Literature mining in PubMed¹ is used to collect uveitis-related genes. In previous studies, Lu et al. (2017, 2018) conducted a PubMed search by using the keywords “uveitis” and obtained 744 relevant articles. A total of 98 review publications with basic summary of uveitis-related genes are manually reviewed. A total of 121 genes are identified from 96 out of 98 review papers reporting functional genes that may be significant for uveitis pathogenesis. Among these genes, 108 are annotated by GO terms and KEGG pathways, and those without annotation information are removed. These 108 genes obtained are used as positive samples. A total of 17,560 other human genes containing GO and KEGG annotation information are collected from Ensembl database and used as negative samples.

Uveitis- and non-uveitis-related genes should be transformed into some features so that they can be processed by generally machine learning algorithms (Zhou et al., 2020; Li et al., 2022; Pan et al., 2022; Tang and Chen, 2022; Wang and Chen, 2022; Wu and Chen, 2022; Yang and Chen, 2022). Here, we use GO and KEGG enrichment theory to generate numerical values for representing each gene (Carmona-Saez et al., 2007; Huang et al., 2011; Huang et al., 2012).

Gene ontology enrichment indicates the association between GO terms and genes. For one gene g, its direct neighbors in STRING and g are termed as neighbors_g genes in this paper. A score is calculated for neighbors_g genes and each GO term GO_j, which is commonly referred to as the GO enrichment score. The score is computed as the −log₁₀ of the hypergeometric P-value for neighbors_g genes and the gene set, in which genes are annotated by the GO_j (called GO_j gene set). The equation is shown as follows:

where k is the number of neighbors_g genes that appear in the GO_j gene set, m is the number of genes in the GO_j gene set, n is the number of neighbors_g genes, and N is the total number of genes considered (gene universe). A high score for neighbors_g genes and GO_j indicates the close relationship between g and GO_j. Finally, 20,681 GO term enrichment features are obtained.

Similarly, for neighbors_g genes and each KEGG pathway KEGG_j, the enrichment score is calculated to evaluate the relationship between g and KEGG_j by using the same method described above. The equation is shown as follows:

where k is the number of neighbors_g genes that are also in the KEGG_j gene set, m is the number of genes in the KEGG_j gene set, n is the number of neighbors_g genes, and N is the total number of genes considered (gene universe). A high score indicates the close relationship between g and KEGG_j. Finally, 297 KEGG pathway enrichment features are obtained.

In summary, a gene g can be presented as a vector v(g) consisting of 20,681 GO term features and 297 KEGG pathway features. v(g) is presented by

Combining the 17,668 samples (108 positive samples and 17,560 negative samples), a two-dimensional matrix is obtained for subsequent analysis.

Lots of GO terms and KEGG pathways are involved in this analysis. Evidently, only a few of them are highly related to uveitis, which can be extracted by advanced machine learning algorithms. Here, the Boruta method is first applied to exclude irrelevant GO terms and KEGG pathways and keep essential ones.

Boruta is a widely used feature filtering method that relies on two concepts, i.e., shadow features and binomial distribution (Carmona-Saez et al., 2007; Kursa and Rudnicki, 2010; Huang et al., 2022; Zhou et al., 2022). Features in Boruta compete with their random versions, which are created by random shuffling. In terms of feature importance, a feature is called “important” if its importance is superior to its random counterpart. In each run, Boruta keeps “important” features. The second strategy, binomial distribution, focuses on obtaining features with the probability of being kept in all runs higher than a certain confidence level. A set of features that are highly correlated with the target variable is obtained by keeping “important” features, which are statistically better than the best random features.

In this study, the Boruta package, retrieved from https://github.com/scikit-learn-contrib/boruta_py, is used to analyze the features mentioned in section “Feature construction.” Such package is performed with its default parameters.

Boruta is a feature selection algorithm that is designed to select features highly related to the classification. However, among them, some may be more important than others. Such task cannot be achieved by Boruta. Some other machine learning algorithms follow to further evaluate the importance of each selected feature. In this study, four feature selection methods [LASSO (Tibshirani, 1996), LightGBM (Ke et al., 2017), MCFS (Micha et al., 2008), and mRMR (Peng et al., 2005)] are employed. These methods can rank the features according to their importance to the classification.

The Boruta algorithm is performed on original features to eliminate nonrelevant features. Finally, 118 features related to the classification are chosen from 20,978 features, which are listed in Supplementary Table 1. These 118 features are functionally linked to the development of uveitis. Among these 118 features, 110 are related to GO terms, where the rest eight features are about KEGG pathways. Considering the features on GO terms are much more than those on KEGG pathways, such results are reasonable. For these 118 features, uveitis is highly correlated with some of them and weakly correlated with others, implying that the importance of the features should be investigated further.

Four methods, namely, LASSO, LightGBM, MCFS, and mRMR, are used to measure the importance of 118 features selected by Boruta. A total of 118 features are ranked in four feature lists (LASSO, LightGBM, MCFS, and mRMR feature lists) based on their ability in distinguishing uveitis genes. These lists can be seen in Supplementary Table 1.

As mentioned in section “Feature ranking algorithms,” each method has its limitations. Some essential features related to uveitis can have high ranks in one feature list, whereas they may be underestimated in another feature list, that is, their ranks in the list are not high. Thus, an investigation on all feature lists is beneficial to fully uncover all essential biological processes of uveitis. For each feature list, top 40 features are picked up to comprise a feature set. Accordingly, four feature subsets are obtained. For each subset, the distribution on GO terms and KEGG pathways is illustrated in Figure 2, from which we can see that features on GO terms are much more than those on KEGG pathways in each feature set. Furthermore, a Venn diagram is plotted to display the intersection of these four sets, as shown in Figure 3. The detailed intersection results on these sets can be found in Supplementary Table 2. Two features appear in all feature sets, that is, they are highly ranked by all four methods. These features are considered to be most essential for distinguishing uveitis-related genes from other human genes, which are discussed in section “Functional features recognized by all methods.” As for the other features identified by three or less methods, they may also be important. Some of them are analyzed in sections “Functional features recognized by three methods,” “Functional features recognized by two methods,” and “Functional features recognized by only one method.”

The first functional characteristic predicted by all methods is GO:0001841 describing neural tube formation. In 1983, researchers observed autoimmune-dependent attack on neural tubes during the pathogenesis of uveitis (Rahi and Addison, 1983), implying that neural tube formation is functionally associated with uveitis-related abnormal immune response. In recent years, an association analyses revealed that folic acid deficiency may contribute to the progression of uveitis by inducing neural tube defects (Sijilmassi, 2019), revealing the potential relationships between neural tube formation and uveitis. As for the next predicted shared feature, hsa04612, a KEGG pathway term, describes another immune-associated pathway, i.e., antigen processing and presentation in human beings. Antigen processing and presentation are confirmed to be related to uveitis by multiple publications from different perspectives (De Smet et al., 2001; Wakefield et al., 2016; Sharma and Jackson, 2017). Specifically, a specific subtype of HLA, i.e., HLA-B27, has been confirmed to participate in the pathogenesis of uveitis and associated with the susceptibility and prognosis of uveitis (Linssen and Meenken, 1995; Tay-Kearney et al., 1996; Wakefield et al., 2011). Therefore, the functional characteristics recognized by all methods have been validated to be functionally associated with uveitis, implying the efficacy and accuracy of our analysis.

For functional characteristics predicted by three methods, GO:0043379 (memory T cell differentiation) has been predicted to be associated with uveitis. According to recent publications, memory T cell has been recognized during uveitis early in 2002 (Imai and Ohno, 2002), and a report in 2021 validates that memory helper T cells trigger the autoreactive immune response during uveitis (Egwuagu et al., 2021). Therefore, such predicted result has been validated by previous publications, implying the accuracy of our analysis. Another recognized GO term GO:0045625 (regulation of T-helper 1 cell differentiation) is also functionally associated with helper T cell-mediated immune responses, which can also be supported by two publications mentioned above. The next predicted term, i.e., GO:0071639, describes the positive regulation of monocyte chemotactic protein-1 production. Positive correlations between monocyte chemotactic protein-1 and pigment epithelium-derived factor, a key factor for uveitis (Zipplies et al., 2009), have been observed (Yoshida et al., 2007). Therefore, as an immune related regulator, monocyte chemotactic protein-1 is a potential biomarker for uveitis.

The first GO term predicted by two methods is GO:0046642, which describes the negative regulation of alpha–beta T cell proliferation. Few publications compared the roles of alpha–beta and gamma–delta T cells during uveitis. Only one mouse model-based uveitis study confirmed that at least in a mouse model, alpha–beta T cells are directly associated with autoimmune-triggered uveitis, validating the efficacy and accuracy of our analysis. Another identified GO term by two methods is GO:0045078 (generally described with GO:0032729), which describes the positive regulation of interferon-gamma production. Interferon-gamma has been shown to be associated with tuberculous-initiated uveitis (Ang et al., 2014). Staphylococcus aureus infection described by hsa05150 has also been predicted to be associated with uveitis. S. aureus as a Gram-positive bacteria has been reported to be identified in various patients with uveitis and has been shown to initiate uveitis-associated autoimmune responses (Rosenbaum et al., 1980; Lin, 2015). Therefore, such bacterial infection is associated with uveitis.

Some characteristics that have only been predicted by one method remain. GO:0042102 describes the positive regulation of T cell proliferation. As discussed above, T cell-associated biological processes are shown to be related to uveitis. Therefore, T cell proliferation is speculated to be associated with uveitis without direct supports. The next predicted GO term is GO:0005151, which describes interleukin-1 type II receptor binding. Interleukin-1 alpha, interleukin-1 beta, and tumor necrosis factor are shown to be associated with endotoxin-induced uveitis (Yoshida et al., 1994; Tsai et al., 2009). Chemokine activity (GO:0008009) has also been identified by one machine learning method. Different from T cell-mediated immune responses, which can be found in almost every uveitis case, chemokine abnormality has been commonly observed in acute active uveitis (Verma et al., 1997; Van Kooij et al., 2006), indicating that such biological function may be associated with the rapid progression of uveitis.

Overall, as discussed above, some top features identified by different machine learning methods have been validated to participate in the initiation and progression of uveitis, implying that machine learning methods are effective tools in recognizing disease-associated biological functions. Among these methods, LightGBM has more features that are also identified by other methods (Figure 3), implying that LightGBM may be the most proper method for investigating uveitis. Therefore, our study recognizes a series of functional characteristics (GO: biological process, cellular components, and molecular function and KEGG: pathways) associated with uveitis and provides a new approach for disease mechanism exploration.