Bench-ML: A Benchmarking Web Interface for Machine Learning Methods and Models in Genomics

Machine learning is a complex but essential technology in genomics data analysis and its popularity has increased the rate of new methodological approaches published but this raises the question of how models should be benchmarked and validated. Bench-ML is a generalizable and easy to use web interface for benchmarking and validation that can preprocess data, train, test, evaluate and compare machine learning algorithms for genomics. It makes benchmarking machine learning methods more accessible by enabling genomics scientists to perform end-to-end analyses, visualize results and evaluate performance or metrics to compare methods and models by providing a point of reference using only a web browser. To improve something it needs to be measured; To benchmark and evaluate models Bench-ML provides several strategies, methodologies, and tools to generate measurements and visualizations to track experiments to help identify areas of opportunity using metrics such as loss and accuracy, model visualization, learning and saturation curves, principal component analysis, feature scoring, confusion matrix, regression for training and test data, mean absolute error, etc. Bench-ML explains the different options to test and validate machine and deep learning models to identify problematic areas and potentially improve performance. Bench-ML provides several strategies to improve performance like showing when a model is not performing or when different hyperparameters values could be needed, it also helps fine tune hyperparameter values and to identify accuracy across multiple classes and from these classes which class could affect performance. The selection, development, and comparison of machine learning methods and models in genomics datasets can be a daunting task based on the goals of a particular study or the target problem. Machine learning is very good at pattern recognition but modeling the world is much more than that so how to know if a machine learning method or model is performing at a good sensitivity and specificity in large genomics datasets is still a big problem and this is where Bench-ML can help.