Golgi is an essential organelle in eukaryotic cells (Yang et al., 2019). Its main function is to store, package, and classify proteins. Golgi proteins are mainly composed of Cis-Golgi proteins and trans-Golgi proteins (Su et al., 2022). The main task of Cis-Golgi is to accept and process at the same time. The main task of trans-Golgi is to release proteins labeled and processed by vesicles. Studies have shown that dysfunction of the Golgi apparatus in cells can cause diseases such as diabetes (Wang and Zou, 2023), Parkinson’s disease (Gonatas et al., 1998), Alzheimer’s disease (Gonatas et al., 1998), and some cancers. The current treatment methods can only partially cure the disease (Elsberry and Rise, 1998), which is challenging to meet the needs.

With the development of machine learning technology, machine learning model has been applied to the related research of protein analysis (Villeneuve et al., 2017; Wei et al., 2017a, 2019; Zeng et al., 2018; Hou et al., 2019; Yuan et al., 2019; Hummer et al., 2020). However, there are few studies on the classification of Golgi protein types, and only a few are used to study the resident proteins of Golgi. In the past few years, Van Dijk et al. (2008) proposed a method to predict the type of type II membrane protein. It utilized a linear kernel support vector machine as a classifier. Ding et al. (2011) Utilized PSEAAC and customized Markov discriminator to identify Golgi protein types with an accuracy of 74.7%. Then, the improved interval dipeptide combination method enhances the accuracy and realizes the prediction accuracy of 85.4% (Ding et al., 2013). Jiao and Du (2016a) utilized the position-specific physicochemical properties (PSPCP) of amino acid residues to extract features and improved the model’s prediction accuracy to 86.9%. After that, they combined PSPCP with Chou’s pseudo amino acid composition Jiao and Du (2016b). Lv et al. (2019) designed a random forest sub-Golgi protein classifier Rfgpt, which utilized 2-gap dipeptide and split amino acid composition as feature vectors, and combined with synthetic minority oversampling technique (SMOTE) and analysis of variance (ANOVA) feature selection method, and the prediction accuracy is 90.5%.

In order to improve the classification effect of Golgi resident proteins, we proposed Gogli_DF model to classify the Golgi proteins with the deep forests model. Firstly, we utilized the UniRep method to achieve 1900-dimensional vector features. Secondly, we utilized the synthetic minority oversampling Technology (SMOTE) to deal with the imbalance issue of the classified samples and then used the light gradient boosting machine(Light GBM) method to reduce the dimension of the feature vector to 200-dimensional. Nextly, one-dimensional convolution, multi-layer LSTM, and PHATE dimensionality reduction are employed to extract the feature information, respectively. At the same time, the same 32-dimensional dense layer is used in the penultimate layer of the three models to ensure the consistency of the dimensions of various extracted vectors, which is convenient for the combination and selection of the features extracted by different models. At the same time, normalized normalization is used to preserve the distribution and eliminate the influence of dimension. Then, these feature vectors are spliced horizontally to achieve the purpose of feature fusion. Next, the above-mentioned three features can be employed in the deep forest classification model. With the 5-fold cross validation, the performance can reach 96.3% in Acc, 93.8% in Sn and 96.9% in Sp, respectively (Figure 1).

To prove that the combination of multi-layer LSTM and one-dimensional convolution network is effective for the deep forest, the feature extraction method is extracted by the machine learning method, and the model’s efficiency is explained. Therefore, these feature extraction methods are connected with deep forests, and the prediction accuracy is compared with the effect of connecting other models with deep forests. Specifically, RESNET, multi-layer CNN, random forest, and elastic net are used to compare with multi-layer LSTM network and one-dimensional convolution network. Table 1 shows the comparison results on the evaluation indexes ACC, MCC, F1 score, AUC, Sn, and SP after connecting the deep forest with the six machine learning models as the means of feature extraction and the matrix spliced with the feature vectors extracted from multi-layer LSTM network and one-dimensional convolution network as the feature extraction results. In the machine learning model for feature extraction, 14-layer CNN consists of one convolution layer of 16 3 * 3 convolution cores, one batch normalization layer, one max-pooling layer, one convolution layer of 32 3 * 3 convolution cores, one batch normalization layer, one max-pooling layer, one drop out layer, one convolution layers of 64 3 * 3 convolution cores, one batch normalization layer, one max-pooling layer and one drop out layer, In addition, it is composed of a global average pooling layer, a 32-dimensional density layer, and a 1-dimensional density layer. Of the last two density layers, the first density layer is 32-dimensional, also the data source of the last feature extraction, while the second-density layer is 1-dimensional. The last two density layers and the same idea of feature extraction and weight correction are adopted in the later RESNET, multi-layer LSTM, and one-dimensional convolution models. In RESNET, a data import part is composed of a convolution layer, batch normalization layer, activation layer, and pooling layer of 3 * 3 convolution kernel, a residual part composed of four residual blocks composed of 64, 128, 256, 512 filters, plus the global pooling layer and two 32-dimensional and 1-dimensional density layers as above. Because the results of some classifiers fluctuate greatly, this paper runs each classifier many times, takes the result with the highest accuracy in each time as the running result of that time, and runs 10 times to take the average value as the final result of the model.

It can be seen from the results that on ACC and MCC, it can be found that both multi-layer LSTM and one-dimensional convolution networks are ahead of other models. After combination, the ACC value is increased by 0.031 compared with multi-layer LSTM and 0.013 compared with a one-dimensional convolution network. At the same time, the score of the combination of multi-layer LSTM and one-dimensional convolution network is the highest on the F1 score, which not only exceeds other machine learning methods but also improves by 0.075 and 0.034 respectively, compared with multi-layer LSTM and one-dimensional convolution network, indicating the robustness of the model. On the evaluation index AUC, the combined feature extraction model is also excellent, which is 0.028 and 0.028 higher than the best multi-layer LSTM and one-dimensional convolution network, respectively, indicating that the combined model has the best generalization performance than other models. The above results show that the stitching feature can improve the model’s classification performance.

In addition, this paper also compares convolution with different structures and LSTM classifiers. As above, we run 10 times and take the average value as the final result of the model. For CNN, this paper compares the 7-layer CNN composed of a 1-layer 16-dimensional convolution, a batch normalization layer, a max-pooling layer, a dropout layer, a global average pooling layer, a 32-dimensional density layer, and a 1-dimensional density layer, and the convolution layer composed of two 16 3 * 3 convolution cores, a batch normalization layer, a max-pooling layer, two 32 3 * 3 convolution cores, and a batch normalization layer, A multi-layer CNN network composed of one max-pooling layer, two convolution layers of 64 3 * 3 convolution cores, one batch normalization layer, one max-pooling layer, plus a global average pooling layer, a 32-dimensional sense layer, and a 1-dimensional sense layer is compared with the 14 layer CNN network previously used. It is found that no CNN of any structure has achieved an available feature extraction effect. Nevertheless, Compared with other CNN networks, the above 14-layer CNN model has achieved better results. Firstly, it shows that the appropriate number of network layers in the CNN network has a specific impact on the results. Secondly, the above 14-layer CNN network can be considered as representative and reference significance; For LSTM, this paper compares the three-layer LSTM network with a single LSTM layer as the feature extraction, and achieves similar results, indicating that increasing the number of layers in Golgi protein classification does not bring ideal results; Finally, for one-dimensional convolution, the 10 layers one-dimensional convolution model is compared. This model consists of a convolution layer of 32 11 * 1 convolution cores, a max-pooling layer, a dropout layer, and a convolution layer of 64 11 * 1 convolution cores, a max-pooling layer, a dropout layer, an average pooling layer, and a dropout layer, plus two sense layers. The first sense layer is also 32 dimensions as the final feature extraction data source, and The second density layer is 1D. It is found that the multi-layer one-dimensional convolution model used in this paper has a better effect.

This paper also uses different dimensionality reduction methods to extract the information of the original data, which are reduced to 32 dimensions, the same as the dimension removed by the classifier before and then directly put into the deep forest classifier for classification. This paper attempts PCA, Lasso, linear regression, PHATE, and ridge ones. These dimensional reduction methods are standard data analysis methods commonly used for dimensional reduction of high-dimensional data and can be used to extract the main feature components of data. To maintain the consistency of the conditions for obtaining the results, each classifier is run several times. The result with the highest accuracy each time is taken as the running result of the current time. It is consistent with the information extracted by the above classifier, and it is run 10 times to take the average value as the final result of the model.

These experiments found that PHATE and ridge regression dimensional reduction have excellent extraction ability for sequence information. Based on the stitching feature information extracted by LSTM and one-dimensional convolution, this paper adopts the same horizontal stitching as above and then further stitches the feature data extracted by PHATE reduction and ridge regression reduction in the same way. To verify the stability of the model, each classifier is run several times, and the result with the highest accuracy each time is taken as the running result of the current time. Due to the increase in data and the relatively stable classification result, the average value is taken as the final result of the model after running five times. Table 2 compares our last model and other models in the evaluation indexes ACC, MCC, F1 score, AUC, Sn, and Sp. For the features extracted from multi-layer LSTM, one-dimensional convolution, and PHATE, after normalization by normalize method, the extraction effect is best using the information fusion method described above, and its ACC value reaches 0.963, while MCC is improved by 0.043. The AUC value of PHATE can reach 0.959, which has excellent generalization performance. Moreover, when comparing the results of LSTM + one-dimensional convolution without PHATE, the ACC value and AUC value is increased by 0.016, proving the effectiveness of PHATE fusion. At the same time, it is also found that the performance of ridge regression decreases after stitching with the extracted features of the two classifiers, which interferes with the classification of the deep forest by the underlying primary classifier (Table 3).

Among them, the model A is the feature extraction combination of multi-layer LSTM, one-dimensional convolution, and ridge regression, the Proposed method is the feature extraction combination of multi-layer LSTM, one-dimensional convolution, and PHATE, which is also the final model, and Model B is the feature extraction combination of multi-layer LSTM, one-dimensional convolution, PHATE, and ridge regression.

This work, the Gogli_DF model has been proposed to classify the Golgi proteins with the deep forests model. Firstly, the UniRep method to achieve 1900-dimensional vector features. Secondly, the SMOTE is employed to deal with the imbalance issue. And then several reconstruction feature methods include Light GBM, one-dimensional convolution, multi-layer LSTM, and PHATE. With the reconstructed features, the deep forest algorithm can be employed as the classification model in this work. With this classification model proposed, several issues can be taken into account.

With the development of big data technology and bioinformatics, the number of available protein sequences has increased significantly. However, due to the complex composition of proteins, it is not easy to classify protein sequences correctly with some traditional methods. Therefore, using machine learning to classify proteins has excellent advantages, and some dimensionality reduction methods of cell sequences can also improve the effect of machine learning model. Firstly, through the pre-training network UniRep method and light GBM dimensionality reduction SMOTE method, this paper unifies the sequences of different lengths into fixed-length feature vectors with relatively fewer features, fully retains various feature information, and solves the problem of the unbalanced classification issue. Through the feature fusion of the multiple information extracted by the machine learning model, including one-dimensional convolution, multi-layer LSTM network, and dimension reduction method PHATE, taking into account the influence of dimension and maintaining the original distribution, we fully mine various information and finally use the deep forest for the final classification. The experimental results show that this method has an excellent performance in the classification of Cis-Golgi proteins and trans-Golgi proteins. At the same time, it is found that the appropriate feature stitching method is helpful to improve the performance, while the effect of feature extraction of some models is good. Still, the performance decreases when combined with other models. Meanwhile, when stitching features, we can consider not only the machine learning model but also the integration of appropriate biological dimensionality reduction methods. The machine learning model can also help improve performance. In the future, in addition to studying better fusion methods, this method and idea can become a powerful tool for bioinformatics and protein research.

Publicly available datasets were analyzed in this study. This data can be found at: https://github.com/baowz12345/golgiDF.

WB and YG can be treated as the co-first authors. WB conceived the method. YG designed the method. BC designed the website of this algorithm. WB and BC conducted the experiments. WB and HY wrote the main manuscript text. All authors contributed to the article and approved the submitted version.

This work was supported by the National Natural Science Foundation of China (Grant No. 61902337), Xuzhou Science and Technology Plan Project (KC21047), Jiangsu Provincial Natural Science Foundation (No. SBK2019040953), Natural Science Fund for Colleges and Universities in Jiangsu Province (No. 19KJB520016) and Young Talents of Science and Technology in Jiangsu and ghfund202302026465, Qing Lan Project in Jiangsu, and Qingmiao project of Xuzhou first People’s Hospital.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.