TABLE 1.
Machine learning methods used for research on the association between miRNAs and neurodegenerative diseases.
| Methods | Positive aspects | Negative aspects | Examples of use in the context of miRNA analysis |
| Ridge | Reduces the impact of variables that are not important for the prediction | Doesn’t eliminate irrelevant variables | PMID: 26947266 |
| Lasso | Reduces overfitting by adding a penalty to coefficients the model overemphasizes and eliminates them | Doesn’t take into account multicollinearity in the model and could eliminate relevant independent variables | PMID: 34048985 |
| PMID: 33316739 | |||
| PMID: 21743061 | |||
| Elastic net | Combines both Lasso and Ridge aspects: It eliminates some variables while reducing the impact of some other variables | Computationally more expensive than LASSO or Ridge | PMID: 35113902 |
| PMID: 29513198 | |||
| Decision tree | Very simple to understand and visualize | Subject to overfitting | PMID: 26649272 |
| Doesn’t work well with imbalanced data | |||
| Very different trees can be generated if a small chance in the data is made | |||
| Random forest (RF) | Can deal with imbalanced datasets and missing data | The number of nodes in decision trees will grow exponentially with depth | PMID: 29056906 |
| Being an ensemble of decision trees, overfitting is not a problem | The prediction needs to be uncorrelated | PMID: 23922946 | |
| Gradient boosted decision trees (GBDT) | More accurate than RF | Sensibility to outliers | PMID: 32604706 |
| Doesn’t need bootstrap sampling like RF | Overfitting can be a problem when too many trees are added | PMID: 35051896 | |
| Support vector machine (SVM) | Works well with 2D, 3D, or higher dimensions | Computationally more expensive for larger datasets | PMID: 29275361 |
| PMID: 24417022 | |||
| Outliers have less impact on the prediction since the hyperplane is influenced by the support vectors (data points closer to the hyperplane) | Works poorly if the dataset has overlapped classes | PMID: 34442108 | |
| Artificial neural networks (ANN) | Work very well with huge amount of data | Can be quickly computationally and time consuming | PMID: 30504368 |
| Can handle unstructured data | Big dependence on the training data, so overfitting can be a problem | PMID: 22349176 | |
| PMID: 30519653 | |||
| k-means | Very simple algorithm to implement | Lack of robustness with big data analysis | PMID: 34879829 |
| Choosing K can be difficult | PMID: 32493067 | ||
| Doesn’t work well with imbalanced data or outliers | PMID: 22255820 | ||
| Weighted correlation network analysis (WGCNA) | Retains connectivity of nodes | Can lack biological precision | PMID: 32699331 |
| PMID: 34225819 | |||
| Bayesian network | Can handle missing data and avoid overfitting | Need for sensitivity analysis, to be applied to the outcome | PMID: 23690582 |
| PMID: 32368197 |