, 17, 18, and 21. The analysed cell types included coelomic epithelial cells, endothelial cells, pre-granulosa cells, germ cells, mesenchymal GATA2-positive cells, mesenchymal LHX9-positive cells, and supporting cells. For the male CTTPs, data were available for GW 8.8, 9, 11, 12, 14, 15, 16, 17, 19, and 21. The selected cell types were coelomic epithelial cells, endothelial cells, fetal Leydig cells, germ cells, mesenchymal GATA2-positive cells, mesenchymal LHX9-positive cells, Sertoli cells, and supporting cells. Subsequently, the CTTP matrix was populated with the average gene expression of each single cell of each cell type at each time point (Fig. 2A). To account for variations in the number of cells per cell type captured at each time point, we adjusted the matrix using linear interpolation, implemented through the Python NumPy library [31] (Fig. 2B). This interpolation applied a stretching or compressing transformation to the expression vector at each time point, ensuring uniform column length and effectively filling in data gaps without falsification, thus avoiding areas with missing information in the CTTP matrix.

Fig. 2.

Schematic representation of 2D image creation steps. First the single cell data is sorted by cell type and time point for male or female scRNA-seq data into a CellTypes and Time Points (CTTP) matrix (2 A). In the next step the matrix is adjusted with the help of linear interpolation (2 B), at last the CTTP matrix is transformed into RGB images (2 C).

Each CTTP matrix represented a specific gene for either the male or female dataset. To use these matrices as CNN inputs, we transformed each CTTP matrix into an RGB heatmap image (Fig. 2 C). CTTP values were log-compressed and normalized to an approximate 0–1 scale and then mapped to colors using a continuous colormap, such that higher relative expression intensities appear red and lower intensities appear blue, with intermediate levels represented by gradual transitions between these extremes. This procedure converts single-cell gene expression dynamics across cell types and developmental time points into standardized RGB images.

As a final step, the RGB images were resized to a height of 100 pixels and a width of 50 pixels to ensure efficient computing time.

We selected the 100 × 50 resolution after evaluating multiple image sizes to balance computational efficiency with preservation of biologically informative patterns. Importantly, this resizing does not simply discard pixels; rather, it compresses the original CTTP matrix via interpolation/local averaging so that each pixel in the resized image summarizes the corresponding neighborhood in the higher-resolution image, retaining the global structure and relative expression relationships. Consistent with the known ability of CNNs and ResNet-based models to learn hierarchical features from moderately compressed inputs [32], [33], [34], and with reports that dimensionality reduction can reduce redundancy and overfitting without harming performance [35], [36]. This resizing was achieved using the scikit-image [31] Python library version 0.20.0. First, the new dimensions in the X and Y directions were computed by multiplying the original dimensions by the scale factor:

New Height=Original Height×Scale FactorNew Height=Original Height×Scale Factor

New Width=Original Width×Scale FactorNew Width=Original Width×Scale Factor

Then, the pixel values in the resized image were determined through interpolation from the original image. As a result, all CTTP images were maintained at the same size of 100 pixels by 50 pixels.

2.3. Overall CNN structure

A deep convolutional neural network, specifically the ResNet-50 architecture [37], was adapted for the analysis of 2D images created from scRNA-seq data. ResNet-50 consists of 50 layers, including 48 convolutional layers, one Max Pool layer, and one average pool layer, with residual connections throughout. Two fully connected layers were added at the end to modify the output for binary classification.

2.4. Transfer learning

In this study, GONAD-ResNet leveraged pre-trained weights from ImageNet [21]. The model was extended with two additional fully connected layers to adapt it to the scRNA-seq data of the developing human female and male gonads. These modifications enabled the model to capitalize on the knowledge learned from generic image data and apply it to the 2D images derived from scRNA-seq data. Fine-tuning was performed to further optimize the network weights, specifically to capture gene expression patterns associated with DSD over time.

2.5. Training and test strategy

Initially, the data was divided into training and validation sets with a ratio of 70 % training and 30 % validation [38]. This approach helps prevent overfitting by monitoring the accuracy and loss on the validation set. Additionally, we employed leave-N-out cross-validation, where 25 genes were excluded from both the DSD and non-DSD group. The remaining genes were then split into 70 % for training and 30 % for testing. The training process was repeated, and the trained the GONAD-ResNet was used to predict the 25 excluded genes. This process was conducted three times with random selections to ensure the robustness and reliability of GONAD-ResNet training. Hyperparameters of the ResNet-50 model, including learning rate, batch size, and optimizer, were optimized using a combination of grid search and random search, each coupled with cross-validation, to enhance the model's performance. GONAD-ResNet was trained using Stochastic Gradient Descent (SGD) for up to 50 epochs, employing early stopping based on validation performance to avoid overfitting. The initial learning rate was set to 0.01. During training, we monitored performance metrics such as accuracy, loss, and validation error to assess the model's convergence and adjust training parameters accordingly.

All analyses and deep learning processes were conducted using Python 3.10 [39] and a UI was developed with Streamlit package version 1.27.0 [40] in python to facilitate interaction with the code. The implementation is available at https://github.com/hAzimi/GONAD-ResNet, where the authors continuously update the code and address open issues.

2.6. Gene selection for training DSD and non-DSD

For the DSD gene class, we included 165 genes previously associated with DSD (see Supplementary Table 1). 2D images were generated separately for males and females following the procedure described in “2.2 Constructing input 2D images”. From this list, we selected only the genes relevant for either the male or female training sets based on expression: genes not expressed in a given dataset were excluded, as some genes are specifically involved in male or female sex development. DSD genes showing no expression in either the male or female dataset were therefore excluded, resulting in 146 genes for the male training set and 149 genes for the female training set. Since the male and female GONAD-ResNet models were trained independently, these differences do not affect model integrity.

For the non-DSD gene class, we defined a list of 837 (Supplementary Table 1) known housekeeping genes based on studies by Caracausi et. al. (2017), Wang et. al (2019) [41], [42]. To ensure equal numbers of training examples for DSD and non-DSD categories, we randomly selected a matching number (male 146 genes, female 149 genes) of housekeeping genes for each dataset. For both male and female datasets, half of the selected housekeeping genes were highly expressed and half were low or non-expressed (based on coloration in the 2D images; see Fig. 3A&B “non-DSD”). The same procedure was applied independently for males and females, ensuring that each training set contained balanced numbers of DSD and non-DSD genes.

Fig. 3.

Representative 2D images utilized as input data for the deep learning algorithm. Fig. 3A illustrates 2D images for the female dataset, while Fig. 3B presents images for the male dataset. In both figures, the upper row showcases 2D images representing non-DSD genes, while the lower row depicts images for DSD genes.

For a detailed list of all DSD and Housekeeping genes used in the male and female GONAD-ResNet training sets, along with example 2D images for both classes, please refer to the GitHub repository: https://github.com/hAzimi/GONAD-ResNet.

2.7. Patient predictions

For prediction, we gathered variants from three XX and three XY genetically undiagnosed patients detected by whole-exome sequencing (WES) of a large international in-house DSD cohort (for the WES analysis strategy and variant filtering strategy, refer to Supplementary Figures 1.). Given the nature of the input “learning” material we exclusively selected patients diagnosed with gonadal developmental defects, i.e. gonadal dysgenesis. For each patient's variants, we created 2D images following the previously described method for all genes. These 2D images were then used as input for the GONAD-ResNet previously trained on DSD and non-DSD genes for female or male, respectively. GONAD-ResNet classified the genes into two groups: probably DSD or probably non-DSD genes, assigning each gene image a value from 0 to 1 for both categories. These predicted values for both categories were added to the variant list and utilized as additional ranking criteria. Subsequently, for each patient, the variants were ranked by Polyphen2 predicted “damaging” and DSD classification from highest to lowest. The 2D images of the top 12 ranked genes of each patient were analyzed and we chose images that shows distinct banded patterns for further validation. Utilizing KEGG pathway (https://www.genome.jp/kegg/pathway.html), Human Protein Atlas (https://www.proteinatlas.org/), Uniprot (https://www.uniprot.org/), Genecards (https://www.genecards.org/), and PubMed (https://pubmed.ncbi.nlm.nih.gov/) we assessed the potential involvement of these genes in DSD.

2.8. Benchmarking the method for performance metrics

For benchmarking the performance of our proposed model, we employed the CNN model from Yuan et.al. to our own data set [22]. The CNN model is designed for scRNA-seq data at a single time point, therefore we adapted our data and selected randomly gw 10 for female and gw 12 for male as examples and followed their protocol to transform our data into gene-pair images that could then be used as input for their CNN model. For our analysis, we labeled genes as 1 for DSD and 0 for non-DSD and paired all genes with each other. To measure the performance of the CNN model on our data we randomly removed 25 genes from DSD and non-DSD gene sets, then we trained the network on the remaining gene pairs. Finally, the trained model was used to predict the DSD or non-DSD labels of the 25 omitted genes. This process was repeated three times.

3. Results

3.1. 2D Images creation for male and female GONAD-ResNet and gene selection for training input

The process of creating 2D images for CNN input entails transforming scRNA-seq data from developing male and female gonads into RGB images. These images visually represent the scRNA-seq data organized by time and cell type as colored patterns. Figs. 3A and 3B display sample 2D images from the female and male scRNA-seq datasets, respectively. Each figure features an upper row for non-DSD genes and a lower row for DSD genes. Additionally for non-DSD genes, we initially selected housekeeping genes with either high (e.g. COX7C, UBC) or low (e.g. NACA2, ABCF2) overall expression levels, leading to predominantly red or blue images. However, due to limited genes within housekeeping genes with very low expression, in this data set, we supplemented these with genes exhibiting low to moderate expression (e.g. YARS2, MRPL27), depicted as turquoise images. This approach considered that consistently high or low gene expression in all cell types of the developing gonads may not contribute to disease manifestation. The DSD gene images used as input exhibited two main structures: colored patterns indicating varying expression levels across cell types (e.g. FOXL2, SOX9), or predominantly blue images with sporadic colored dots representing low expression in a few cells (eg. FIGLA, NR5A2). We excluded genes entirely blue in either dataset but retained those showing expression in a few cells.

3.2. Evaluation of the male and female GONAD-ResNet

GONAD-ResNet is a versatile computational framework designed for supervised single cell gene expression data analysis across various cell types and time points in the developing gonads. Based on a modified ResNet-50 architecture, GONAD-ResNet incorporates additional layers for transfer learning. Its primary objective was to distinguish between genes with potential association to DSD and non-DSD genes. To evaluate the performance of GONAD-ResNet, we conducted a rigorous 3-fold cross-validation separately for male and female datasets. We allocated 30 % of the data for testing and 70 % for training. Performance metrics included learning curve accuracy, loss, and area under the Receiver Operating Characteristic (ROC) curve. As depicted in Fig. 4, the average AUC of the ROC for 3-fold cross-validation across female (Fig. 4A) and male (Fig. 4B) datasets was calculated. Upon 50 iterations, employing early stopping to mitigate overfitting, GONAD-ResNet achieved an accuracy of 89.8 % for female and 84.4 % for male datasets in distinguishing DSD-related genes from non-DSD genes. Additionally, loss curves were monitored to ensure avoidance of overfitting to the training dataset. When comparing the performance of GONAD-ResNet to the CNN model proposed by Yuan et al. (2019) [22], GONAD-ResNet demonstrated superior accuracy in predicting genes related to DSD. Using our scRNA-seq dataset, the CNN model achieved an average AUC of 53 % for female data and 57 % for male data (Fig. 5). In contrast, our novel GONAD-ResNet approach showed an accuracy of 89.8 % for female and 84.4 % for male datasets, with significant more accurate predictions of p = 0.00011 for female (Fig. 5A) and p = 2e-05 for male (Fig. 5B)performed in triplicate.

Fig. 4.

Receiver Operating Characteristic (ROC) curves for the female (Fig. 4A) and male (Fig. 4B) datasets are presented. These curves illustrate the performance evaluation of the DSD-trained deep learning algorithm. The curve plots the True Positive Rate (Sensitivity) against the False Positive Rate (Specificity) across various decision thresholds. Each point on the curve represents a different threshold setting, indicating the trade-off between sensitivity and specificity. The diagonal dashed line represents the line of no-discrimination, indicating the performance expected by random chance. The area under the ROC curve (AUC) provides a single metric to quantify the overall performance of the algorithm, with higher values indicating better discriminative ability.

Fig. 5.

Predictive Performance Comparison of CNN [22] and our suggested method using male and female datasets. A) Box plots showing the area under the curve (AUC) of the receiver operating characteristic (ROC) for CNN (in red) and Gonad-ResNet (in blue) in predicting genes related to DSD, using female datasets. B) AUC of ROC for the male dataset, comparing CNN (in red) and GONAD-ResNet (in blue).

3.3. Visualization of genes predicted by GONAD-ResNet as DSD and non-DSD genes.

After training the female and male GONAD-ResNet using the input images depicted in Fig. 3, we proceeded to generate 2D images of three female and three male patients with gonadal dysgenesis without genetic diagnosis. 2D images were created for all genes in which patients exhibited variants.

GONAD-ResNet assigns a score from 0 to 1 for each class, 'DSD' and 'non-DSD'. For simplicity, we concentrate on the score for the predicted class. In both female and male GONAD-ResNet, genes with alternating blue and red bands or predominantly blue with scattered dots were classified as DSD genes, like the patterns in the DSD gene training set. Conversely, genes with predominantly red images were classified highest as non-DSD genes, consistent with the non-DSD gene training set where housekeeping genes were used as input. For example, female GONAD-ResNet, assigned OLFML3 of P19 a probability score of 0.99999344 for the DSD class (Fig. 6A). The corresponding 2D image exhibited a banded pattern starting with turquoise, transitioning to blue, then red, and returning to turquoise. The accompanying UMAP illustration on the right side indicates expression in Mesenchymal-LHX9 and Mesenchymal-GATA2 positive cells, and Endothelial cells (the cell type identifications are taken from the original publication of the “Single-cell roadmap of human gonadal development” [1]. Similarly, FAM71A was predicted as a DSD gene with a probability score of 0.99999595, displaying an overall blue image with scattered colored dots, and the accompanying UMAP presentation indicated overall absence of expression across all displayed cell lines. Conversely, gene UQCRB was predicted with a total probability score of 1 as a non-DSD gene. The corresponding 2D image displayed an overall red color, with the UMAP presentation indicating high expression across all displayed cell types.

Fig. 6.

Fig 6A represents 2D images predicted by the female DSD-model, while Fig 6B depicts predictions made by the male DSD model. In both figures, the two upper images correspond to genes predicted as DSD, while the lower image represents a gene predicted as non-DSD. For male and female the left images in the figure are representative 2D images created from the scRNA-seq data. The right images depict a Uniform Manifold Approximation and Projection (UMAP) presentation of the original dataset, only showing the cell-types used for this study [1].

The same pattern seems to hold true for the male GONAD-ResNet (Fig. 6B). For instance, in the case of P69, the gene CHD11 received a total probability score of 0.988354, indicating its classification as a DSD gene. The corresponding 2D image showcased a banded pattern, featuring alternating narrow red and blue stripes, transitioning into a larger red band, followed by a blue band, and ending with a narrow turquoise band. The accompanying UMAP illustration highlighted expression in cell types such as Mesenchymal-LHX9, Mesenchymal-GATA2, and Fetal Leydig cells. Similarly, gene CBLC of P49 was predicted as a DSD gene with a probability score of 0.9999838. Its 2D image exhibited an overall blue image with scattered colored dots, while the accompanying UMAP presentation depicts only a few purple dots, indication relative expression in only a few cells across all displayed cell types. In contrast, gene SON of male P74 was predicted as a non-DSD gene, garnering a total probability score of 0.9988501. The corresponding 2D image displayed an overall red coloration, and the UMAP presentation revealed high expression across all depicted cell types.

3.4. Genes predicted as DSD genes in patients

Here, we predicted the probability of genes with variants detected in three female and three male patients with gonadal dysgenesis. By ranking each patient's gene list by the predicted DSD score from highest to lowest and selecting only SNPs predicted as "damaging" by PolyPhen-2, we reduced the variant list for female patients by 10-fold from average 1150 variants to 102 and for male patients by more than 10-fold on average from 1465 variants to 127. This represents a reduction of approximately 91 % in genes needing further investigation for their potential role in the patient’s phenotype. By prioritizing the top 12 genes for both female and male patients, we streamlined the process of diagnosing and potentially discovering new DSD genes. Subsequently, amongst these top-ranked genes, we focused on genes that display distinct patterned images (Table 2, Table 3, underlined) while setting aside genes exhibiting predominantly blue 2D images with scattered colored dots (as the blue indicates negligible expression).

Table 2.

The table presents the top 12 genes identified from the analysis of three female patients using the male deep GONAD-ResNet. Underlined genenames exhibit a banded pattern on 2D images, while non-underlined genes show an all-over blue appearance. Columns include patient ID, clinical features, karyotype, gene name, chromosomal position (GRCh38), codon and amino acid changes, zygosity, allele frequency (gnomAD v4), DSD-score (probability score of DSD association), rs number and OMIM. Variants are identified by WES.

patient id	clinical features	karyotype	gene name	chromosome position (GRCh38)	codon change	amino acid change	zygosity	allel frequency (gnomAD v4)	DSD-score	rs number	OMIM
Patient 19	Gonadal dysgenesis	46, XX	OLFML3	NC_000001.11: g.113980434 C>T	NM_020190.5: c.217 C>T	NP_001273281.1: p.Arg73Trp	0/1	0.000007	0.9433915	rs766941607	610088
			GIPC2	NC_000001.11: g.78046290 C>T	NM_017655.6: c.196 C>T	NP_060125.4: p.Leu66Phe	0/1	0.000001861	0.9336186	rs1189946562	619089
			GARIN4	NC_000001.11: g.212624909 C>T	NM_153606.4: c.32 C>T	NP_705834.2: p.Thr11Met	0/1	0.008840	0.91940624	rs139614117	619852
			CAV1	NC_000007.14: g.116559034 C>T	NM_001753.5: c.284 C>T	NP_001744.2: p.Thr95Met	0/1	0.0000300	0.91736263	rs372416448	601047
			HMX3	NC_000010.10: g.124895846 C>T	NM_001105574.2: c.280 C>T	NP_001099044.1: p.Pro94Ser	0/1	0.000007	0.9104608	rs367615610	613380
			LOXL2	NC_000008.11: g.23316985 C>T	NM_002318.3: c.1600 G>A	NP_002309.1: p.Gly534Arg	0/1	0.011005	0.90307707	rs149587418	606663
			TSHZ2	NC_000020.11: g.53254963 G>A	NM_173485.6: c.1505 G>A	NP_775756.3: p.Arg502Lys	0/1	0.006430	0.89861244	rs45479792	614118
			ABCC3	NC_000017.11: g.50683692 G>A	NM_003786.4: c.3890 G>A	NP_003777.2: p.Arg1297His	0/1	0.0552409	0.8915396	rs11568591	604323
			PRDM7	NC_000016.10:g. 90060585 A>G	NM_001098173.2: c.989 T > C	NP_001091643.1: p.Leu330Pro	0/1	0.0574438	0.88863504	rs150196149	609759
			DMBT1	NC_000010.11: g.122599067 C>T	NM_001320644.2: c.3250 C>T	NP_015568.2: p.His1084Tyr	0/1	0.0346231	0.8878883	rs2277244	601969
			CARD14	NC_000017.11: g.80202245 C>T	NM_001257970.1: c.2044 C>T	NP_001353314.1: p.Arg682Trp	1/1	0.0144406	0.86727524	rs117918077	607211
			FXYD4	NC_000010.11: g.43375710 G>A	NM_001184963.1: c.188 G>A	p.Cys63Tyr	0/1	0.033846	0.86295485	rs41307500	616926
Patient 69	Gondal dysgenesys	46, XX	PIFO	NC_000001.11: g.111347753 G>A	NM_181643.6: c.239 G>A	NP_857594.2: p.Arg80Lys	0/1	0.000426	0.9363028	rs150508940	614234
			SPNS3	NC_000017.11: g.4445094 C>T	NM_182538.5: c.328 C>T	NP_872344.3: p.Arg110Cys	0/1	0.0019088	0.9213433	rs146383415	611701
			DNAJC5G	NC_000002.12: g.27276807 G>A	NM_001303128.2: c.79 G>A	NP_775921.1: p.Gly27Ser	0/1	0.005267	0.88895154	rs61754191	613946
			DSC3	NC_000018.9: g.28604374 T > C	NM_001941.5: c.716 A>G	NP_001932.2: p.Asn239Ser	0/1	0.0189470	0.8714577	rs35630063	600271
			RPL3L	NC_000016.10: g.1947003 C>T	NM_005061.3: c.784 G>A	NP_005052.1: p.Val262Met	0/1	0.019736	0.8713923	rs113956264	617416
			AK7	NC_000014.9: g.96486941 T > C, g.96478651 C>T	NM_001350888.2: c.2018T>C, c.1742 C>T	NP_689540.2: p.Leu673Pro, p.Pro581Leu	0/1 0/1	0.002008 0.0000100	0.85286933	rs116298211 rs201624359	615364
			PRG4	NC_000001.11: g.186311103 G>A	NM_001127709.3: c.3569 G>A	NP_005798.3: p.Gly1190Asp	0/1	0.000805	0.8449645	rs150072104	604283
			LPA	NC_000006.12: g.160601075 C>T	NM_005577.4: c.2969 G>A	NP_005568.2: p.Arg990Leu	0/1	0.008176	0.8406379	rs41259144	152200
			CASQ1	NC_000001.11: g.160193801 A>T	NM_001231.5: c.419 A>T	NP_001222.3: p.Tyr140Phe	0/1	0.0018377	0.82990694	rs34489853	114250
			SLC6A13	NC_000012.12: g.224060 G>A	NM_016615.5: c.1243 C>T	NP_057699.2: p.Arg415Trp	0/1	0.000162	0.82847995	rs199634252	615097
			MYO18B	NC_000022.11: g.25890783 G>A	NM_001318245.2: c.4345 G>A	NP_001305174.1:p.Asp1449A	0/1	0.0005031	0.824654	rs202224239	607295
			ADAMTS2	NC_000005.10: g.179273037 C>G	NM_014244.5: c.562 G>C	NP_055059.2: p.Glu188Gln	0/1	0.000094	0.82439125	rs146064587	604539
Patient 74	Gondal dysgenesys	46, XX	COL5A1	NC_000009.12: g.134750808 G>A	NM_001278074.1: c.1588 G>A	NP_000084.3: p.Gly530Ser	0/1	0.030107	0.96980506	rs61735045	120215
			EPN3	NC_000017.11: g.50537095 G>A	NM_017957.3: c.704 G>A	NP_060427.2: p.Arg180Gln	0/1	0.0064487	0.9510341	rs149893296	607264
			ZNF703	NC_000008.11: g.37697820 C>T	NM_025069.3: c.919 C>T	NP_079345.1: p.Pro307Ser	0/1	0.0191953	0.9337613	rs79707182	617045
			CLEC1B	NC_000012.12: g.9996905 T > C	NM_001393342.1: c.379 A>G	NP_057593.3: p.Lys127Glu	0/1	0.0000514	0.91740155	rs763938177	606783
			IRX4	NC_000005.10: g.1878211 C>T	NM_001278632.1: c.1318 G>A	NP_001265562.1:p.Asp466Asn	0/1	0.004652 (	0.90807813	rs201914951	606199
			SYTL2	NC_000011.10: g.85734322 G>C	NM_001162951.4: c.1004 C>G	NP_116561.1: p.Ser335Cys	0/1	0.0109292	0.9036023	rs74718633	612880
			ZC3H12D	NC_000006.12: g.149461959 T > C	NM_207360.3: c.317 A>G	NP_997243.2: p.Lys106Arg	0/1	0.0552348	0.9003535	rs61997220	611106
			TMEM255B	NC_000013.11: g.113801690 G>A	NM_001348663.2: c.547 G>A	NP_872420.1: p.Val183Ile	0/1	0.011235	0.8936016	rs41284482	NA
			WNT10A	NC_000002.12: g.218890289 T > A	NM_025216.3: c.682 T > A	NP_079492.2: p.Phe228Ile	0/1	0.0216768	0.891	rs121908120	606268
			NPHS1	NC_000019.10: g.35848345 C>T	NM_004646.4: c.1223 G>A	NP_004637.1: p.Arg408Gln	0/1	0.0633005	0.8864285	rs33950747	602716
			DSC3	NC_000018.10: g.31001635 G>A	NM_001941.5: c.2218 C>T	NP_001932.2: p.Pro740Ser	0/1	0.0033210	0.8714577	rs114935867	600271
			ATP10B	NC_000005.10: g.160634558 C>A	NM_001366652.1: c.1177 G>T	NP_001353584.1:p.Gly393Trp	0/1	0.018297	0.8693947	rs149397148	619791

Table 3.

The table presents the top 12 genes identified from the analysis of three male patients using the male deep GONAD-ResNet. Underlined gene names exhibit a banded pattern on 2D images, while non-underlined genes show an all-over blue appearance. Columns include patient ID, clinical features, karyotype, gene name, chromosomal position (GRCh38), codon and amino acid changes, zygosity, allele frequency (gnomAD v4), DSD-score (probability score of DSD association), rs number and OMIM. Variants are identified by WES.

patient id	clinical features	karyotype	gene name	chromosome position (GRCh38)	codon change	amino acid change	zygosity	allel frequency (gnomAD v4)	DSD-score	rs number	OMIM
Patient 13	Gonadal dysgenesis	46, XY	C7	NC_000005.10: g.40936439 T > C	NM_000587.4: c.382 T > C	NP_000578.2: p.Cys128Arg	0/1	0.0113272	0.99998856	rs2271708	217070
			FAM186A	NC_000012.12: g.50354272 A>G, g.50330667 A>T	NM_001145475.3: c.2560 T > C, c.6940 T > A	NP_001138947.1:p.Tyr854His, p.Trp2314Arg	0/1 0/1	0.033148 0.012600	0.9999753	rs77002718 rs80144666	NA
			KRT71	NC_000012.12: g.52552810 C>T	NM_033448.3: c.268 G>A	NP_258259.1: p.Gly90Arg	0/1	0.0026650	0.99995565	rs61734905	608245
			TAF1L	NC_000009.12: g.32632534 G>A	NM_153809.2: c.3046 C>T	NP_722516.1: p.Arg1016Cys	0/1	0.0273016	0.9999529	rs35905429	607798
			TSPYL6	NC_000002.12: g.54255416 G>A	NM_001003937.3: c.736 C>T	NP_001003937.2:p.Arg246Cys	0/1	0.023573	0.9999106	rs17189743	NA
			ITGA10	NC_000001.11: g.145899262 G>A	NM_001303040.2: c.2002C>T	NP_003628.2: p.Arg668Trp	0/1	0.0026750	0.99987364	rs36073645	604042
			MSLN	NC_000016.10: g.765549 T > C	NM_001177355.3: c.727 T > C	NP_005814.2: p.Ser243Pro	0/1	0.0120335	0.99978524	rs75279195	601051
			RGMA	NC_000015.10: g.93045455 T > A	NM_001166283.2: c.920 A>T	NP_001159755.1:p.Glu307Val	0/1	0.0001849	0.9997675	rs201874145	607362
			NT5DC4	NC_000002.12: g.112725234 T > C g.112722743 T > A	NM_001350494.2: c.880 T > C, c.421 T > A	NP_001380584.1:p.Ser326Pro, p.Phe141Ile	0/1 0/1	0.001592 0.005590	0.9997477	rs58616713 rs61743392	NA
			MAMDC2	NC_000009.12: g. 72723238 T > A	NM_001347990.2: c.260 T > C	NP_001334919.1: p.Ile87Thr	0/1	NA	0.99967146	NA	612879
			GRIN3B	NC_000019.10: g.1000531 C>T	NM_138690.3: c.94 C>T	NP_619635.1: p.Arg32Cys	0/1	0.000487	0.9996325	rs542744852	606651
			MAATS1	NC_000003.12: g.119732405 G>A	NM_033364.4: c.1130 G>A	NP_203528.3: p.Arg377His	0/1	0.0012674	0.99953794	rs72965006	609910
Patient 49	Gondal dysgenesys	46, XY	NPHS1	NC_000019.10: g.35839554 C>G	NM_004646.4: c.2869 G>C	NP_004637.1: p.Val957Leu	0/1	0.000724	0.99999595	rs114849139	602716
			SLCO1A2	NC_000012.12: g.21304500 T > G	NM_001386878.1: c.516 A>C	NP_066580.1: p.Glu172Asp	0/1	0.0556477	0.99998486	rs11568563	602883
			TDRD5	NC_000001.11: g.179640414 C>T	NM_001199085.3: c.1769 C>T	NP_001186014.1:p.Pro590Leu	0/1	0.004148	0.9999838	rs147735255	617748
			CBLC	NC_000019.10: g.44792407 A>G	NM_012116.4: c.1030 A>G	NP_036248.3: p.Met344Val	0/1	0.0082791	0.99996173	rs149074838	277400
			ABCB1	NC_000007.14: g. 87196176 A>G, g.87550493 C>T	NM_001348944.2: c.455 T > C, c.1199 G>A	NP_001335873.1:p.Phe152Ser, p.Ser400Asn	0/1 0/1	0.000001239 0.0329587	0.9999517	NA rs2229109	171050
			ZNF732	NC_000004.12: g.295483 C>T	NM_001137608.3: c.181 G>A	NP_001131080.1:p.Glu61Lys	0/1	0.0535233	0.9999131	rs112132883	NA
			CDH11	NC_000016.10: g.64982262 C>G	NM_001308392.2: c.1039 G>C	NP_001788.2: p.Val347Leu	0/1	0.01900	0.99988484	rs76181686	600023
			TSHZ3	NC_000019.10: g.31278387 T > C	NM_020856.4: c.1406 A>G	NP_065907.2: p.Glu469Gly	0/1	0.0088348	0.9998796	rs143453460	614119
			SVEP1	NC_000009.12: g.110407495 T > C	NM_153366.4: c.8105 A>G	NP_699197.3: p.Asp2702Gly	0/1	0.024024	0.99985003	rs111245230	611691
			PKP3	NC_000011.10: g.397331 G>C	NM_001303029.2: c.830 G>C	NP_009114.1: p.Arg277Pro	0/1	0.00003	0.9998235	rs200371913	605561
			CHIT1	NC_000001.11: g.203225751 C>T	NM_001256125.2: c.175 G>A	NP_003456.1: p.Ala59Thr	0/1	0.000035	0.9997445	rs113204979	600031
			CMKLR1	NC_000012.12: g.108292778 A>C	NM_001142343.2: c.185 T > G	NP_001135815.1:p.Ile62Ser	0/1	0.0055270	0.99999595	rs141421422	602351
Patient 131	Gondal dysgenisis	46, XY	ZNF732	NC_000004.12: g.271758 A>G	NM_001137608.3: c.1099 T > C	NP_001131080.1:p.Cys367Arg	0/1	0.0144415	0.9999517	rs150738695	NA
			GLB1L3	NC_000011.10: g.134312359 C>T g.134314335 G>A	NM_001080407.3: c.1298 C>G, c.1673 G>A	NP_001073876.2:p.Ser433Trp, p.Arg558His	0/1 0/1	0.001333 0.002796	0.9999486	rs200082150 rs201605779	NA
			SCARF1	NC_000017.11: g.1635267 G>A	NM_003693.4: c.1984C>T	NP_003684.2: p.Arg662Trp	0/1	0.021973	0.9999219	rs8072430	607873
			SCN3A	NC_000002.12: g.165140923 G>A	NM_001081676.2: c.1747 C>T	NP_001075145.1:p.Arg583Trp	0/1	0.0000043	0.9999025	rs745365212	182391
			CD109	NC_000006.12: g.73785452 T > C	NM_001159587.3: c.2312 T > C	NP_598000.2: p.Phe771Ser	0/1	0.0001841	0.9998241	rs139108193	608859
			TRIM31	NC_000006.12: g.30110488 A>G g.30112756 G>C	NM_007028.5: c.704 T > C, c.50 C>G	NP_008959.3: p.Leu235Pro p.Pro17Arg	0/1 0/1	0.0019843 0.0019898	0.999818	rs35775852 rs36063651	609316
			POU4F3	NC_000005.10: g.146340176 C>T	NM_002700.3: c.749 C>T	NP_002691.1: p.Ala250Va	0/1	0.000019	0.99978215	rs371875449	602460
			UBASH3A	NC_000021.9: g.42404027 C>T	NM_001001895.3: c.82 C>T	NP_061834.1: p.Leu28Phe	0/1	0.0556854	0.99977404	rs2277800	605736
			STX19	NC_000003.12: g.94015157 G>A	NM_001001850.3: c.113 C>G	NP_001001850.1:p.Ala38Gly	0/1	0.003608	0.99970573	rs61739250	NA
			ABCC2	NC_000010.11: g.99836218 G>T	NM_000392.5: c.3542 G>T	NP_000383.2: p.Arg1181Leu	0/1	0.001692	0.9997048	rs8187692	601107
			CRISP2	NC_000006.12: g.49695854 A>G	NM_001142407.3: c.586 T > C	NP_003287.1: p.Cys196Arg	0/1	0.0188155	0.99969375	rs36069724	187430
			FN1	NC_000002.12: g.215370410 G>A	NM_001365522.2: c.6737 C>T	NP_997647.2: p.Thr2246Ile	0/1	0.0000021	0.99966085	rs1416042517	135600

In female patients for example, OLFML3, which is top-ranked for Patient 19, shows expression in a small proportion of pre-granulosa cells starting from gw 8 and is highly expressed in most mesenchymal-LHX9 (essential for genital ridge formation [2]) and mesenchymal-GATA2 cells (which help to establish the environment that helps guide Sertoli or granulosa cell specification [2]), with expression gradually increasing from gw 6–21. Genes such as PDRM7, FXYD4, DNAJC5G, and AK7 are predominantly expressed in germ cells starting from gw 10, shortly after the beginning of ovarian differentiation (gw 8 [6]). Interestingly, AK7, a gene that is among the top 12 ranked genes for patient 69, harbored two mutations, both predicted to be damaging. AK7 was previously linked to primary male infertility [43] and is mentioned by the disease-phenotype association tool “DISEASES” in the context of Bardet-Biedl Syndrome and infertility (https://diseases.jensenlab.org). WNT10A, which codes for a secreted signal protein of the WNT family, is predominantly expressed in the coelomic epithelium (the origin tissue of the gonad [6]) from gw 6, with its expression increasing through week 21. It is also expressed to a lesser extent in pre-granulosa cells starting from gw 8. PIFO is expressed in pre-granulosa cells starting from gw 8 and in germ cells from gw 6, with its expression decreasing through gw 21. PIFO is associated with Alström syndrome, which is linked to endocrine disruptions such as hypergonadotropic hypogonadism. ZNF703 is predominantly expressed in supporting cells, a novel embryonic gonadal cell population that gives rise to either granulosa cells or Sertoli cells [44], as well as in pre-granulosa cells throughout gw 6–21.

For male patients, GONAD-ResNet primarily prioritized genes expressed in germ cells, Sertoli cells, endothelial cells, coelomic epithelium, fetal Leydig cells, mesenchymal-LHX9 cells, and mesenchymal-GATA2 cells, either throughout all gw of gonadal development or during specific periods. For example, C7 is increasingly expressed in mesenchymal-GATA2 cells from gw 7–21, while its expression in mesenchymal-LHX9 cells decreases from gw 6–21. Additionally, it is expressed in fetal Leydig cells starting from gw 7–8, which is a crucial time point in testis differentiation [6]. The same is true for CDH11 which is expressed in fetal Leydig cells from gw 8–21. Mutations in CDH11 are a known cause for Elsahy–Waters Syndrome which presents among other symptoms with hypospadias [45], [46]. Another group of genes that were prioritized by the machine were genes predominantly expressed in endothelial cells, for instance CMKLR1 whose expression is mostly present at gw 14, or gene CD109 which is expressed between gw 6–9 and FN1 whose expression is starting from week 6 and increasing until week 21. Endothelial cells have previously been shown to play a crucial role in the germline stem cell niche in male mice [47]. The last group of genes that got prioritized by the machine are genes that are mainly expressed in germ cells for instance SLCO1A2, TDRD5, RGMA and ZNF732. Interestingly, TDRD5 which is expressed during gw 7–21 was previously associated with azoospermia and infertility by the disease-phenotype association tool “DISEASES” (https://diseases.jensenlab.org).

4. Discussion

In this study, we introduce a novel method that leverages the pattern recognition capabilities of deep learning algorithms, specifically CNN-ResNet, to predict unknown gene-disease associations in human sex development, in a way comparable to how face recognition systems identify patterns.

GONAD-ResNet is built on a large collection of scRNA-seq data from developing human male and female gonads, spanning gw 6–21 [1].

This dataset is used to train a CNN-ResNet, enabling it to learn the unique spatial and temporal gene expression signatures associated with sex development and its differences referd to as GONAD-ResNet. We demonstrate that GONAD-ResNet can identify previously unknown associations between genes and DSD in genetically undiagnosed patients by detecting their distinctive expression profiles during gonadal development.

To train the CNN-ResNet on scRNA-seq data, it was necessary to convert the tabular dataset into images. For each gene and sex, the data was transformed into heatmap representations, providing a novel and robust input for CNN models. This enabled our GONAD-ResNet to effectively learn and interpret gene expression patterns critical for identifying novel genes with similar expression profiles within the developmental dataset for male and female separately.

After transforming known DSD associated genes into 2D images, we excluded genes from the male and female gene lists that showed zero expression, based on the assumption these genes are less involved in sex development within the timeframe of this study (gw 6–21). For non-DSD genes, we chose housekeeping genes under the assumption that disruptions in these genes are less likely to contribute to those specific phenotypes. Furthermore, using housekeeping genes provided the advantage of selecting well-known and well-studied genes, minimizing the likelihood of their involvement in sex development. This strategy allowed GONAD-ResNet to clearly distinguish between genes associated with DSD and those that are not.

Due to the limited number of genes -and thus limited 2D images- available for training GONAD-ResNet in the DSD context, we employed a pre-trained CNN-ResNet network as the backbone model and applied transfer learning (TL) to tailor GONAD-ResNet specifically to DSD and increase its selectivity. This approach is particularly advantageous in biological research, where data acquisition is often labor-intensive, expensive, and limited. Using a pre-trained model allowed us to jump-start the DSD-specific analysis and improve GONAD-ResNet’s performance and accuracy despite the limited datasets.

With our innovative approach, we achieved a prediction accuracy of 89.8 % for the female dataset and 84.4 % for the male dataset in distinguishing DSD-related genes from non-DSD genes. In contrast, when applying the model proposed by Yuan et al. [22], which also uses a convolutional neural network to predict disease-related genes from scRNA-seq data, GONAD-ResNet demonstrated superior performance for our dataset: GONAD-ResNet F: 89.8 %; M: 84.4 % vs. CNN F: 53 %; M: 57 %.This difference might be attributed to the design of the original CNN architecture, which is tailored to gene networks rather than spatial and temporal expression data, potentially limiting its performance in developmental diseases.

A comparison between the recently introduced STIGMA approach and our GONAD-ResNet is warranted STIGMA [46], a single-cell tissue-specific gene prioritization framework learns temporal gene-expression dynamics across cell types and integrates gene-level features. STIGMA effectively models developmental trajectories and cell-population heterogeneity; however, it represents genes as vectors or trajectories and does not explicitly exploit local spatial relationships among features. In contrast, GONAD-ResNet embeds gene attributes into 2D feature maps and applies a convolutional ResNet architecture, enabling the model to detect motif-like interactions among neighboring features. Combined with transfer learning to accommodate the limited number of labeled DSD genes, this design allows GONAD-ResNet to capture complex feature-interaction signatures that may complement trajectory-based approaches such as STIGMA.

Both the male and the female GONAD-ResNet successfully prioritized genes within the variant lists in DSD patient cohort that displayed patterns consistent with the DSD-training sets. The prioritized genes showed distinct expression in critical cell types such as Sertoli and Granulosa cells, as well as at crucial time points for human gonadal development (gw 8–15) [6].

One limitation of the model is that GONAD-ResNet prioritized a subset of genes with predominantly blue images as DSD-associated. We assume these genes may have a lower likelihood of being involved in a patient`s phenotype due to their negligible expression in early female and male gonadal tissues. However, GONAD-ResNet likely prioritizes these genes because a subset of the DSD training set images also displayed predominantly blue patterns. We anticipated this possibility when selecting DSD gene images as input. To avoid introducing excessive bias or overly reducing the training set, we included images that were predominantly blue (only excluded images with zero expression). Despite their sparse or brief expression in the tissue, the disruption of these genes might still contribute in the patient's phenotype and could be considered within second-tier strategy.

To address this limitation, future versions of the model could incorporate weighting schemes or normalization strategies that consider for the overall expression intensity of each gene across developmental stages. Another possibility would be to integrate additional modalities- such as temporal gene co-expression patterns or functional annotations- to help the model distinguish biologically relevant low-expression genes from those included due to image similarity. Such enhancements would improve model interpretability and reduce the influence of expression sparsity on gene prioritization.

Another limitation inherent to the data source is that single-cell RNA sequencing may fail to detect very low-abundance transcripts, potentially leading to an underestimation of genes with also subtle but potentially biologically relevant expression.

Another limitation inherent to the data source is that single-cell RNA sequencing may fail to detect very low-abundance transcripts, potentially underestimating genes with subtle but biologically relevant expression.

GONAD-ResNet also learned the patterns of the non-DSD category images. Genes with predominantly red images (high expression across all cells and time points) were classified as non-DSD genes with the highest probability, consistent with the highly expressed genes in the non-DSD training set, where housekeeping genes were used. Predominantly blue images selected as non-DSD were also classified as non-DSD, although with slightly lower probability - likely due to the presence of predominantly blue images in DSD category as well.

Although ResNet-based models typically benefit from large and diverse training sets, integrating multi- single-cell transcriptomics datasets is only reliable when studies are homogeneous in both -omic modality and cell-type definitions. In practice, many public datasets differ in sequencing platforms, preprocessing pipelines, and annotation strategies, meaning that direct pooling can introduce batch effects and label inconsistencies that obscure biological signals and ultimately degrade model performance. For this reason, we trained and evaluated GONAD-ResNet on the Garcia-Alonso et al. [1] dataset, which offers high-quality single-cell profiles with consistent annotation across relevant gonadal cell types. Expanding GONAD-ResNet with additional compatible datasets remains promising direction, but it would require careful harmonization -such asbatch correction and mapping to standardized cell-type ontologies- followed by retraining and independent validation. To support future extensions, the project’s code base and workflow are available on the paper’s GitHub page, enabling other researchers to collaborate on or adapt the pipeline as new datasets become available.

In summary, our original approach demonstrates that ResNet can serve as a powerful tool for interpreting genomic data, reducing patient potentially clinically relevant variant lists by 90 %. We developed an innovative method for transforming scRNA-seq data into 2D images and successfully used these images to train a ResNet within a biological disease context, specifically DSD. Our results show that GONAD-ResNet enables efficient identification of unknown and novel genes potentially associated with gonadal developmental defects. GONAD-ResNet can also be applied to other developmental diseases, provided that a similar time-dependent scRNA-seq dataset is available to construct comparable 2D images. Identifying such candidate genes is the first step toward validating their role in experimental animal and stem cells models, ultimately accelerating the discovery of disease mechanisms and improving patient care in the future (see Fig. 1).

Author statement

In response to the reviewers’ valuable feedback, we have thoroughly revised our manuscript. The language has been clarified throughout, and all comments have been addressed in detail in the rebuttal document. To improve the transparency and readability of our workflow, we have added a new Table 1 at the beginning of the Methods section, summarizing the main steps of our protocol together with their aims and outcomes.

All authors have reviewed and approved the revised manuscript and support its resubmission to Computational and Structural Biotechnology Journal. The authors declare no conflicts of interest.

Ethical statement

This study was conducted in accordance with the principles outlined in the Declaration of Helsinki. All procedures involving human participants and their genetic material were reviewed and approved by the Cantonal (Vaud, Switzerland) Ethics Commission of the human research (CER-VD) (Protocol no. 42/15, 2015). Written informed consent was obtained from all participants or their legal guardians prior to sample collection. All DNA samples were anonymized to ensure the privacy and confidentiality of participants. Data were handled and stored in compliance with applicable data protection regulations. Participants were informed about the purpose of the study, and their right to withdraw at any time without consequences.

CRediT authorship contribution statement

Von der Decken Isabel: Writing – review & editing, Writing – original draft, Visualization, Validation, Software, Resources, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Anna Lauber-Biason: Writing – review & editing, Supervision, Resources, Project administration, Funding acquisition, Conceptualization. Hamid Azimi: Writing – review & editing, Writing – original draft, Visualization, Validation, Software, Methodology, Investigation, Formal analysis, Data curation, Conceptualization.

Declaration of Generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the author(s) used ChatGPT to improve the language and readability of the manuscript, as English is not our native language. The ideas, research, and scientific content presented in this work are entirely original and were developed solely by the author(s). After using this tool, the author(s) carefully reviewed and edited the content as needed and take full responsibility for the final version of the publication.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary material

Supplementary material

Supplementary material

Data Availability

The datasets analysed during the current study are available online https://www.reproductivecellatlas.org/ published by [1]

The WES data from the DSD patients analysed during the current study available from the corresponding author on reasonable request. All code for generating 2D images, as well as training and testing the model, is available at https://github.com/hAzimi/GONAD-ResNet, and the authors continuously address open issues.