Single-cell profiling datasets are downloaded from the transcriptomic atlas of human and mice (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE165371). The total numbers of samples and features in the dataset are 321,387 and 74,593, respectively, which are composed of four different cell sample datasets, corresponding to four sell types: Golgi cell, GC, interneuron cell, and UBC. The breakdown of each dataset, including number of mouse cells, number of human cells, total number of cells and number of gene features, is provided in Table 1. A binary classification problem was employed to investigate each dataset, where mouse cells were termed as positive samples and human cells were considered as negative samples.

Although mRMR produces a feature list, users still cannot know which features should be selected for model building. In our study, such procedure is fulfilled by the integration of the IFS (Liu and Setiono, 1998) approach with RF (Breiman, 2001) and DT (Safavian and Landgrebe, 1991) algorithms.

The IFS (Liu and Setiono, 1998) is a feature selection approach that aims to select optimal features for the creation of a supervised classifier. To perform IFS on a descending list of features, we first construct a series of feature subsets, each of which contains some top features in the list. On each feature subset, a classifier is built and its performance is evaluated by ten-fold cross-validation (Kohavi, 1995; Ding et al., 2022; Tang and Chen, 2022). After testing all possible feature subsets, the classifier with highest performance is discovered. The features for such classifier is the optimal features, and the classifier is the optimal classifier.

As IFS method needs one classification algorithm, we employ two classic algorithms in this study. They are RF and DT. The RF is a meta-classifier containing a large number of DTs, where each DT is built based on randomly selected samples and each node in such DT computes the output through a random subset of features. The outputs of the DTs are aggregated to generate the final output class. RF is quite powerful. Thus, it has wide applications in tackling many biological and medical problems (Bifsha et al., 2014; Zhao et al., 2018; Jia et al., 2020; Liang et al., 2020; Liu et al., 2021; Pan et al., 2021; Li et al., 2022; Yang and Chen, 2022). RF reduces errors by averaging the predicted outputs of all DTs because of some variations between DTs. This phenomenon loses some interpretability, slightly increases bias, improves performance, and avoids overfitting.

Different from the RF algorithms, which acts as a kind of “black-box” classifier, DT can construct classification and regression models that are understandable by humans. Although it is not very powerful, it can provide novel insights to uncover underlying mechanism. This algorithm has also been used to deal with some important problems (Liang et al., 2020; Zhang et al., 2021a; Chen et al., 2021; Pan et al., 2021). The tree generated by DT generally consists of a set of interpretative rules, indicating the contributive roles of features to the final model in the format of “IF–THEN” conditions.

In our study, RF and DT are implemented using the Scikit-learn package in Python.

As listed in Table 1, four datasets are imbalanced. Models directly built on such datasets are always not efficient. Here, we adopt SMOTE (Chawla et al., 2002) to process this problem, which is a technique for oversampling based on the principle of creating synthetic data by using the k-nearest neighbor algorithm. First, a sample, denoted by x, is randomly selected from the minor class. Second, the k-nearest neighbors of x, which are also in the minor class, are found, and one neighbor is randomly selected, denoted as y. Third, a synthetic sample is created by x and y, which is defined as the linear combination of x and y with randomly generated combination coefficients. This sample is poured into the minor class. Such procedures are executed several times until minor class has equal number of samples to the major class. In this study, SMOTE is only used in the evaluation of classifiers in IFS method.

In our study, several machine algorithms are applied on four datasets of cerebellum cells. We can obtain some essential gene features. The interpretation of gene features includes two parts, i.e., interpretations of single- and combined-gene rules. The single-gene interpretation focuses on the optimal gene selected by the mRMR and IFS, whereas the interpretation of the combined-gene rule focuses on the predictive rules given by DT. Our interpretation is based on a comprehensive literature review of a previous work.

The Matthew’s correlation coefficient (MCC) (Matthews, 1975; Chen et al., 2017; Zhao et al., 2018; Jia et al., 2020; Liang et al., 2020; Zhang et al., 2021b) is used to evaluate the performance of training models. MCC is the correlation coefficient between the observed categories and predictions. MCC serves as an indicator that can be applied to samples with large imbalance. MCC is defined as: M C C = T P × T N − F P × F N ( T P + F P ) ( T P + F N ) ( N P + F P ) ( N P + F N ) (1) where TP, FP, TN, and FN represent the numbers of true-positive, false-positive, true-negative, and false-negative samples, respectively. The value range of MCC is distributed between −1 and 1. A high MCC indicates good performance of the classifier.

In additional, we also employ other measurements, including sensitivity (SN) (same as recall), specificity (SP), prediction accuracy (ACC), precision and F1-measure. They can be computed by S N = T P T P + F N (2) S P = T N T N + F P (3) p r e c i s i o n = T P T P + F P (4) F 1 − m e a s u r e = 2 × r e c a l l × p r e c i s i o n r e c a l l + p r e c i s i o n (5)

These measurements are provided for reference.

In our study, 321,387 cerebellum cell samples were collected. These single-cell samples could be divided into four categories, i.e., Golgi cell (5,048, 1.57%), GC (250,307, 77.88%), interneuron cell (60,526, 18.83%), and UBC (5,506, 1.72%). In the original cell sample data, each cell sample was represented by expressions on lots of genes, which were quite difficult for machine learning analysis due to the dimensionality curse. Therefore, the Boruta filtering method was first used to do the optimal compression of features (i.e., gene expressions) for each of 4 cell datasets. After compression, each Golgi cell, GC, interneuron cell, and UBC was represented as a compressed vector of 1,276, 1,271, 1924, and 1,252 features, respectively. The gene ID corresponding to the feature in the sample vector could be found in Supplementary Table S1.

Remaining features were analyzed by the mRMR method. An mRMR feature list was obtained for each dataset. Four feature lists are also provided in Supplementary Table S1. These lists were further investigated in the following procedures.

For each dataset, an mRMR feature list was obtained. Afterward, RF and DT classification algorithms were used to construct the classification models in the IFS method with a step size of one on the basis of such list. Each classification model was evaluated by ten-fold cross-validation. The predicted results were counted as measurements listed in Performance Evaluation section. To clearly show the performance of DT and RF on different feature subsets, an IFS curve was plotted for each classification algorithm, as shown in Figure 2, where x-axis represents the number of features in the subset and y-axis stands for the main measurement, MCC. The key information extracted from these IFS curves is listed in Table 2.

When RF was used in the IFS method, it achieved perfect performance with MCC = 1.00000 when proper feature subsets were adopted. In detail, for Golgi cell, 518 top features in the mRMR feature list were used, whereas this number was 2, 100 and 28 for other three types of cells, respectively. These features constituted the optimal features for each dataset. Accordingly, an optimal RF classifier was built on each dataset using corresponding optimal features. Their detailed performance, including SN, SP, ACC, precision and F1-measure, is illustrated in Figure 3. Evidently, each measurement reached its perfect value, suggesting extreme good performance of these classifiers. They can be efficient tools to classify mouse and human cerebellum cells.

Although the optimal RF classifiers yielded perfect performance, they cannot provide useful clues to uncover essential differences between mouse and human cerebellum cells because RF is a black-box algorithm. In view of this, DT was employed in this study, which can provide more clear insights to study mouse and human cerebellum cells. It can be observed from Figure 2 that the highest MCC on each dataset yielded by DT was also very high (>0.99000). These MCC values were obtained by using top 34, 5, 1 and 28 features in the list on four datasets, respectively. Accordingly, these top features comprised the optimal features for four datasets, respectively, and an optimal DT classifier with corresponding optimal features was built on each dataset. Other measurements of these optimal DT classifiers are shown in Figure 3. Clearly, on each dataset, the optimal RF classifier was always superior to or equal to the optimal DT classifier. This result conformed to the general fact that RF is more efficient than DT.

The optimal DT/RF classifiers constructed above shown good even perfect performance. This section elaborated that these results were not incidental. To this end, for each optimal classifier, we did the following test. According to the number of features in the optimal classifier, i.e., the number of optimal features, we randomly selected same number of features from all features. These selected features were used to represent mouse and human cerebellum cells. A classifier with DT or RF was built on such representation. Ten-fold cross-validation was employed to evaluate its performance. To give a full test, above procedures were conducted twenty times. Obtained MCC values were shown in a box plot, as illustrated Figure 4. It can be observed that these classifiers with randomly selected features all provide lower performance than the corresponding optimal classifier. This indicated that the optimal features were really important for classifying mouse and human cerebellum cells and can be significant single-genes to distinguish mouse and human cerebellum cells.

Important genes for distinguishing human and mouse cerebellar cells based on the mRMR ranking are presented in Table 3. These genes can be significant single-genes to distinguish mouse and human cerebellum cells, which would be discussed in Analysis of Significant Single-Genes section.

Furthermore, we adopted the rule learning algorithm DT to generate combined-gene rules, interpret the classification rules of features. According to the optimal DT classifier on each dataset, a tree was learnt on all cell samples, which were represented by the optimal features of this DT classifier. Such tree was represented by some rules, which are listed in Table 4. A total of 8, 2, 2, and 11 rules were obtained for the Golgi cell, GC, interneuron cell, and UBC, respectively. In Analysis of Combined-Gene Rules section, we would discuss these rules.

According to features used in the optimal RF classifiers, we identified 648 (518+2+100+28) considerable features with the mRMR method to distinguish human and mouse cerebellum cells. Next, we further introduced research evidence related to the most important features in each cerebellar cell type to confirm the reliability of previous forecasts.

The protein encoded by leucine-rich repeat and immunoglobulin (Ig) domain containing 2 (LINGO2) is identified as a member of the leucine-rich repeats (LRR) gene family. The expression of the LINGO2 gene in the hypothalamus and cerebral cortex hypothalamus is higher than that in other regions. The Lingo2 is in the top relevant feature with the mRMR method and has been linked with essential tremor (ET) and Parkinson’s syndrome (Wu et al., 2011). ET is the most common movement disorder and adult patient accounts for the vast majoritys. Studies showed that LRR and LINGO2 protein structural variations containing Ig domains may be related to ET. Compared with those of the control group, the protein levels of LINGO1 in the cerebellar cortex and cerebellar white matter of patients with ET are significantly increased. Changes in the LINGO2 expression in the diseased brain appear to occur as the disease progresses, starting in the cerebellar cortex before reaching the white matter. Compared with those of normal individuals, the LINGO1 protein levels in the cerebellar cortex and white matter of patients are significantly increased. In addition, the expression of LINGO2 of patients shows consistent changes with the progression of ET, which starts from cerebellar cortex and then reaches the white matter. The upregulation of LINGO expression is likely to be a potential pathological indicator of neurodegenerative diseases (Jasinska-Myga and Wider, 2012; Delay et al., 2014). Studies confirmed that a tSNP variant of LINGO2 is associated with Parkinson’s syndrome (p < 0.05) (Chen et al., 2015). Our results show that LINGO2 is significantly different in the cerebellum of humans and mice, indicating that the LINGO2 expression may be associated with the development of the nervous system. The above research results remind us that the expression level of LINGO2 may be related to the nervous system especially the evolutionary level of cerebellum. In addition, the LINGO2 mutant may be an important indicator of neurodegenerative diseases.

The protein-coding gene UBE3A encodes the E3 ubiquitin protein ligase, which can bind to the ubiquitin of the E2-binding enzyme in the form of a thioester and then transfer ubiquitin to E2-binding enzyme substrate (Kumar et al., 1999; Dhananjayan et al., 2006; Mishra et al., 2009; Shimoji et al., 2009; Gossan et al., 2014; Ronchi et al., 2014). In addition, UBE3A can accelerate the degradation of misfolded proteins, thereby achieving cell quality control. The UBE3A gene presents a biallelic expression pattern in some tissues, but its transmission mode in the brain is maternal inheritance. The Ube3A mutation can cause the Angelman syndrome (Buiting et al., 2016), a neurological disease accompanied by severe developmental delay, hypotonia, epilepsy, aphasia, and other complications. In addition, reports showed that the UBE3A protein binds to the E6 protein of papillomavirus, causing p53 ubiquitination and hydrolysis. Other studies showed that UBE3A can mediate the activity-regulated cytoskeleton-associated protein (ARC) ubiquitination and degradation to regulate synaptic growth (Greer et al., 2010). In addition, mutations in Ube3A are related to autism. Researchers found that Ube3A dysfunction can increase ARC expression and reduce the quantity of α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA) receptors in synapses. Therefore, researchers inferred that AMPA dysregulation may be related to Angelman syndrome or other cognitive disorders (Greer et al., 2010). In our analysis results, Ube3A, as one of the top features, has significant differences in expression in human and mouse cerebellum cells. Combined with existing studies, our study suggests that the expression level of Ube3A may serve as a powerful indicator for predicting the function and evolution of the nervous system.

Neuroligin 1 (NLGN1) is a protein-coding gene, and its translation product is a member of the neuron cell surface protein family. The NLGN1 family can be used as specific ligands for β-neuroproteins, which may be related to the formation and remodeling of synapses. The NLGN1 protein interacts with neuroproteins to promote synaptic transmission signals and recruits and accumulates other synaptic proteins. Studies showed that NLGN1 can promote the de novo formation of synaptic structures in vitro and may participate in the regulation of excitatory synapses. The protein encoded by NLGN1 has hydrolase and protein dimerization activities and plays an important role in protein–protein interaction at the synapse and signal transmission process across the synapse. NLGN1 variants may result in autism and Asperger’s syndrome. We speculated that the differential expression pattern of NLGN1 in Golgi cells between human and mouse cerebellum may imply a potential functional evolution.

The Raly-like recognition motif (Ralyl) is identified as an important characteristic gene by the mRMR, and its encoded RNA binding protein affects embryonic development (Ji et al., 2003a). A previous study pointed out that Ralyl may be related to Alzheimer’s disease (Zhang et al., 2020). Researchers revealed that Ralyl is a hub gene in the brain transcriptome module of patients with Alzheimer’s disease and is highly associated with Alzheimer’s reserve-related phenotypes. Notably, the Ralyl expression decreases with Alzheimer’s progression. Subjects with Alzheimer’s disease reserves show significantly higher Ralyl expression compared with those without Alzheimer’s disease reserves (Zhang et al., 2020). Ralyl is related to cancer cell metastasis and poor prognosis in patients with liver cancer. Ralyl is specific for liver progenitor cells and regulates hepatocellular carcinoma stem cells by upregulating the stability of TGF-β2 mRNA through the reduced N6-methyladenosine modification (Wang et al., 2021). In addition, the overexpression of Ralyl can inhibit the MAPK and CDH1 signaling pathways, thereby inhibiting the development of ovarian cancer (Xia et al., 2021). Compared with those in nontumor tissues and epithelial ovarian cancer cells, the expression level of Ralyl in ovarian clear cell cancer cells is lower. The pathological stage and prognosis of patients with ovarian clear cell carcinoma and high Ralyl expression are improved. Other diseases associated with Ralyl include Bardet–Biedl syndrome 1. Ralyl can be regarded as a prognostic marker for certain tumors and a monitoring target for central nervous system disorders.

The protein-coding gene Fgf14 belongs to the fibroblast growth factor (FGF) family. Members of this family can promote cell mitosis and are closely related to other biological processes. The mutation of this gene is related to autosomal dominant-inherited brain ataxia (Miura et al., 2019). Fgf14 related pathways include ERK signaling and apoptosis pathways in synovial fibroblasts. FGF14 is an intracellular protein that controls neuronal excitability and synaptic transmission and is suggested for use in the nervous system and mental diseases. Studies showed that male Fgf14 knockout mice have significantly reduced aggressiveness, sexual behavior, and other behaviors driven by spontaneous initiatives. The fine-tuning of neuronal function by Fgf14 is an important mechanism for controlling such behaviors. FGF14 can control the excitability and synaptic transmission of neurons and has certain diagnostic indicators in neurological and mental diseases. Further molecular studies revealed that Fgf14 can affect individual behaviors by regulating the function of neurons (Hoxha et al., 2019). Recent data indicate that Fgf14 can modulate multiple ion channels and the localization of the potassium voltage-gated channel subfamily Q member 2 (KCNQ2) protein in hippocampal neurons (Pablo and Pitt, 2017). All these results proved the important role of Fgf14 in regulating nervous functions. Our analysis proposed a new sight that Fgf14 shows differential expression in GCs between human and mouse cerebellum, and implying a linkage between gene Fgf14 and nervous system evolution in mammals.

NRXN3 belongs to the neuroprotein (NRXN) family, which can act as a cell adhesion molecule in the process of synaptogenesis and intercellular signaling. NRXN3 has a wide range of alternative splicing and alternative promoters. Thus, the gene has multiple transcription variants and protein isoforms. Previous research suggested that NRXN3 variants are associated with abnormal behavioral phenotypes, such as alcohol dependence, nicotine addiction, and autism spectrum disorders. However, new research showed that NRXN3 also plays a potential role in disorders of synaptic transmission. NRXN3-related pathways include muscular dystrophy and protein interactions at synapses. Studies showed that mice with missense variants of NRXN3 show increased fear. The possible change in NRXN3 from arginine to tryptophan is a pathogenic variant of empathy and fear (Keum et al., 2018). Researchers collected one data set related to healthy aging and 3 data sets related to Alzheimer’s disease in the hippocampus from the Gene Expression Omnibus database. The results of functional analysis showed that NRXN3-led synaptic dysfunction plays a prominent role in the process of aging and Alzheimer’s disease-related cognitive decline. In addition, when the expression of NRXN3 in an individual decreases, the risk of Alzheimer’s disease increases, but its underlying mechanism needs to be further elucidated (Zheng et al., 2018). NRXN3 encodes an important part of synaptic function related to autism and other neurodevelopmental/neuropsychiatric diseases (Südhof et al., 2008). The chromosome microarray analysis is used to identify rare exon deletions affecting the NRXN3 alpha isoform in three-generation Chinese families. The results of family cosegregation studies indicate that NRXN3 affects autism and neurodevelopment/neuropsychiatric disorders. Moreover, schizophrenia and facial deformities are potential new features of NRXN3 haploid deficiency (Yuan et al., 2018). A study showed that the NRXN3 gene is a potential factor affecting the risk of nicotine addiction and that the NRXN3 marker rs1004212 is significantly related to the amount of smoking (Novak et al., 2009). Combining our analysis results and existing research, we further speculate that the abnormal expression of NRXN3 is the cause of neurological diseases that cannot be ignored, and its expression level may also be an important marker for representing the evolution of the nervous system of different species.

The protein encoded by Ankyrin 2 (ANK2) belongs to the ankyrin family and connects integral membrane proteins with the cytoskeleton. Ankyrin is involved in cell proliferation and movement and the maintenance of special domains. ANK2 can promote the localization of ion transporters and channels and maintain the stability of cell membranes. For example, in cardiomyocytes, ANK2 can coordinate the assembly of ion exchangers to maintain and promote the targeting and stability of ion exchangers. In addition, in neonatal cardiomyocytes, ANK2 is indispensable for regulating the contraction rate. In the skeletal muscle, ANK2 is involved in the correct positioning of DMD and DCTN4 and in the formation and/or stabilization of microtubule subsets related to the ribs and neuromuscular junctions. In the rod-shaped photoreceptor, ANK2 participates in the coordinated expression of Na/K atpase, Na/Ca exchanger, and β-2-spectrin. In addition, ANK2 is involved in important life processes, such as endocytosis and intracellular protein transport. ANK2 variants can cause long QT syndrome four and arrhythmia syndrome (Watanabe and Minamino, 2016; Gessner et al., 2019). The association between gene ANK2 and interneuron cells of cerebellum have not been reported so far. Our study demonstrated the significant difference in ANK2 expression between human and mouse cerebellum, and it suggested a potential role of ANK2 in nervous system development and evolution.

The phosphodiesterase (PDE) 1A gene encodes a Ca²⁺/calmodulin-dependent PDE, which includes 23 exons and 9 subtypes. The PDE1A gene belongs to the cyclic nucleotide PDE family. PDE, a phosphohydrolase, catalyzes the hydrolysis of adenosine (cAMP) and/or guanine (cGMP) 3′,5′-cyclic phosphate in the 3′-cyclic phosphate bond. PDE1A can regulate the concentration of cyclic nucleotides in the cell and influence signal transduction. The cyclic nucleotide PDE has dual specificity for cAMP and cGMP and is involved in the regulation of some important physiological processes. PDE1A can bind calmodulin and cGMP, has higher affinity for cGMP than for cAMP, and occupies an important position in the GPCR and calcium signaling pathway. Current studies found that PDE1A9 is highly expressed in the brain tissue, but its expression may lead to functional changes depending on age. Studies showed that compared with young controls, the phosphorylation level of the transmembrane regulatory protein in the hippocampus of aged rats is significantly reduced (Kelly et al., 2014), which may be related to the expression of PDE1A in individuals of different ages. Other diseases associated with PDE1A include Fraser’s syndrome 1.

The encoded product of the G protein signal regulator 6 (RGS6) belongs to the G protein signal transduction regulator protein family. The RGS6 protein is characterized by DEP and GGL domains. The latter is the G beta 5 interaction domain, and these proteins can activate the gtpase activity. Many alternatively spliced transcripts of this gene have long or short N-terminal domains, complete or incomplete GGL domains, and isotypes of different C-terminal domains. The RGS protein may regulate G protein-mediated signal transduction through negative feedback, thereby affecting the activity of neurons, cardiovascular system, and lymphocytes and may even increase the risk of cancer. Mutations in RGS6 may cause Hirschsprung’s disease 1, night blindness and congenital quiescence. RGS6-related pathways include GPCR signal transmission and protein metabolism. When the G protein surface receptor is activated, the G protein initiates a signal cascade in the host cell. The RGS protein inactivates the G protein and turns off this signaling cascade. RGS6 belongs to the R7 subfamily and regulates the G protein function, which is essential for the transmission of a variety of neurotransmitters and neuronal responses. Genetic variations in RGS6 may disrupt normal GPCR signals, leading to disease or subtle features. For example, studies reported that RGS6 abnormalities may be related to diseases, such as alcohol dependence, Parkinson’s syndrome, and neurological or affective disorders (Ahlers et al., 2016). In human dopamine neurons, the expression of RGS6 is restricted but can regulate the D2R-Gi/o pathway and can prevent Parkinson’s neurodegeneration, resulting in the accumulation of α-neurite nuclein (Luo et al., 2019). RGS6 has an important effect on the differentiation of microtubules and neurons. RGS6 induces neuron differentiation through a new mechanism involving the interaction of SCG10 with its GGL domain (Liu et al., 2002). Existing studies are consistent with our analysis results. RGS6 is expressed in humans and mice and is regarded as an important characteristic gene in cerebellar cells. Compared with the existing research, our analysis locates the specific cell type of the RGS6 expression site from the single-cell level, laying a foundation for in-depth mechanism research.

We built total 23 combined-gene rules through the DT method. A total of 8, 2, 2 and 11 decision rules are observed for distinguishing human from mouse cerebellum in Golgi cell, GC, interneuron cell, and UBC, respectively. Given the biological significance of these traits, how do we determine the stage of evolution and development on the basis of the expression of these traits? Here are some studies to introduce their experimental evidence.

The protein coding gene LINGO2 is one of the four important families related to the nervous system (LINGO1–4) (Llorens et al., 2008; Homma et al., 2009). This gene is enriched in early spermatids, late spermatids, and bipolar cells. The LINGO2 expression is detected in the neuronal tissues of the brains of adult mice (Vilarino-Guell et al., 2010). The LINGO2 variant detected in the Chinese population may increase the risk of gestational diabetes. The results of a large number of Asian population studies showed that LINGO2 may be a susceptibility gene for ET and Parkinson’s syndrome and that the increased expression of LINGO is a characteristic pathological response of neurodegenerative diseases (Delay et al., 2014). When the expression of LINGO2 gene is high, a low degree of nervous system evolution is observed. By contrast, the expression of the LINGO gene is low when the species evolves at a high level especially when the nervous system is developed. On the basis of existing research and our analysis results, we speculate that LINGO2 may show specific expression patterns at different stages of neurodevelopment and species evolution, which may provide references for interpreting neurological disorders.

Thsd7b is another important gene involved in the decision rules by our analysis. The thrombospondin type 1 domain containing 7B (Thsd7b) is a protein-coding gene. Thsd7b is enriched in human brain regions especially the pons and medulla. Thsd7b is closely related to O-linked glycosylation, glycosylation diseases, and Ehlers–Danlos syndrome. Researchers found a correlation between Thsd7b and the formation of cisplatin resistance. A large number of studies on pancreatic cancer in Japan showed that the Thsd7b gene is significantly associated with the risk of pancreatic cancer. In addition, this gene is related to the prognosis of non-small cell lung cancer with chemotherapy intervention (Lee et al., 2013). At present, no study is available on the mechanism of Thsd7b related to the cerebellum. Our analysis results fill in the gaps in the development of the cerebellum especially the evolution of the Golgi cell.

A relatively higher expression of Ccdc85a in UBCs was required to indicate human cerebellum in decision rules. The protein-coding gene coiled-coil domain containing 85A (Ccdc85a) is expressed in various brain regions especially in the cerebral cortex. Diseases associated with Ccdc85a include hydrocephalus. Studies confirmed that the Ccdc85a protein is required in the AppNL-F interaction group. The results of colocalization analysis indicate that the Ccdc85a protein may endogenously regulate the function of the amyloid β-protein. In the early stage of Alzheimer’s disease, the upregulation of Ccdc85a expression may be a compensation for the increase in amyloid β-protein and the elimination of amyloid β-protein metabolism. These findings suggest that Ccdc85a may play an important role in nervous system development and evolution, and become a new target and biomarker for clinical intervention in neurological dysfunction diseases (Aladeokin et al., 2019).

We use the mRMR feature screening method to conduct an in-depth analysis of the existing single-cell transcriptome data set and select the key and characteristically expressed genes. Subsequently, we use DT and SMOTE tools to determine the expression rules of characteristic genes. In the end, we obtain key genes that may be related to evolution and neurodevelopment and confirm the decision-making rules which reflect the heterogeneity between species in different cerebellar cell types. Our analysis results are consistent with many existing research conclusions, but the specific pathogenic molecular mechanism of each characteristic gene needs further verification. Overall, this research has obtained representative species evolution genes and their expression differences in various cerebellar cells. The remarkable potential of these features and rules in studying species evolution and are highlighted and provide insights into new key genes. The excellent performance of our classifiers can be attributed to the strong specificity of gene expression at species and tissue levels, which can significantly distinguish human and mouse. In addition, our research methods and strategies have a good guiding role in exploring genetic evolution. Therefore, the characteristic genes we have identified can be used to identify specific cell groups and the evolutionary level of species and can be regarded as biological indicators to provide research directions for disease-related molecular mechanisms.