Species-agnostic transfer learning for cross-species transcriptomics data integration without gene orthology

Abstract

Novel hypotheses in biomedical research are often developed or validated in model organisms such as mice and zebrafish and thus play a crucial role. However, due to biological differences between species, translating these findings into human applications remains challenging. Moreover, commonly used orthologous gene information is often incomplete and entails a significant information loss during gene-id conversion. To address these issues, we present a novel methodology for species-agnostic transfer learning with heterogeneous domain adaptation. We extended the cross-domain structure-preserving projection toward out-of-sample prediction. Our approach not only allows knowledge integration and translation across various species without relying on gene orthology but also identifies similar GO among the most influential genes composing the latent space for integration. Subsequently, during the alignment of latent spaces, each composed of species-specific genes, it is possible to identify functional annotations of genes missing from public orthology databases. We evaluated our approach with four different single-cell sequencing datasets focusing on cell-type prediction and compared it against related machine-learning approaches. In summary, the developed model outperforms related methods working without prior knowledge when predicting unseen cell types based on other species’ data. The results demonstrate that our novel approach allows knowledge transfer beyond species barriers without the dependency on known gene orthology but utilizing the entire gene sets.

INTRODUCTION

Model organisms such as mice and zebrafish play a crucial role in developing and validating new hypotheses in biomedical research, particularly in studying disease mechanisms and treatment responses. However, due to biological differences between species, translating these findings into human applications remains challenging. Therefore, additional intensive experimental validation is required to confirm whether the knowledge gained from one species can be translated into humans [1]. Although researchers may generate similar data on different model organisms during this process, it is often infeasible to integrate and analyze due to data heterogeneity. This data heterogeneity originates from different sources, such as signal noise, batch effects, technological heterogeneity and biological heterogeneity. This presents a challenging task in data integration, requiring domain-specific noise handling, a multi-modal analysis methodology capable of handling different data types [2] or a cross-species transfer learning methodology [3]. In biomedical research, biological heterogeneity where each species has a different set of genes responsible for certain biological functions is a critical hurdle to be addressed. The current approach to handle this issue is to utilize gene orthology information. In this inter-species data integration, biological heterogeneity necessitates a gene-id conversion with human-curated data such as an orthologous gene. However, this approach faces two major challenges: a many-to-many mapping for various gene families and severe information loss during the gene-id conversion.

In single-cell RNA sequencing data analysis, one of the most frequent tasks is cell type classification and identification of cell clusters. This task is often done by latent feature analysis through dimensionality reduction and feature selection of scRNA-seq data. The best practice in single-cell RNA sequencing analysis suggests performing a dimensionality reduction using PCA after standardized data preprocessing with normalization, as suggested by Seurat [4] and others [5]. More recently, different deep neural network approaches were also developed for scRNA-seq data analysis. For instance, a Deep Count Autoencoder network (DCA) is designed to denoise scRNA-seq datasets [6]. The DCA model captures gene–gene dependencies from count matrices with a negative binomial noise model. This autoencoder (AE)-based model can compress gene expression profile to low-dimensional representations and captures the underlying true data manifold. The Single-Cell Embedded Topic Model (scETM) approach applied a topic modeling concept from natural language processing to embed the gene expression profile of single-cell RNA-seq data. Similar to DCA, scETM can compress gene expression into low-dimensional features [7].

Subsequently, in cross-species data analysis, cell-type label projection is one of the main tasks for recent transfer learning approaches. TransCompR utilizes a PCA-based method for the interspecies translation model. It aligns feature space with gene homology and integrates knowledge of mouse and human data. It shows a better prediction of human Crohn’s disease on mouse’s latent space [8]. The recent bone-marrow cell type classification study with mice and humans found that transfer learning can improve classification accuracy when the target data are limited in a small sample size [9]. They utilized orthologous gene information, representing meta-information about genes in two species. scAdapt [10] implemented an adversarial domain adaptation network for the cross-species label projection. In the scETM (Single-Cell Embedded Topic Model) work, the features of two different species are transformed by orthologous gene pairs. Although this work showed the possibility of cross-species transfer learning, this is limited to homogeneous domain transfer learning using orthologous gene set information [7].

Transfer learning approaches are commonly differentiated into two categories: (1) homogeneous and (2) heterogeneous domain adaptation (HDA). In homogeneous transfer learning, the domains share the same feature space but have different distributions. Subsequently, most of the recent transfer learning approaches focus on homogeneous transfer learning, when a gene ID needs to be mapped to the same IDs in a different species with gene homology. Heterogeneous transfer learning, in contrast, assumes that the domains originate from different feature spaces and distributions [11].

The SATL method falls into the category of heterogeneous transfer learning and thus is closely related to generalized zero-shot learning (GSZL) and HDA. The GZSL is mainly studied in computer vision and aims to integrate text data for image classification. Therefore, during the training phase, the text and image features are projected into an aligned feature space. One of the most widely used approaches is the CADA-VAE model, which is developed to learn shared latent representations of different modality domains, text and images [12]. This GZSL methodology, which involves multi-modal analysis by finding common latent spaces, has fundamental similarities to an HDA task with a manifold-alignment approach [13]. Moreover, the LPJT is developed for unsupervised domain adaptation [14]. The DDACL (Discriminative Distribution Alignment with Cross-entropy Loss) reduces distribution divergence and maximizes distances between class centroids during training [15]. Thus, similar to heterogeneous transfer learning, these approaches could allow knowledge integration and translation across datasets of various species without relying on gene-ID conversion or external knowledge. We developed a novel approach called Species-Agnostic Transfer Learning (SATL) that identifies a common latent space for predicting out-of-sample data from different species without relying on external knowledge. The approach is built on the Cross-Domain Structural Preserving Projection (CDSPP) method, where the model learns a projection matrix for a domain-invariant feature subspace to reduce the discrepancy between domains [16]. This allows the incorporation of the entire dataset in cross-species analysis. Finally, in order to benchmark the performance of the newly developed SATL method in the area of heterogeneous transfer learning, we evaluated it in comparison to approaches closest to it, GSZL as well as HDA.

MATERIALS AND METHODS

Dataset and preprocessing

To show cross-species knowledge transfer, we analyzed four datasets. The first dataset originates from an lps-stimulation experiment with four species, mouse, rat, rabbit and pig (E-MTAB-6754) [17]. There are three biological replicates for each species, and each dataset has four different cell-differentiation statuses. The other three datasets comprise single-cell sequencing data with cell labels of bone marrow [9] (E-MTAB-8629 and E-MTAB-8630), pancreas [18] and brain (https://portal.brain-map.org/) both from human and mouse. Each dataset has 11 different cell populations. For the test with an individual external test set, two additional human pancreas datasets are used (GSE85241 and E-MTAB-5061) [19, 20]. Prior to the main integration step, all cross-species datasets are processed with two steps, conventional single-cell RNA-seq preprocessing and feature extraction.

The following steps are employed during the single-cell RNA-seq preprocessing. First, cells with fewer than 200 expressed genes were discarded, as these are often of low quality due to their low gene capture rate. Only genes expressed in at least five cells were considered to reduce the number of features [21]. Cells with inflated gene counts, typically caused by doublets during sequencing [22], were also filtered out. In addition, cells with high counts for mitochondrial genes, indicative of perforated cells, were excluded from the analysis [23]. The gene count matrix is normalized with ‘Total-count with 1e6’ and log-transformed. Table 1 shows details of the preprocessed datasets of human and mouse used in this study. For the utilized lps-stimulated cell dataset, Table 2 shows details of the preprocessed dataset.

Table 1.

Description of datasets after quality control. Datasets are used for the comparison of SATL performance in cell type label projection analysis

Tissue	Species	Total cells	Genes	Cell types
Bone marrow	Human	9234	20 887	11
Bone marrow	Mouse	5504	14 909	11
Pancreas	Human	8567	15 858	11
Pancreas	Mouse	1883	13 653	11
Brain	Human	9564	29 499	11
Brain	Mouse	15 396	22 428	11

Table 2.

Description of mononuclear phagocyte stimulation dataset used for SATL comparison with MNN and common latent space analysis. It shows the number of genes after single-cell preprocessing

	mouse1	mouse2	mouse3	rat1	rat2	rat3
#Gene	12 962	13 070	12 826	13 350	12 821	12 613
unst	2126	3519	2936	2650	1748	1960
lps2	2537	4321	3169	2094	2682	1962
lps4	2366	3293	3536	2779	1718	1079
lps6	1703	2126	1453	3909	1362	2526
	rabbit1	rabbit2	rabbit3	pig1	pig2	pig3
Genes	10 899	10 694	10 853	12 041	11 846	12 121
unst	3773	3275	3399	2221	1748	2179
lps2	3690	1820	4401	1630	1614	2353
lps4	2965	1522	2163	1948	1899	2758
lps6	4196	1660	1664	1941	1381	1797

The feature extraction methods in this study were informed by recent standard analysis pipelines and encompassed both linear and non-linear feature extraction models. Specifically, we employed several methods, including the Seurat pipeline [4] as the standard method, DCA [6] as a de-noising AE and scETM [7] for a deep feature extractor. These methods were chosen due to their established efficacy in de-noising and feature extraction. The scripts used for each preprocessing and feature extraction can be found in the codebase. Details of feature extraction are available in Supplementary materials. We compared the performance of SATL with related machine learning models.

SATL

SATL utilizes the common labels between two species datasets (seen classes) to align the latent spaces and predicts out-of-sample class labels (unseen classes). The general workflow of SATL is shown in Figure 1. The single-cell RNA sequencing data are preprocessed and normalized, and the gene expression profile is converted into lower-dimensional features. For dimensionality reduction, different approaches, such as principal component analysis (PCA) or AE-based models, can be employed (see Data and preprocessing section for details). Subsequently, the feature spaces of both species are integrated with HDA methods. During this step, a set of common cell labels available in both species (also called seen classes) are incorporated and used to align the two different feature spaces into a common latent space. In order to deal with unseen classes and the class imbalance in the test set, SATL employed a two-step pseudo-labeling scheme (see Supplementary Method 1.3 for details). This algorithm is inspired by recent works [24, 25]. This procedure considers that the initial model is strongly biased toward the source domain in the zero-shot setting. Therefore, we use an initial burn-in phase for out-of-sample classes. In the first step, only predictions for unseen classes are considered. Therefore, in each iteration, an increasing number of test samples with their predicted pseudo-labels are added to the training step. In the second step, the process is repeated on all class predictions. Finally, in the common latent space, it is possible to predict the labels of the target species with the transferred model that is trained on the source species. For a more detailed method description, see the Methods section in the supplementary material.

Figure 1.

Overview of cross-species single-cell data integration and label-projection utilizing the developed SATL workflow. Different species’ gene expression profiles can be used without gene-id conversion. Extracted features from each species are not in the same domain space. HDA aligns these different domains into a common latent space with a projective matrix (P_{source/target}). The latent space is further aligned with two pseudo-labeling steps. During the first pseudo-labeling step, the latent space is updated with the train set and only unseen classes of the test set. The second pseudo-labeling step utilizes all classes in the test set for the pseudo-labeling and further updates the alignment of latent spaces. In this common space, unlabeled data in the target species can be predicted with knowledge of the source species.

Task details: out-of-sample label prediction with a latent feature from single-cell sequencing data

The learning and testing scheme follows a transductive generalized zero-shot learning (GZSL) setup where the labels of a certain fraction of classes in the target domains are unknown during training. The machine learning model has fully labeled data from a source domain. However, in the target domain, only a partial set of classes is labeled and available during the training step. The main task of this work is an out-of-sample prediction in a target species. Without gene-ID conversion, which incurs severe information loss (Figure 2), we first extract features and then integrate the features of different species to project labels from one species to another.

Figure 2.

Comparing the strategy of previous approaches with the newly developed SATL method. (A) Previously, transfer learning across species data is done using external gene homology data. However, this approach often results in severe information loss. (B) In the proposed SATL method, a feature extractor is used to embed a gene set into latent features. After feature extraction, heterogeneous features are aligned using an HDA algorithm. The aligned features can then be used directly in the integrative analysis.

In particular, for the task with sequencing data, the domain is the species, and the data are the gene expression profile. Domain heterogeneity comes from biological heterogeneity, as different species have different sets of genes, and the overlapping functions/genes between species vary depending on the evolutionary distance of species. The label projection with single-cell sequencing data is done by cell types (Bone Marrow, Brain, Pancreas dataset) or experimental labels (LPS-stimulation dataset). More information is available in the supplementary method.

Evaluation metric

The performance measures used in this work are balanced accuracy for each predicted seen and unseen class as well as the h-score thereof [26]. Average accuracy may mislead generalization performance when a dataset is highly imbalanced and the classifier is biased, which is common in real-world biological datasets [27]. Thus, we used balanced accuracy (bAcc). It is the average of all recalls calculated for each class. Recall is a true-positive rate for each class. When the number of classes is C,

(1)

(2)

(3)

bAcc_unseen is the balanced accuracy for the prediction of unseen classes. Likewise, bAcc_seen is balanced accuracy for the prediction of seen classes. It is calculated with scikit-learn ‘balanced_accuracy_ score’. This ensures that a low accuracy for unseen classes or for seen classes leads to a low score. Balancing the class-wise accuracies is of particular importance since the data sets are highly imbalanced. Additionally, both unseen and seen class accuracies are of interest.

We tested entire scenarios with different splits of unseen and seen classes. Thus, balanced accuracy and h-score in this work are averaged scores of all possible split scenarios.

Feature importance and gene ontology analysis

The feature importance is obtained by using the explained variance from PCA and the weights of the projection matrix for each species. With this approach, it is possible to investigate which gene contributes more to the common latent space for cross-species analysis in the SATL model. After obtaining the list of genes, we sorted them based on the absolute value of weight and obtained the top 100 contributing genes. The gene lists were further processed using gene ontology analysis for the Gene Ontology (GO) biological process. We used gProfiler 0.77 [28] and PantherDB 17.0 [29] for gene ontology analysis.

RESULTS

While there is a large set of bioinformatics and data science methodologies that allow the data-driven integration of biological and biomedical datasets, data-driven knowledge transfer between species only recently became possible with novel transfer learning approaches. However, most of these studies have a limitation: they need to employ external data to homogenize heterogeneous feature spaces in order to transfer learned models among different domains. One of the biggest issues with this methodology is the inevitable loss of information during the feature conversion based on gene orthology information [30–33]. Thus, we are presenting a new methodology to investigate biological knowledge in a completely data-driven way without severe information loss.

SATL adapts heterogeneous domains of feature spaces from different species

The main goal of our work is to explore an alternative approach to transfer classifiers beyond the barriers between species. Therefore, we investigated different HDA methodologies and related fields such as domain adaptation (DA) and GZSL for their capabilities to model shared cell types in single-cell sequencing data across two species. In contrast to the current approaches, no information about orthologous genes will be used. Instead, the newly developed SATL method will align the data space of the source and target domain with the aid of a subset of labeled data from the target domain and source domain. A detailed description of the methodology is provided in the supplementary material.

The main idea of SATL is to extend a semi-supervised HDA algorithm to enable GZSL on heterogeneous biomedical data sets, such as single-cell sequencing data from different species, for instance. The evaluation scheme, therefore, follows the strategy of a transductive GZSL setting, where the labels of a certain fraction of classes of the target species are unknown during training. In the following results, we demonstrate and evaluate cross-species transfer learning. At first, with lps-stimulated macrophage dataset from the four species, we show how the proposed method works and how could the result be interpreted in a biologically meaningful way by functional analysis. SATL is compared with mutual nearest neighbors (MNNs), 3.2, and identified latent space is investigated, 3.3. After that, we further evaluated the performance of SATL with other models having a similar concept and purpose. Three human–mouse pair datasets are used for this comparison. This comparison works contains comparisons to GZSL models and DA 3.5 methods. The detailed results from three single-cell sequencing data are also investigated, 3.4.

SATL outperformed mutual nearest neighbors in HDA task

Here, we explore the performance of SATL in comparison to the MNN, a popular method for single-cell data integration and batch effect mitigation. While the MNN is intended for homogeneous domain adaptation, requiring the same order and genes, the methodology is closely related to CDSPP. For the cross-species integrative analysis without gene-ID conversion, we first extract features from the species with different numbers of genes. These features consist of different latent spaces for each species dataset, each corresponding to heterogeneous domains. We compared SATL to MNN in an HDA task, with the lps-stimulated macrophage dataset [34].

Previously, scGEN has shown that vector arithmetics on latent space can predict gene signatures of unseen cell types in different species in the lps-stimulated macrophage dataset [35]. Likewise, we analyzed SATL on the same dataset. The stimulation dataset has labels about experimental conditions instead of cell types. Training and testing are done in the same procedure as other single-cell datasets. There are four species and each species has four labels, unst, lps2, lps4 and lps6. Compared to the cross-species analysis of the original work, where only 2336 genes were investigated [17], and the scGEN study, where 6619 genes were investigated [35], our analysis can target all expressed genes from both species, see Materials and methods for details. In this analysis, analyzed genes were 10 000–13 000 genes after single-cell RNA data preprocessing (Table 2).

Here, one group of cells from target species is masked and predicted based on the model trained with data from source species. The evaluation is done in all pair-wise scenarios among all four species. At first, we extracted features using PCA and applied MNN correction to mitigate batch effects and align the feature spaces, assuming that the two species had homogeneous domains. Subsequently, we trained a random forest (RF) model on the processed source species data to predict cell types of target species (Figure 3). Simultaneously, SATL was employed on the same features extracted from PCA. In most of the scenarios, our SATL model predicts the masked cells with the correct label. However, the comparison between both approaches shows that, the RF model trained with MNN-corrected features, designed for homogeneous domain adaptation, cannot correct heterogeneous domains in cross-species data integration tasks. In most of the scenarios, it failed to predict the masked cell type of the target species. Additionally, we extracted locality preserving projection (LPP) features instead of PCA features and combined them with the MNN and RF classification approach. However, no improvement in prediction was achieved (Supplementary Figure S1).

Figure 3.

Comparison of out-of-sample label prediction with heterogeneous features. The xy-plot shows the results of the cross-species integration analysis comparing SATL with the MNN approach for HDA. For evaluation, each RF model is trained on the source and tested on the target data with a masked label across species.

SATL prediction models identify aligned latent spaces with similar biological function

As shown in section SATL, SATL models allow for the prediction of unseen classes in cell type labels across species. In order to evaluate the biological context of the aligned latent feature spaces derived by SATL based on the principal components of the single-cell RNA sequencing data obtained from different species, we performed a functional analysis using the GO biological processes, see Materials and methods for details. We exemplary examined the common latent spaces of pig and rabbit as shown in Figure 4. Figure 4a shows the aligned latent spaces of the prediction scenario where the Pig1 dataset was used as source species and Rabbit 1, 2 and 3 are each used as target species (‘Pig1 → Rabbit1,2,3’) while ‘lps4’ cells were masked and predicted for each target dataset. One can see that masked cells from the rabbit data are correctly predicted based on the pig data on the common latent space. To validate whether SATL can find biologically plausible common latent spaces from all genes, we further investigated the variable importance.

Figure 4.

Visualization and interpretation of the common latent space. Shown is an example between Pig as a source and Rabbit 1–3 as a target in the lps-stimulated cell dataset. (A) Visualization of the Source True Label (Pig), Masked Cells (Rabbit), Target Prediction (Rabbit), and Target True Label (Rabbit) in corresponding latent spaces: pig1 → rabbit1, pig1 → rabbit2 and pig1 → rabbit3. SATL-aligned latent spaces are visualized via t-SNE plot. It shows the location of cells in the common latent space. (B, C) Overlap of 100 most important genes comprising latent space in Pig1 and Rabbit 1-3 depicted as Venn diagrams. (B) Pig1’s most important latent space genes aligned with different rabbit batches. (C) Rabbit1 3’s most important latent space genes aligned with Pig1. (D) GO (biological process) analysis from gProfiler. The union of the 100 most important contributing genes on latent axis 1 from Rabbit1-3 is analyzed. Top enriched GO biological processes are shown.

Therefore, the explained variance from PCA and the weight of the projection matrix from SATL are multiplied to find the most influential genes on each axis of the latent space.

The top 100 genes from the first axis of Pig1 and the Rabbit1, 2 and 3 data were examined. Each rabbit sample resulted in a different common latent space when aligned with ‘Pig1’ data (Figure 4b,c). By comparing the gene sets, we could observe a large within-species inter-individual variability in the latent space representation. In particular, when investigating the top 100 important genes of the first axis of the common latent space, one can discover distinct sets of genes across the different biological replicates (Figure 4B). In the rabbit dataset, all three individuals share 22 genes out of the top 100 genes. In the paired Pig1 result, they share 62 genes out of the top 100 genes. This can be seen in Figure 4C depicting the Venn diagrams of important genes composing the first axis of latent spaces in pig and rabbit, respectively. This demonstrates the potential of SATL to point out individual variances and the quality of normalization of the selected feature space.

Furthermore, we examined the functional relevance of the selected feature space, we performed a functional annotation of the top 100 genes in all three rabbits by GO enrichment of biological processes using g:Profiler [28]. As expected, it reveals a general tendency toward immune-related GO terms despite inter-individual variances in gene sets (Figure 4D). Moreover, the GO enrichment analysis shows that all three rabbits share the following biological processes: ‘Immune system process’, ‘Regulation of immune system process’ and ‘Defense response’, for instance. All these terms are associated with the immune system. This finding indicates that, although identified important genes were different, they have similarities in their biological function; for instance, the presented rabbit samples share similar immune system–related GO terms.

SATL captures biologically meaningful latent space from mouse to human

To further validate our approach we utilized another dataset consisting of single-cell sequencing data from paired mouse and human organs, bone marrow, pancreas and brain. Each dataset comprises different cell types and numbers thereof, see Dataset and preprocessing. We excluded the randomly selected two labels (here cell types) during the training process and predicted them in the target domain. The overall steps of this SATL analysis were as follows: first, we reduced the dimensionality and extracted features in an unsupervised manner with scETM [7]. Subsequently, we applied SATL to identify a common latent space between the mouse and human datasets, see Materials and methods for details.

When investigating individual cell type–wise accuracy, we observed a drop in accuracy for specific cell types. Among the three datasets, the SATL model trained on the brain dataset performs best when cell labels are transferred from mouse to human (Table 3). This is likely due to the fact that the brain dataset shows a balanced class distribution compared to the other two datasets. For the prediction of L2/3 IT cells, the scETM-SATL method shows slightly better performance than scAdapt, which uses external gene-pair knowledge. On closer inspection, the misclassified L2/3 IT cells were assigned to L5 IT cells (Figure 5, Supplementary Table S1). Literature search suggests that both cell types are closely related to intratelencephalic neurons and therefore share similar gene expression.

Table 3.

Comparison of the performance. (1) SATL, combined with different feature extraction methods (Seurat, DCA and scETM) and (2) a deep neural network approach for generalized zero-shot learning, CADA-VAE, with different feature extraction methods. The evaluation is based on the balanced accuracy of predicted seen, and unseen classes and h-score. The last rows show the accuracy of two gene homology-based deep neural network approaches, scAdapt and scNym, applied to a homogeneous transfer learning task on the same datasets

	Bone marrow				Pancreas				Brain
Method	S	U	H	Acc	S	U	H	Acc	S	U	H	Acc
Seurat:SATL	0.908	0.427	0.506	–	0.984	0.531	0.576	–	0.966	0.646	0.699	–
DCA:SATL	0.914	0.417	0.511	–	0.869	0.403	0.479	–	0.966	0.617	0.690	–
scETM:SATL	0.916	0.726	0.783	–	0.986	0.615	0.673	–	0.965	0.822	0.861	–
Seurat :CADA-VAE	0.946	0.090	0.153	–	0.980	0.229	0.322	–	0.984	0.438	0.576	–
DCA :CADA-VAE	0.947	0.180	0.275	–	0.899	0.302	0.407	–	0.982	0.390	0.514	–
scETM :CADA-VAE	0.934	0.138	0.201	–	0.980	0.349	0.475	–	0.977	0.524	0.645	–
scAdapt*	–	–	–	0.734	–	–	–	0.924	–	–	–	0.952
scNym*	–	–	–	0.780	–	–	–	0.538	–	–	–	0.623

S: balanced accuracy of seen classes, U: balanced accuracy of unseen classes, H: h-score, Acc: transfer learning accuracy of all classes. ^*Methods utilize gene orthology.

Figure 5.

SATL latent space result representation using t-SNE from mouse to human cell-type label transfer analysis. scETM feature is used here. Masked cells are represented in the second row. Among masked cells, misclassified cells are highlighted with dashed circles in the third and fourth rows. In brain data, ‘Lamp5’ and ‘L2/3 IT’ cells are masked. ‘L2/3 IT’ cell is predicted to ‘L5 IT’ cell. In bone marrow data, ‘Monoblasts’ and ‘Pro-B’ cells are masked. ‘Pro-B’ is predicted to ‘HSPCs’. In pancreas data, ‘quiescent stellate’ and ‘activated stellate’ are confused. The average expression profiles of these cell types, it becomes clear that macrophages, T and Schwann cells are less correlated between different species [18], further increasing the difficulty of zero-shot classification with species agnostic transfer learning. These examples demonstrate that the generalized zero-shot classification errors of our approach are not simply random, but rather are based on biological reasons.

In the bone marrow dataset, the accuracies of Pro-B and Pre-B cells were outliers among all other cell types (Supplementary Table S2). The two cell types, Pre-B cells and Pro-B cells, are hard to distinguish in the latent space, resulting in Pre-B cells being confused with Pro-B cells and Pro-B cells being classified as HSPC cells. When monoblasts and Pro-B are masked, Pro-B is predicted to HSPC (Figure 5). While this is a clear limitation of our model, it can be explained by looking at the cell differentiation lineage tree, where these cell types are closely related cell states. Specifically, Pro-B cells are on a developmental level located between Pre-B and HSPC cells. However, this differentiation was also challenging for the method using external knowledge, scAdapt (Supplementary Table S2).

In the pancreas dataset, several cell types are rare in the mouse dataset, namely, t cells, Schwann, activated stellate cells and macrophages. Consequently, all methods, SATL and methods with external knowledge show difficulties identifying these cell types (Supplementary Table S3). While the performance is relatively stable with the other preprocessing techniques, in combination with scETM-features, the model fails to recognize two out of these four cells. When activated stellate is masked, it is predicted to quiescent stellate (Figure 5). A second reason for this bad performance is that there are different degrees of biological similarity. When studying the correlations between.

To assess further robustness and generalizability of SATL, we evaluated the performance of the SATL model using two individual external test datasets. These datasets consist of single-cell RNA sequencing data of human pancreas (GSE85241 and E-MTAB-5061). We specifically selected common cell types in these datasets, and each dataset encompasses four and five cell types, respectively. The SATL model was trained using the previously mentioned mouse–human paired training datasets of the pancreas. During this training phase, the same cell types were masked in both human pancreas datasets. Following training, the individual external test datasets were transformed using the same PCA transformation and predicted with the trained SATL model. The average h-score for GSE85241 and for E-MTAB-5061 were 0.728 and 0.744, respectively (Figure 6). The results, where ‘alpha’ and ‘ductal’ cells are masked, are visualized. Unseen and seen cells from individual datasets are positioned in similar coordinates in the space determined by t-SNE. This result demonstrates that SATL can identify a generalizable latent space between species, making it possible to analyze different datasets by projecting them into this latent space.

Figure 6.

Individual test set results. (A) Illustration of individual dataset testing. The model is trained with a mouse–human pancreas pair dataset (GSE84133). Individual human pancreas dataset is transformed and predicted with a trained model. (B) A visualized t-SNE plot of GSE84133 to GSE85241 test is shown. Circles highlight unseen classes. (C) A visualized t-SNE plot of GSE84133 to E-MTAB-5061 test is shown. Circles highlight unseen classes.

SATL comparison with transfer learning method

We further evaluated the performance of SATL in combination with different feature extraction methods. Similar to the previous analysis, we used the paired data of mouse and human tissues: bone marrow, pancreas and brain. We excluded two randomly selected labels (here cell types) during the training process and predicted them in the target domain. The overall steps of this SATL evaluation were as follows: First, we reduced the dimensionality and extracted features in an unsupervised manner. Subsequently, we applied SATL to identify a common latent space between the mouse and human datasets, see Materials and methods for details.

Due to SATL being based on a feature-mapping HDA method, the feature extraction step affects the model performance [36]. Therefore, we combined SATL with a common preprocessing strategy and deep learning models for single-cell sequencing data, namely, Seurat [4], DCA [6] and scETM [7]. Seurat with PCA is a common single-cell data analysis method, including a preprocessing step for filtering highly variable genes. The top 2000 genes are selected for analysis. DCA is an AE-based model. It additionally includes a step for filtering variable genes after their denoising process. Lastly, we extracted features using scETM without a variable gene-filtering step. scETM is also a deep learning model. Subsequently, we applied SATL to identify a common latent space between the mouse and human datasets, see Materials and methods for details.

Table 3 presents the balanced accuracy of seen and unseen classes and h-score across all three datasets with three different methods for feature extraction. Despite the different feature spaces in paired species, SATL reasonably predicts the unseen cell types in the target domain. Moreover, the results show a strong influence of the feature extraction method on the SATL analysis, indicated by the large differences in performance. The balanced accuracies of unseen cell types are 0.417 0.726 in the bone marrow, 0.403 0.615 in the pancreas and 0.617 0.822 in the brain data. The performance varies depending on the dataset and feature extraction method. However, SATL with scETM achieves the highest h-score on all three datasets with 0.822. This result confirms the finding that the feature-mapping HDA methods strongly rely on the extracted features’ quality [25, 36]. In particular, if the feature extractor fails to embed an appropriate gene signature, the performance of SATL to classify similar cell types will be limited.

We performed a comparison of SATL with related machine learning methods, with and without prior knowledge, on the same data. Table 3 presents a performance comparison of SATL and CADA-VAE and two gene homology–based approaches. First, we compared our SATL results to CADA-VAE [12], a well-known method frequently employed in generalized zero-shot learning for image classification. The CADA-VAE model shows good performance for the prediction of seen classes. However, the model was not able to predict unseen classes well. The CADA-VAE shows the best h-score of 0.645 in combination with the scETM feature selection on the brain dataset. However, it was not better than the SATL model. Additionally, we compared our results with recent deep neural network–based methods that use prior knowledge of gene homology to map data into the same feature space. Both scNym [37] and scAdapt [10] work completely supervised but utilize prior knowledge of other species with orthologous genes. Thus, the scAdapt and scNym algorithms handle a typical transfer learning classification task, rather than a GZSL. Therefore, comparing the accuracy directly with the averaged h-score of SATL is unsuitable. However, their accuracies were comparable to or better than the averaged h-score of the SATL model. See Table 3 for details. This is likely due to the high quality of the gene homology database on a mouse–human pair. Both human and mouse are among the most intensively studied organisms. Thus, in this case, the information is sufficient for the cell type classification task.

Nevertheless, this result demonstrates the possibility of directly integrating different species datasets without external knowledge. Further comparison with other GZSL, HDA and DA methods are available in Supplementary Tables S4–S12.

DISCUSSION

In this paper, we present SATL, a novel methodology for species-agnostic transfer learning with HDA in biomedical research. To the best of our knowledge, SATL is the first algorithm that allows cross-species data integration in a phenotype-centric way without relying on any prior knowledge. The CDSPP, baseline model of SATL, solves an HDA using a latent embedding approach with a supervised eigenvalue problem by utilizing seen-label information. This CDSPP-based method can have a significant advantage in computational efficiency compared to recent deep neural network models that require training for numerous parameters. Furthermore, it is important to point out that SATL does not require large-scale training data, it is particularly applicable in various biomedical scenarios where high-throughput screening is not available.

In contrast to other methods, the key idea of SATL is to utilize a latent feature space for HDA instead of a homogenized feature space through gene homology. This characteristic enables a versatile integration of various species datasets using different feature extraction methods on gene-expression profiles while minimizing information loss, which is a consequence of gene-id conversion. The feature, latent-feature-based integration, gives additional benefit to our model. This SATL can integrate any deep features extracted from various high-throughput data, such as transcriptome, genome, epigenome, proteome or metabolome. This means SATL shows its possibility on cross-species data integration; however, it is possible to integrate multi-modal data for a model organism study. A drawback of SATL is its limitation toward applications in the transductive generalized zero-shot learning setup, which requires both labeled and unlabeled test samples to improve the accuracy of recognizing unseen classes. Additionally, we have constrained our analysis to situations where both domains have the exact same set of classes, meaning that the two domains have the same set of cell labels. This constraint can limit the ability to identify novel cell types in cross-species studies. In the next step, we could extend the analysis with a dataset having highly imbalanced classes.

Although SATL shows prominent performance in the classification of cell types in cross-species transfer learning, a conventional approach with prior knowledge had higher performance. However, it needs to be clarified that the results are obtained from the human–mouse pair, which are two of the most intensively studied species. Thus, that leads to a higher data quality and density compared to other species pairs. We think that SATL can be beneficial in research with less well-studied organisms on the evolutionary tree of life. Furthermore, our approach can be beneficial for the systematic analysis of novel analogous gene pairs. Current bioinformatics algorithms can predict orthologous genes based on sequence homology but cannot predict analogous genes that have similar functions but dissimilar sequences. One of the alternative approaches is structure-based protein function prediction, such as AlphaFold [38] or a protein language model [39]. However, functional prediction from a structure is still challenging. To the best of our knowledge, our SATL is the first method that enables the investigation of analogous genes in a data-driven way.

Additionally, we think that a combination of the proposed method with prior knowledge can lead to even more powerful approaches. For instance, The SATURN employed a ‘Macrogenes’ concept to successfully solve many-to-many mapping in gene orthology [40]. The SAMap algorithm integrates transcriptomes from different species by exploiting gene orthology information with a graph-based method [41]. It, therefore, improved the gene-id conversion method to have less information loss. Those approaches utilizing prior knowledge can guarantee a more biologically meaningful feature extraction than PCA or Autoencoder where latent features could represent noise rather than true biological signals. Thus, we envision that the SATL model integrated with such prior knowledge-boosted feature extraction methods could lead to improved performance and a better understanding of genes’ function across various species.

The proposed training and testing ”chem’ for cross-species transfer learning in GZSL application is easily expanded and applied to various biomedical research problems. SATL learns a projection matrix for common latent space to handle the GZSL task where partial common data with labels are available. Compared to the unsupervised/semi-supervised task, this GZSL task has great benefit in expanding knowledge based on accumulated datasets in biomedical domains. Our first analysis on the lps-stimulated dataset with the four species dataset demonstrates that SATL learns a biologically meaningful latent space and can predict unknown cell types through this latent space. Another interesting aspect is intra-species individual variability, while the biological replicates, rabbit1, rabbit2 and rabbit3, share a similar GO term related to ‘innate immune response’, which is the target biological process the original study designed.

As described, the SATL approach has an advantage in integrating all features in the data. When evaluating the 10 most important genes in the first axis of the latent SATL space of Pig one, two genes, namely, ENSSSCG00000028525 and ENSSSCG00000000246, have no match in databases (Not found in PantherDB 17.0 [29]; Retired in Ensembl release 110—July 2023). However, further analysis with Foldseek [42] detected two structurally very similar proteins (Figure 7). Moreover, these two proteins, ENSSSCG00000028525-CXCL2 and ENSSSCG00000000246-SAA2, are reported to be immune system related [43, 44]. Moreover, when the 10 most important genes of the latent space of Rabbit one are queried in the orthologue analysis database, Better Bunny (Aug. 2020) [45] and Ensembl (release 110—July 2023), ENSOCUG00000022364 is reported as a novel, yet unknown rabbit gene orthologues to interferon-induced transmembrane protein. Foldseek search found the same protein as a structurally similar protein. Those genes are dropped in a gene-id conversion preprocessing. However, in SATL analysis, they are found as important genes to investigate lps-stimulated response on macrophages. This result can illustrate a limitation of the current gene-id conversion method and a potential new approach to discovering and investigating the functional similarity of genes between different species.

Figure 7.

Phenotype-centric orthology inference via SATL. A proposed approach for ortholog analysis with SATL is shown. SATL exploits labels of data and finds latent space related to experimental design. When important genes composing latent space are investigated, some genes are not found in the current orthologous database. Two genes are retired in the Ensembl database. Therefore, structural similarity analysis of protein is done. Foldseek search identified ‘ENSSSCG00000028525’ as a similar protein to CXCL2 with sequence identity: 75.7 and E-Value: 3.87e−14. ‘ENSSSCG00000000246’ is identified as a similar protein to SAA2 with sequence identity: 72.3 and E-Value: 2.47e-13. ‘ENSSSCG00000017721’ and ‘ENSSSCG00000000242’ are identified as ‘CCL8’ and ‘CCL2’.

In summary, we present a new data-driven method, SATL, to enable the integration of heterogeneous high-throughput data across various species. In particular, SATL shows great potential for HDA algorithms to expand our species-dependent knowledge to more general pan-species biological knowledge. SATL can dissect functional groups of genes in the species-common latent space when presented with large-scale sequencing data, such as the responses of organisms under various experimental conditions.

Key Points

• SATL can align heterogeneous latent space of different species by utilizing common labels in data.

• SATL enables out-of-sample cell type prediction via cross-species transfer learning.

• SATL can overcome information losses originating in gene ID conversion with orthologous gene information.

• SATL can be utilized in the analysis of inter-species functional similarity genes.

Youngjun Park is a PhD candidate at the International Max Plank Research School for Genome Science in Georg-August-Universität Göttingen Göttingen, Germany.

Nils P. Muttray is a Master student at Georg-August-Universität Göttingen Göttingen, Germany.

Anne-Christin Hauschild is a Junior Professor at the Medical Informatics Department at the University Medical Center Göttingen, Germany.

AUTHOR CONTRIBUTIONS

The conceptualization of the study was conceived by Y.P. and A.-C.H. Computational experiments were performed by Y.P. and N.P.M. Y.P. and N.P.M. generated results. Y.P. prepared figures. Y.P. and A.-C.H. wrote the manuscript and performed review and editing.

FUNDING

Y.P. and A.C. are supported by the German Ministry of Education and Research (BMBF) under grant agreement No. 01KD2208A (project FAIrPaCT). Y.P. is supported by the International Max Planck Research School for Genome Science, part of the Göttingen Graduate Center for Neurosciences, Biophysics, und Molecular Biosciences.

DATA AVAILABILITY

The main analysis code is available in a GitHub repository, https://github.com/iron-lion/HDA4SATL. The code for a Python implementation of CDSPP for GZSL is available in a GitHub repository, https://github.com/iron-lion/cdspp-hda-py.