To collect the information of meiotic genes, specific keywords were used to search in PubMed (Supplementary Table 1). Then all the collected papers were curated manually and genes that had been validated by experiments were deemed as functional meiotic genes.

Gene expression information was retrieved from the ArrayExpress database.² Datasets from Affymetrix GeneChip platform were downloaded and were divided into different categories, including “developmental stages,” “gene disturbance,” “before and after treatment,” and “tissues and cell types” (Supplementary Table 2). Gene expression data combined with category information are provided as annotation information in MeiosisOnline and applied for predicting genes that may function in meiosis.

Annotation information for each gene in MeiosisOnline contains “basic information,” “function annotation and classification,” “protein–protein interaction (PPI) and gene expression.”

To execute more jobs in parallel, a Dell 730 server with LAMP architecture is used to host the MeiosisOnline database. The server is equipped with 128 GB RAM and two 12-core Intel processors (2.2 GHz). The jQuery is used to render the interface and Python and R languages are employed to supply the backend.

MeiosisOnline is aiming to construct a functional annotation pipeline about meiosis-associated genes from published articles. After keywords querying in PubMed, about 45,000 research articles published before January 1, 2021, were collected. All collected papers were manually curated, and functional meiotic genes are only included as those with functional experimental validation (Supplementary Table 3). In total, 2,052 unique meiotic genes with experimentally verified functions from 84 species were curated along with functional information in MeiosisOnline. We found that the functional meiotic genes are firstly derived from mice, which accounts for 28.74% of the total reported genes, followed by human (5.16%) and rat (5.07%). Furthermore, other species comprise the rest of 61.03% (Supplementary Table 4). To be noted that the genes always have preferably expression profiles, for example, Mlh3 (mg0000873) expresses during both male and female meiosis, Sun5 (mg0000693) only presents in male germ cells, while Bmp15 (mg0000982) is specially expressed in oocytes.

MeiosisOnline is developed in a user-friendly manner and the major functional modules of the database (Figure 1) include:

Besides the general information including gene ID, protein ID, taxonomy ID, and basic descriptions, MeiosisOnline also provides high-quality functional annotation for the collected genes. Based on the function annotation information, the experimental verified genes were classified into different categories. Additionally, figures and/or tables illustrating the function of the collected meiotic genes were also incorporated. Moreover, manually annotated functions, signaling pathways, and associated protein complexes of the collected meiotic genes were provided (Figure 2). The functional distribution of these genes in various stages of meiosis and fecundity is also listed (Supplementary Tables 4, 5).

For instance, using “Stra8” for query, results will be listed in tabular form, including MeiosisOnline Gene ID (MG ID), gene names, UniProt ID, etc. (Figure 2A). Once clicking the MG ID (MG0001089), detailed information for mouse Stra8 is available that includes the following: (1) basic information (gene name, nucleotide and protein sequences, etc.), (2) functional annotation and classification from related literature (developmental stages, experimental methods, literature abstract, relevant figures, etc.), and (3) PPI and gene expression information (Figure 2B).

To expand the utilization of our MeiosisOnline database, a prediction model was constructed and used to predict the candidate functional meiotic genes. As mouse is one of the best studied animal models, the GAS algorithm (Zhang et al., 2013) was used to predict potential meiotic functional genes from Mus musculus (Supplementary Figure 4).

To verify the efficiency of GAS, we randomly separated the training data into two equal parts: one as a new training dataset and the other as a testing dataset. The model was split into three stages: stage 1 was constructed with features of the category “developmental stages,” stage 2 included features from the categories “tissues and cell types” and “before and after treatment,” and stage 3 is the features of the category “gene disturbance.” As shown in Supplementary Figure 5, the performances of GAS models were better when more features were added to them. Then, based on the experimentally verified meiotic genes and gene expression data, 590 mouse genes with experimentally verified function in meiosis were used as the positive training dataset. The negative training dataset contained 5,868 genes from MGI (Mouse Genome Informatics),¹⁰ of which knockout mice did not have any abnormalities in the reproductive system. Three hundred and ninety-four features used for GAS construction and prediction were extracted from the 85 microarray data (Supplementary Table 2).

Ultimately, 165 candidate genes (GAS probability > 0.5) having potential role were sorted out and annotated in MeiosisOnline (see text footnote 8). For the candidate genes, information that implicate their function in meiosis, including gene expression, protein localization, structure, and protein interactions, are included in MeiosisOnline.

We also performed GO enrichment analysis on both literature-reported genes in M. musculus (Supplementary Table 6) and GAS-predicted candidate genes (Supplementary Table 7). Compared with whole genome data, we statistically calculated the distribution of MGs in cellular components, biological processes, and molecular functions by R (hypergeometric distribution, p < 0.05) (Yu et al., 2012). Among all GO terms, 118 in biological processes, 1 in molecular functions, and 32 in cellular components are overlapped in both sets of reported and predicted genes (Figure 3A). Considering that predicted genes enriched by overlapped GO term have more potential in regulating meiosis, interestingly, meiotic cell cycle (GO: 0051321) was enriched from both sets of genes (reported and predicted) (Supplementary Table 8).

Furthermore, as genes are mostly regulated through network structure in meiosis, we mapped out the PPIs among all of the genes and constructed a potential meiosis network with 1,083,566 reported PPIs among 26,569 proteins in MeiosisOnline. For example, Cct5 and Sf3a3, which have not been reported, show very high connectivity with reported genes (Figure 3B). Further investigation of Cct5 and Sf3a3 would disclose how these two genes interacted with reported genes and what is the function of these interactions. Moreover, we also found that reported genes like Ccna2 interacted wildly with predicted genes (Figure 3C).