Soft computing methods have proved to be a viable way of tackling some of the most difficult scientific challenges and practical problems we face nowadays. Recent advances in machine learning have led to an increased usage of artificial intelligence in various domains, such as medicine, mechanical engineering, robotics, and transportation. The quality of machine learning models is usually assessed only through their capability of solving a specific task; however, we argue that energy consumption ought to be taken into consideration as well, as the current trend of creating larger and more complex models has had a big influence on the energy footprint such machine learning models exhibit. Depending on the problem being tackled, the amount of data used, the complexity of the model, and the utilized hardware, the energy consumption can rise to \(3.6 \times 10^{9}\)  J. One way to reduce such high energy consumption is to pre-process the training data prior to model training and retain only the most important features describing each data instance. Here, we conduct three case studies examining the energy efficiency of machine learning models for multivariate regression, multiclass classification, and binary classification. Using our previous work on therapeutic peptide prediction as an experiment baseline, we conclude that by reducing the number of features characterizing each peptide from 94 to 45, 119 J of energy are saved per model training cycle while retaining a similar degree of accuracy. Considering the iterative nature of finetuning the machine learning models in numerous training cycles and retraining when additional data is gathered, we believe such an approach is of crucial importance for the sustainability of developing machine learning-based systems.