Many neuroscience experiments involve manual scoring of animal behavior by a human operator. This operation is time consuming, is liable to bias and hinders reproducibility. For these reasons, automatic systems to score animal behavior from videos have gained increasing attention in recent years. Here, we present a machine learning approach to automatically score behavior in the Forced Swim Test (FST). During FST, animals perform three different behaviors: climbing, swimming and immobility. So far, these behaviors have been evaluated by an operator (a bias-prone method) or using automatic systems that extrapolate the degree of mobility of the animal and classify it in three levels (which cannot distinguish between the three behaviors). In our work, a total of 503 minutes of videos were used to train a ML algorithm to automatically classify behaviors from videos. First, videos were labeled by two independent trained operators to decrease bias and any mismatch was evaluated by both. Videos were split into 3-seconds segments and the main behavior was selected for each segment. Different ML models were trained and that reaching the best accuracy (88%) on the test set was trained on the whole dataset. To validate the model, male Wistar rats were administered desipramine, which has been shown to increase climbing and reduce immobility in the FST. Behavior was scored both by an operator and the ML model. Both analyses showed a significant increase of climbing and reduction of immobility in rats treated with desipramine, indicating the robustness and the consistency of our model.