The data used in this study are obtained from the NCBI Sequence Read Archive (SRA045646, SRA050230) provided by Qin et al. (2012) for T2D; accession number PRJNA398089 in the SRA for the Integrative Human Microbiome Project for IBDMDB (Beghini et al., 2021). IBD dataset is obtained from the MetaHit project (Marco-Ramell et al., 2018) (ERA000116). The CRC metagenomic dataset containing 1,262 samples was created by Beghini et al. (2021). Microbiome sequencing data is classified into disease states based on the metadata associated with them. To ensure data quality, we applied quality filtering to meet the standards outlined in the Human Microbiome Project Consortium SOP (2012), as referenced in Thomas et al. (2019). This procedure allowed us to categorize the raw sequencing data according to relevant disease states, enabling our subsequent analyzes. The microbiome samples were associated with the microbial species of origin (taxa) using the MetaPhlAn tool, and the relative abundance composition for each taxon was generated accordingly. These taxa and their relative abundances serve as features or variables in our machine learning approaches. MetaPhlAn first assigns reads to microbial clusters using clade-specific genes for assignment. It then presents the relative abundance of microbial taxa based on these readings. In this study, the assignment to microbial species of origin (taxa) was determined for each DNA sequence using the MetaPhlAn tool. The relative abundance value is normalized by dividing the number of reads for each taxonomic level by the total number of reads for only one sample. In this way, the taxonomic abundance values are expressed as real numbers in the range [0,1] with a sum of 1 for each sample. Samples with less than 1 million total reads were not included in our study. For each sample, we determined the diversity of disease-relevant microbiomes, where diversity represents the presence and relative abundance of microorganisms (Alatawi et al., 2022).

The four microbiome datasets used to evaluate the microBiomeGSM tool are listed in Table 1. The table presents the number of samples in each dataset and the number of samples that are labeled as positive. Positive samples refer to patients, while negative samples refer to controls. Each dataset contains the abundance values of the species, which we consider as features. We have considered 3 taxonomic levels for creating the groups, i.e., genus, family, and order. For each dataset, the number of extracted groups is listed in the corresponding column, while ‘-’ denotes missing information.

Statistical information regarding the numbers of features in each group is given in Table 2. For each data set and for each taxonomic level (genus, family, and order), the average, maximum, and minimum numbers of features within a group are given.

Supplementary Table S1 shows the distribution of the groups based on their sizes for the IBDMDB dataset. The numbers in the table indicate the number of groups that have the specified number of species for that specific taxonomic level. There are 187, 77, and 43 groups for genus, family and order levels, respectively. About 90% of the groups at the order level, about 90% of the groups at the family level, and about 97% of the groups at the genus level contain 20 or fewer species for the IBDMDB dataset.

Our proposed method, microBiomeGSM, consists of three main components: Grouping, Scoring, and Modeling (G-S-M). The G-S-M approach has been used in other studies that consider the pre-existing biological knowledge (Yousef et al., 2019, 2021a,c, 2022a; Qumsiyeh et al., 2022; Yousef and Voskergian, 2022; Ersoz et al., 2023; Jabeer et al., 2023). Additionally it was modified to integrate two-omics datasets such as the miRcorrNet and miRModuleNet tools (Yousef et al., 2021a, 2022b); and even to integrate 3 omics datasets such as 3Mint tool (Unlu Yazici et al., 2023). Interested readers can find further details about those approaches in our recent reviews (Yousef et al., 2021b; Kuzudisli et al., 2023).

Utilizing the G-S-M approach, microBiomeGSM performs a search to identify the most important taxonomic groups in disease-associated metagenomic datasets. The relative abundance values of the species within the group can be checked for each sample; and the generated model decides whether the sample has the disease or not. By focusing on a specific taxonomic level, we can use the G component to find the most significant group for the disease under study. This approach provides the advantage of focusing on either the macroscopic or microscopic view of the most important group to distinguish between healthy samples and patient samples. An overview of the steps performed in microBiomeGSM is presented in Figure 1.

Let X be the two-class dataset consisting of the species in the columns, and samples in the rows including the class labels (1 denoting the disease state and 0 denoting the healthy state). To understand the approach in detail, let us assume that the taxonomic level is selected as “genus” for the “Select taxa rank” step in Figure 1. The input X_abd (abundance matrix) is first split into a training set (X_train) and a test set (X_test) with a ratio of 80:20 based on the class labels. Denote by S the feature space of all species in X_abd and by U_genus all unique genera for S. Grp{} denotes the selection function of each U_genus in S, grouping all species on the basis of similar genuses. Grp{U_genusⁱ for S} represents each genus in S, with all the species grouped by genus. For example, if we take Alistipes as one of the genus in U_genus, we get the following when we apply the Grp function.

Grp{U_genusⁱ}, where i = Alistipes and ∈ S.

Grp{Alistipes} = {alistipes_finegoldi, alistipes_indistinctus, alistipes_inops, alistipes_shahii}.

Similarly, this approach is applied to all genuses that are present in X_abd, and a list of genus groups is created, as shown in Figure 1 after the select taxa rank step. This is repeated for the three taxonomic levels identified.

When Figure 1 is examined, firstly, in the grouping component G, for all the groups of genus, we partition X_train into sub data denoted as sub_d_x. Following the earlier example of Alistipes, this group yields sub_d_alistipes which is created from X_train. The sub_d_alistipes contains the labels of the samples, but the feature space is restricted only to species within the Alistipes genus. This is applied to all different genera created in the prior step, so we have multiple subsets of data with a feature space specified by genus. Secondly, in the scoring step S, the generated sub_d is trained on a Random Forest classifier with 5-fold cross-validation with randomized stratified shuffling. Each sub_d is given a score equal to the mean of the accuracy over all foldings based on the prediction of the labels. Each sub_d is scored and then sorted based on the score. The top k groups with the highest score are used for the subsequent step. The value chosen for k is 10, but other values for k have been tested. Following the example of selecting genus as the taxonomic level, the top 10 genus groups that show strong discriminative ability are used to build the classification model. Thirdly, in the modeling component, the species from the top 10 genus groups are used to train a Random Forest model with 100-fold Monte Carlo Cross-Validation (MCCV). The top ranking set of species corresponding to the top ranked group is trained on X_train and then tested on X_test. Then, the second set of species corresponding to the second highest scoring group is aggregated with the top scoring set of species; and then used to train and test the model. This process is repeated until all species in the top 10 ranked genus groups are aggregated; and used to train and test the classifier. This whole process is repeated 100 times, stratifying the initial X_abd and randomly splitting it into X_train and X_test without replacement. The classification performance metrics are determined as the average of the metrics obtained in 100 folds. Similarly, the top ranked groups and the top ranked species are retained for each run.

The microBiomeGSM tool utilizes the pre-existing biological knowledge of the assignment of the species into different taxonomic levels, such as genus, family, and order. Experiments with the microBiomeGSM tool were conducted on the open-source KNIME platform (Berthold et al., 2009). This platform can handle a wide range of data types and operations. The user can configure the number of iterations, the rank function, and the number of iterations for MCCV. All rows with missing values are removed within the workflow.

In metagenomics research, it is observed that in studies using taxonomic features, the number of observations used for training data is higher than the number of observations used for testing data. This situation is undesirable if studies are to produce more effective results, and researchers are proposing various methods of resolution, particularly feature selection methods. Although the process of feature selection in disease prediction problems based on metagenome data has not been well studied, the literature suggests that this process may be as important as the choice of a classification method (LaPierre et al., 2019). The process of feature selection in metagenome-based disease prediction could help us learn more about disease development mechanisms. Therefore, further research in this direction is warranted. In metagenomics studies, in order to reduce the number of taxa, i.e., to select informative species (features), min Redundancy Max Relevance (mRMR) (Ding and Peng, 2005), Lasso (Tibshirani, 1996), Elastic Net (Zou and Hastie, 2005), and the iterative sure select algorithm (Duvallet et al., 2017) have been used extensively. Another feature selection method, called Fizzy, addresses the challenge of using classification techniques to identify important functional elements for downstream analysis (Ditzler et al., 2015). Oudah and Henschel presented an alternative taxonomy-based method for feature selection (Oudah and Henschel, 2018). Bakir-Gungor et al. (2021) applied CMIM (Fleuret and Ch, 2004), FCBF (Senliol et al., 2008), mRMR (Ding and Peng, 2005), and Select K best (SKB) (Pedregosa et al., 2011) to type 2 diabetes-associated metagenomics datasets and obtained powerful performance metrics (Bakir-Gungor et al., 2021). Jabeer et al. also proposed a robust classification method for evaluating colorectal cancer associated metagenomic datasets using a combination of feature selection methods and machine learning methods (Jabeer et al., 2022). Bakir-Gungor et al. (2022) also proposed a powerful method for IBD classification with fewer features by combining feature selection methods and machine learning methods (Bakir-Gungor et al., 2022). While these feature selection approaches have produced effective results in a variety of fields, they have only recently been applied to microbiome-based disease prediction problems.

In this study, we have comparatively evaluated microBiomeGSM with different classifiers and with different feature selection methods. As the feature selection methods, we have utilized Select K best (SKB), Fast Correlation Based Filter (FCBF), Extreme Gradient Boosting (XGBoost), Min Redundancy Max Relevance (mRMR), Information Gain (IG), and Conditional Mutual Information Maximization (CMIM). Wang and Liu (2020) compare the performance of classifiers with traditional methods and ensemble methods for disease prediction based on human microbiome data. They use Elastic Network and SVM as traditional methods and Random Forest and Extreme Gradient Boosting (XGBoost) as ensemble methods. In their study, they find that the XGBoost algorithm shows superior performance compared to other algorithms (Wang and Liu, 2020). In another study, Marcos-Zambrano et al. (2021) conducted an important review paper to reveal the links between the microbiome and diseases. In this study, which included information on the performance of machine learning methods, they found that the Support Vector Machines (SVM), Random Forest (RF), k-Nearest Neighbors (k-NN), and Logical Regression (LR) algorithms were widely used. They concluded that when selecting a machine learning algorithm, several factors should be considered such as the set of observations, the set of features, the type of data, and the quality of the data. They suggest using several different methods, comparing them, and choosing the one that provides the best performance value (Marcos-Zambrano et al., 2021).

Accuracy, F1 score, sensitivity, specificity, and AUC were used to evaluate the predictive performance of the proposed models. AUC score is a common measure for performance evaluation and a reliable metric for evaluating balanced datasets. Other metrics such as F1 score, sensitivity, specificity, and accuracy, were used to evaluate the performance of the created models because the dataset for this study has an uneven distribution of classes. When a balance between precision and recall is desired and there is an uneven distribution of classes, the F1 score is a good option among the performance metrics (many true negatives). Several classifiers report the probability values for their predictions, which can also be considered as confidence values for the prediction. The AUC often uses this information to figure out how often incorrect predictions occur at different confidence levels. In real life, test results from positive and negative examples overlap. AUC illustrates how the threshold or cut-off value for identifying positive examples affects the relationship between recall and precision. In this study, all of the above-mentioned metrics were calculated as the mean of 100 times MCCV. After each iteration, we obtain lists of significant taxonomic groups and species associated with these taxa groups for a given disease. To assign scores to the entities in the taxonomic groups list and in the species lists, a prioritization approach is used. For this purpose, we integrated the RobustRankAggreg algorithm (Kolde et al., 2012) and microBiomeGSM. RobustRankAggreg algorithm is available as an R package. Each entity (taxonomic group or species) in the lists is given a value of p by the RobustRankAggreg technique, indicating how highly ranked that entity. Using the RobustRankAggreg tool, microBiomeGSM outputs a list of species to which it has assigned a significance value (value of p) for a specific taxonomic group. Each taxa group is assigned a significance value and the species associated with that group are assigned the same value.

One approach to evaluate model performance in the context of microBiomeGSM is to compare model performance between different values of the parameter k. k represents the number of groups (taxa) used in microBiomeGSM models. This approach can help researchers determine the optimal value of k that balances model complexity and predictive power, ultimately leading to more effective and interpretable models in microbiome-related research. It provides insight into how the inclusion or exclusion of specific taxa affects the overall performance of microBiomeGSM models.

Supplementary Table S2 shows the performance metrics obtained with 100-fold MCCV for the aggregated top 10 groups for four different datasets compared at three different taxonomic levels (genus, family, order) for grouping. For the IBDMDB dataset, microBiomeGSM achieved an AUC of 93% using the top 1 group at the family level. Performance metrics are shown for the top 2 groups via combining species from the first and second highest scoring groups. We obtained an AUC of 97% when the top 2 groups are combined at the family taxonomic level for the IBDMDB dataset. In this way, microBiomeGSM provides cumulative performance results for the top 10 highest scoring groups. For the IBDMDB dataset, the highest performance metric (an AUC of 98%) is obtained using the species from the top 10 groups at the order taxonomic level. For the IBD dataset, the highest performance metric (an AUC of 93%) is obtained using the species from the top 9 groups at the order taxonomic level. For the T2D dataset, the highest performance metric (an AUC of %74) is obtained using the species from the top 9 groups at the order taxonomic level. For the CRC dataset, the highest performance metric (an AUC of %83) is obtained using the species from the top 10 groups at the family taxonomic level. While examining other performance metrics (such as accuracy, sensitivity, specificity in Supplementary Table S2), it is noteworthy that satisfactory results are obtained with microBiomeGSM for each taxonomic level, especially for the IBDMDB dataset. The high sensitivity values that are reported for the CRC, IBDMDB, and IBD datasets display the success of the microBiomeGSM tool in terms of detecting the patient samples. In the CRC, IBDMDB, and IBD datasets, the strikingly high specificity values indicate that the microBiomeGSM tool correctly identifies the negative samples (i.e., individuals who do not have the disease). However, in the T2D dataset, the specificity rate appears to be relatively low compared to the other datasets. Nevertheless, the ability to detect negative samples remains at a reasonable level.

In addition, Figures 2, 3 show the sensitivity and specificity values obtained with the microBiomeGSM tool for all datasets. Figure 2 shows the sensitivity values obtained using the microBiomeGSM tool across all datasets. One can notice from Figure 2A that for the CRC data set the highest sensitivity value (73%) is obtained for the order taxon level using 10 cumulative groups. In particular, the sensitivity values calculated for the IBDMDB dataset were quite impressive, especially in group 1 and group 6, both at the family taxon level, reaching 99% sensitivity value, as shown in Figure 2B. Figure 2C shows another impressive set of results for the IBD data set. In Figure 2C, we observe high values for sensitivity, in particular 87% sensitivity at the taxon level in group 1. As shown in Figure 2D, the highest sensitivity value for the T2D data set is 69%. This result is obtained for the genus taxon level using 10 cumulative groups. A sensitivity value of 69% is also obtained for the family taxon level using 4 cumulative groups.

Figure 3 shows the specificity values obtained using the microBiomeGSM tool for all datasets. As shown in Figure 3A, the specificity value obtained for the CRC dataset is remarkable, reaching an impressive specificity value of 94% at the family taxon level for 1 group. Figure 3B depicts that the highest specificity value obtained for the IBDMDB dataset is 93% for 1 group at the order taxon level. As displayed in Figure 3C, the highest specificity value obtained for the IBD dataset is 85% for the 4 cumulative groups at the order taxon level. The same result is also obtained at the order taxon level for the 5 cumulative groups. One can notice in Figure 3D that the highest specificity value that is obtained for the T2D dataset is 71% for the 6 cumulative groups at the order taxon level.

The number of significant groups used to train the model could affect the performance of microBiomeGSM. Table 3 shows the influence of the number of groups and the number of species at family, genus and order levels on four datasets. Table 3 presents the performance of the top 10 cumulative groups and top 1 group for each taxonomic level on different tested datasets. For the IBDMDB dataset, for the family taxonomic level, one can observe that the AUC increases by 5% when we consider the top 10 significant groups cumulatively, while we increase the number of species from 34 to 205. On the same dataset, an increase of 8% in AUC score is observed at the Genus taxonomic level via increasing the number of species from 34 to 119. For the same dataset, a decrease of 1% is observed at the Order taxonomic level. Order taxonomic level using the top group that includes 98 species achieves the highest AUC success rate of 98% for the IBDMDB dataset. Similarly, family taxonomic level using the top 10 combined groups achieves 97% AUC on the IBDMDB dataset, but these 10 combined groups include a much higher number of species (205 species). For the IBD dataset, the highest AUC value of 91% was obtained using the microBiomeGSM tool. This value at the family taxonomic level was obtained by cumulatively combining 10 groups, using an average of 260.4 species. For the T2D dataset, the highest AUC value of 72% was obtained using the microBiomeGSM tool. This value, obtained at the order taxanomic level, was obtained by combining 10 groups cumulatively. For 1 group, an average of 138.28 species are used at the taxonomic level, while for 10 groups, an average of 596.99 species are used. For the CRC dataset, the highest AUC value of 87% was obtained using the microBiomeGSM tool. This value at the order taxanomic level was obtained by cumulatively combining 10 groups, using an average of 604 species.

microBiomeGSM reports important groups of features that are detected at different taxonomic levels for the disease under study. Table 4 lists the top 10 important groups that are identified by microBiomeGSM for three different taxonomic levels on four different datasets. The identified features are ranked by their importance scores from high to low. The feature with the highest importance value is the strongest candidate to be announced as potential taxonomic biomarker for the disease under investigation.

The microBiomeGSM tool lists a number of associated species for each identified group. The species included in the top 5 significant groups are listed in Supplementary Tables S3–S5 for family, order, and genus taxonomic levels, respectively for four different datasets. All species for the family, order, and genus taxonomic levels for the T2D, IBDMDB and CRC datasets can be found in Supplementary Tables S6–S14, respectively.

For the IBDMDB dataset, the changes in the AUC score when the number of groups is increased from 1 to 10 are shown in Supplementary Figure S1. For the IBDMDB dataset, a high AUC score is obtained at the order taxonomic level. When the number of groups was increased, the AUC score decreased relatively, and no significant change was observed after 5 groups. At the genus and family taxonomic levels, there is a significant increase in the AUC score until 5 groups are combined and no significant change after 5 groups.

Our Grouping-Scoring-Modeling (G-S-M) approach emerges as a paradigm shift from traditional feature selection methods. Instead of pinpointing individual informative features, the GSM methodology groups these features. These groups are then scored, and a classification model is built using these top-ranking feature conglomerates. The versatility of the GSM method, as detailed in our prior work (Yousef et al., 2021b), lies in its adaptability. Groups can be created either by computational/statistical methods or by using domain-specific knowledge. In order to use the GSM strategy for a given dataset, a deep domain expertise is required to skillfully define these groups, which makes each application different. The modifications required to tailor the G-S-M approach to the unique needs of microbiome research highlight the adaptability of the G-S-M method and the novelty of our current study.

We have comparatively evaluated the performance of microBiomeGSM against 4 different classifiers and 6 different feature selection methods using the same datasets. All algorithms are run with default parameters. The developed approach and feature selection methods were executed multiple times, and the results were averaged and shared. Table 5 shows the performance of the different feature selection algorithms and different classifiers on the same disease associated microbiome datasets. In these experiments, the number of features was set to 100. The best result for the IBDMDB dataset is obtained by using the XGBoost feature selection algorithm in combination with the Random Forest classification algorithm with 98% AUC. For the CRC dataset, the best result is obtained by using the XGBoost feature selection algorithm in combination with the Random Forest classification algorithm with an AUC of 85%. For the IBD dataset, the best result is obtained using the Random Forest classification algorithm with 92% AUC and the SKB feature selection algorithm. For the T2D dataset, the best result is obtained by using the XGBoost feature selection algorithm in combination with the Random Forest classification algorithm with 70% AUC.

We would like to note that the primary objective of microBiomeGSM is not to compete with other feature selection methods (FS). Even if microBiomeGSM’s performance is on par with or slightly less favorable than other FS methods, its fundamental contribution lies in identifying the most informative microbiomes. These microbiomes play a pivotal role in aiding researchers in gaining a deeper understanding of the biological underpinnings of the disease under investigation. In essence, microBiomeGSM’s value lies in its ability to contribute to the advancement of biological knowledge, rather than merely outperforming other feature selection techniques.

Table 6 shows the performance metrics of microBiomeGSM for each taxonomic level for four different datasets. The # of species column shows the number of species (features/variables) used to train and test the model. Since the number of species changes in each iteration of MCCV, we also report the standard deviation. Performance metrics are reported as the average of 100 iterations with the corresponding standard deviation. For the CRC dataset, among different classifiers the RF algorithm has the highest performance for all calculated metrics including the accuracy, sensitivity, specificity, precision, and AUC metric. The AdaBoost, LogitBoost and DT models show lower performance compared to the RF model. The performance metrics of these three algorithms are similar but not as high as RF model. At the order taxonomic level, the mean values of the performance metrics are stable and the standard deviations are low. This indicates that the order level is a more appropriate choice for CRC classification. Comparing the RF model and the microBiomeGSM model, similar performance metrics are obtained for the CRC dataset, but it is worth mentioning that the number of features used in the proposed tool is lower. In other words, for the CRC dataset the microBiomeGSM model can accurately classify using fewer taxonomic features. For the IBDMDB dataset, among different classifiers the RF algorithm has the highest accuracy, sensitivity, specificity, precision, and AUC values. In particular, RF model achieved very high sensitivity and AUC values. For the IBDMDB dataset, the microBiomeGSM tool achieves an AUC of 98% for the order taxon level, the same performance metrics as obtained by the RF classification algorithm. However, the microBiomeGSM tool uses 341 features for the order taxon level, while the RF model uses 579 features. For IBD dataset, the RF algorithm generates the highest performance on several metrics, including accuracy, sensitivity, specificity, precision, and AUC. It performs particularly well on sensitivity and AUC. In our analysis, microBiomeGSM achieved an impressive AUC value of 91% at the family taxon level. Equally remarkable is the similar performance of the RF classification algorithm (an AUC of 92%) for the same task. However, it is important to highlight an important difference between these two approaches. For IBD dataset the RF classification algorithm achieved an AUC of 92% by using a much larger set of features (1,456 features) for the classification task. For the same dataset, the microBiomeGSM tool also showed remarkable performance (an AUC value of 91%). In stark contrast, microBiomeGSM achieved nearly equivalent AUC performance while using a much smaller set of features, only 260 features. This divergence in feature usage highlights the effectiveness and potential advantages of the microBiomeGSM tool in extracting meaningful information from microbiome data while optimizing computational resources. For T2D dataset, the RF classification algorithm outperforms other classification algorithms on several performance metrics including accuracy, sensitivity, specificity, precision and AUC. microBiomeGSM achieved an AUC value of 72% at the order taxon level. Interestingly, a similar level of performance is observed using the RF classification algorithm, which achieves an AUC value of 75%. However, it is important to note that the underlying mechanisms of these two methods are very different. The RF classification algorithm achieves this AUC value by incorporating a much larger set of features, 1,456 features, into its classification process. In contrast, the microBiomeGSM tool achieves comparable AUC metric by using a leaner set of 596 features. This difference in feature usage is worth highlighting as it shows that the microBiomeGSM tool is able to deliver competitive results with a lower computational load, making it an efficient and resource-efficient choice for the classification task at hand. These results highlight the nuanced trade-offs in selecting the appropriate tool or algorithm for the specific data analysis requirements.

As shown in Table 7, the performance of our proposed method varies depending on the taxonomic level considered. For the order taxonomic level, for all tested datasets, the proposed method outperforms other models in terms of the AUC score, except for the RF classifier. Similarly, for all datasets, at the family and genus taxonomic levels, the AUC values are also highly competitive, outperforming those of the other four machine learning algorithms used in this study, with the sole exception of the RF classifier. These results highlight the robust performance of our method across different taxonomic levels. A remarkable performance of our proposed method was observed when it is applied on the IBDMDB dataset. Here, we obtained an exceptionally high AUC value of 0.98 ± 0.03 at the order taxonomic level using a 100-fold MCCV approach. This remarkable result demonstrates the exceptional performance and the potential of the microBiomeGSM tool.

This section discusses the biological relevance of the features discovered by microBiomeGSM at different taxonomic levels for all tested datasets. T2D is a metabolic disease characterized by high glucose levels in blood and caused primarily by cellular resistance to the activity of insulin (Sedighi et al., 2017). There are several studies in the literature that have demonstrated the relation of different microorganisms at the genus, family, and order levels with T2D development. For the T2D dataset, the top 10 microbiomes identified by our method at the genus, family, order levels and the relevant literature can be summarized in Supplementary Table S15. On the other hand, inflammatory bowel diseases (IBDs), which include primarily ulcerative colitis and Crohn’s disease, but also non-infectious inflammation of the bowel, have puzzled gastroenterologists and immunologists alike since their first modern descriptions around some 75–100 years ago (Ni et al., 2018; Bakir-Gungor et al., 2022). For the IBDMDB dataset, the top 10 microbiomes identified by our method at the genus, family, and order levels and the relevant literature can be summarized in Supplementary Table S15. CRC is a prevalent malignancy affecting the colon and rectum. It constitutes approximately 10% of all newly diagnosed cancer cases worldwide (Li X. et al., 2023). For the CRC dataset, the top 10 microbiomes identified by our method at the genus, family, and order levels and the relevant literature can be summarized in Supplementary Table S15.

Numerous studies have investigated the relationship between microbiomes and diseases like T2D, CRC, and IBD using similar datasets as used within this study. Upon examination of these studies, it becomes evident that while their experimental designs may vary, they consistently yield comparable results when it comes to identifying microbiomes linked to these diseases. These findings align with the important microbiomes identified by microBiomeGSM for T2D, CRC, and IBD, showcasing the tool’s effectiveness in accurately identifying relevant microbiomes associated with these diseases. These congruent findings reinforce the reliability and validity of the microbiome associations detected by the microBiomeGSM tool. It also underscores the tool’s capacity to identify microbiomes that are consistently linked to specific diseases, providing valuable insights for disease characterization and prediction. Hassouneh et al. (2021) conducted a series of experiments aimed at uncovering microbiomes associated with IBD. In their analysis using the same dataset as used by the microBiomeGSM tool, they observed differences in Clostridium microbiota among IBD patients. Additionally, another microbiome identified for IBD in their study is Ruminococcus. Remarkably, these microbiomes align with the important microbiomes detected for the IBD disease by the microBiomeGSM tool. This correspondence in findings highlights the capacity of microBiomeGSM in identifying relevant microbiomes linked to IBD. Zhang Y. et al. (2022) conducted a study with the goal of identifying disease-associated microbiome species for Inflammatory Bowel Disease Microbiome Database (IBDMDB), employing the same dataset (PRJNA289734) as used in microBiomeGSM. In their research, they highlighted the significance of the Bacteroides microbiome. Interestingly, the Bacteroides microbiome is also identified as one of the important microbiomes by the microBiomeGSM tool proposed in our study. This alignment in findings underscores the effectiveness of microBiomeGSM in recognizing key microbiomes associated with diseases like IBD. Bai et al. (2022) conducted a series of experiments aimed at identifying microbiomes associated with T2D. In their research, they utilized the SRA4565 data for T2D and highlighted the significance of the methanobacteriales microbiome. Notably, methanobacteriales is among the top 10 microbiomes identified by the proposed microBiomeGSM tool. This convergence of findings underscores the effectiveness and utility of the proposed tool in uncovering microbiome associations with diseases like T2D. Forslund et al. (2015) conducted experiments utilizing the same T2D dataset employed by microBiomeGSM to investigate microbiomes associated with T2D. Upon close examination of their experiments, they underscored the significance of the Clostridiales microbiome in relation to T2D disease. Interestingly, Clostridiales also emerges as one of the important microbiomes identified by microBiomeGSM. This convergence in findings highlights the relevance and effectiveness of microBiomeGSM in identifying crucial microbiomes associated with T2D. Ma et al. (2021) conducted a study that investigated the microbiomes associated with CRC using the same dataset as in our study. Among the various microbiomes they examined, the Prevotella microbiome stood out as strongly linked to CRC. This association aligns with the findings of microBiomeGSM, underscoring the significance of the Prevotella microbiome in the context of characterizing CRC. Chen et al. (2023) conducted research using the same dataset to investigate microbiomes in the context of colorectal cancer, akin to the proposed microBiomeGSM tool. Similar to the findings of microBiomeGSM, their study also identified Peptostreptococcus, Fusobacterium, and Porphyromonas microbiomes as valuable and effective biomarkers for CRC. This convergence in results underscores the potential significance of these specific microbiomes in CRC characterization and their importance as potential biomarkers for the disease.

In summary, via analyzing the raw microbiome data of specific diseases, this study aims to identify taxonomic biomarkers that may have a role in the associated diseases. Three different taxon levels (genus, family, and order) are studied and disease prediction is performed by building effective machine learning models using the G-S-M approach. Four different datasets are analyzed and the identified microorganisms at genus, family and order levels are compared with the existing literature.

The quality and the scope of our study have been significantly influenced by several primary limiting factors. These factors encompass the nature of the data set, the tools employed for data preprocessing, the specific taxon groups considered, and the overall volume of data under examination. First and foremost, the data set itself plays a pivotal role in shaping the outcomes and conclusions of our study. Its size, diversity, and representativeness directly impact the generalizability of our findings. Furthermore, the quality of data, its sources, and any potential biases within the dataset significantly affect the reliability of our results. Equally significant is the role of the tools employed for data preprocessing. The choices made in data cleaning, feature selection, and data transformation can introduce variability and influence the robustness of our analytical pipeline. It is paramount to acknowledge how these preprocessing steps can shape the study’s outcomes. Additionally, our study’s focus on specific taxon groups within the dataset should be considered. The selection of these taxonomic levels and the criteria used for their inclusion or exclusion has bearing on the granularity and relevance of our findings. Finally, the number of data points utilized in our analysis is another crucial factor. A larger dataset provides a broader and potentially more representative sample, which can enhance the reliability and statistical power of our results. Conversely, a smaller dataset may limit the generalizability of our conclusions. A comprehensive understanding of these limiting factors is essential for contextualizing our study’s outcomes and conclusions.