Summary

Machine learning (ML) for classification and prediction based on a set of features is used to make decisions in healthcare, economics, criminal justice and more. However, implementing an ML pipeline including preprocessing, model selection, and evaluation can be time-consuming, confusing, and difficult. Here, we present mikropml (pronounced “meek-ROPE em el”), an easy-to-use R package that implements ML pipelines using regression, support vector machines, decision trees, random forest, or gradient-boosted trees. The package is available on GitHub, CRAN, and conda.

Statement of need

Most applications of machine learning (ML) require reproducible steps for data pre-processing, cross-validation, testing, model evaluation, and often interpretation of why the model makes particular predictions. Performing these steps is important, as failure to implement them can result in incorrect and misleading results (Teschendorff 2019; Wiens et al. 2019).

Supervised ML is widely used to recognize patterns in large datasets and to make predictions about outcomes of interest. Several packages including caret (Kuhn 2008) and tidymodels (Kuhn et al. 2020) in R, scikitlearn (Pedregosa et al. 2011) in Python, and the H2O autoML platform (H2O.ai 2020) allow scientists to train ML models with a variety of algorithms. While these packages provide the tools necessary for each ML step, they do not implement a complete ML pipeline according to good practices in the literature. This makes it difficult for practitioners new to ML to easily begin to perform ML analyses.

To enable a broader range of researchers to apply ML to their problem domains, we created mikropml, an easy-to-use R package (R Core Team 2020) that implements the ML pipeline created by Topçuoğlu et al. (Topçuoğlu et al. 2020) in a single function that returns a trained model, model performance metrics and feature importance. mikropml leverages the caret package to support several ML algorithms: linear regression, logistic regression, support vector machines with a radial basis kernel, decision trees, random forest, and gradient boosted trees. It incorporates good practices in ML training, testing, and model evaluation (Topçuoğlu et al. 2020; Teschendorff 2019). Furthermore, it provides data preprocessing steps based on the FIDDLE (FlexIble Data-Driven pipeLinE) framework outlined in Tang et al. (Tang et al. 2020) and post-training permutation importance steps to estimate the importance of each feature in the models trained (Breiman 2001; Fisher et al. 2018).

mikropml can be used as a starting point in the application of ML to datasets from many different fields. It has already been applied to microbiome data to categorize patients with colorectal cancer (Topçuoğlu et al. 2020), to identify differences in genomic and clinical features associated with bacterial infections (Lapp et al. 2020), and to predict gender-based biases in academic publishing (Hagan et al. 2020).

mikropml package

The mikropml package includes functionality to preprocess the data, train ML models, evaluate model performance, and quantify feature importance (Figure 1). We also provide vignettes and an example Snakemake workflow (Köster and Rahmann 2012) to showcase how to run an ideal ML pipeline with multiple different train/test data splits. The results can be visualized using helper functions that use ggplot2 (Wickham 2016).

While mikropml allows users to get started quickly and facilitates reproducibility, it is not a replacement for understanding the ML workflow which is still necessary when interpreting results (Pollard et al. 2019). To facilitate understanding and enable one to tailor the code to their application, we have heavily commented the code and have provided supporting documentation which can be read online.

Running ML

The main function in mikropml, run_ml(), minimally takes in the model choice and a data frame with an outcome column and feature columns. For model choice, mikropml currently supports logistic and linear regression (glmnet: Friedman et al. 2010), support vector machines with a radial basis kernel (kernlab: Karatzoglou et al. 2004), decision trees (rpart: Therneau et al. 2019), random forest (randomForest: Liaw and Wiener 2002), and gradient-boosted trees (xgboost: Chen et al. 2020). run_ml() randomly splits the data into train and test sets while maintaining the distribution of the outcomes found in the full dataset. It also provides the option to split the data into train and test sets based on categorical variables (e.g. batch, geographic location, etc.). mikropml uses the caret package (Kuhn 2008) to train and evaluate the models, and optionally quantifies feature importance. The output includes the best model built based on tuning hyperparameters in an internal and repeated cross-validation step, model evaluation metrics, and optional feature importances. Feature importances are calculated using a permutation test, which breaks the relationship between the feature and the true outcome in the test data, and measures the change in model performance. This provides an intuitive metric of how individual features influence model performance and is comparable across model types, which is particularly useful for model interpretation (Topçuoğlu et al. 2020). Our introductory vignette contains a comprehensive tutorial on how to use run_ml().

mikropml pipeline

Acknowledgments

We thank members of the Schloss Lab who participated in code clubs related to the initial development of the pipeline, made documentation improvements, and provided general feedback. We also thank Nick Lesniak for designing the mikropml logo.

We thank the US Research Software Sustainability Institute (NSF #1743188) for providing training to KLS at the Winter School in Research Software Engineering.

Funding

Salary support for PDS came from NIH grant 1R01CA215574. KLS received support from the NIH Training Program in Bioinformatics (T32 GM070449). ZL received support from the National Science Foundation Graduate Research Fellowship Program under Grant No. DGE 1256260. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation.

Author contributions

BDT, ZL, and KLS contributed equally. Author order among the co-first authors was determined by time since joining the project.

BDT, ZL, and KLS conceptualized the study and wrote the code. KLS structured the code in R package form. BDT, ZL, JW, and PDS developed methodology. PDS, ES, and JW supervised the project. BDT, ZL, and KLS wrote the original draft. All authors reviewed and edited the manuscript.

Conflicts of interest

None.