The Forced Swim Test (FST) is a widely used preclinical model for assessing antidepressant efficacy, studying stress response, and evaluating depressive-like behaviours in rodents. Over the last 10 years, more than 5,500 scientific articles reporting the use of the FST have been published. Despite its widespread use, the FST behaviours are still manually scored, resulting in a labor-intensive and time-consuming process that is prone to human bias and variability. Despite eliminating some biases, existing automated systems are costly and typically only able to distinguish between immobility and active behaviours. Therefore, they are often unable to accurately differentiate the major subtypes of movement patterns, such as swimming and climbing. To address these limitations, we propose a novel approach based on machine learning (ML) using a three-dimensional residual convolutional neural network (3D RCNN) that processes video pixels directly, capturing the spatiotemporal dynamics of rodent behaviour. Our ML model was validated against manual scoring in rats treated with fluoxetine and desipramine, two antidepressants known to induce distinct behavioural patterns. The ML model successfully differentiated among swimming, climbing, and immobility behaviours, demonstrating its potential as a standardized and unbiased tool for automatized behavioural analysis in the FST. Subsequently, we successfully validated our model by testing its ability to distinguish between drugs that predominantly evoke climbing (i.e., amitriptyline), those that preferentially facilitate swimming (i.e., paroxetine), and those that evoke both in a more balanced manner (i.e., venlafaxine). This approach represents a significant advancement in preclinical research, providing a more accurate and efficient method to analyze forced swimming data in rodents. We anticipate that due to its adaptability, in addition to the FST, the model could be applied to various behavioural tests in laboratory animals.

Volatile Organic Compounds (VOCs) are organic molecules that have low boiling points and therefore easily evaporate into the air. They pose significant risks to human health, making their accurate detection the crux of efforts to monitor and minimize exposure. Infrared (IR) spectroscopy enables the ultrasensitive detection at low-concentrations of VOCs in the atmosphere by measuring their IR absorption spectra. However, the complexity of the IR spectra limits the possibility to implement VOC recognition and quantification in real-time. While deep neural networks (NNs) are increasingly used for the recognition of complex data structures, they typically require massive datasets for the training phase. Here, we create an experimental VOC dataset for nine different classes of compounds at various concentrations, using their IR absorption spectra. To further increase the amount of spectra and their diversity in term of VOC concentration, we augment the experimental dataset with synthetic spectra created via conditional generative NNs. This allows us to train robust discriminative NNs, able to reliably identify the nine VOCs, as well as to precisely predict their concentrations. The trained NN is suitable to be incorporated into sensing devices for VOCs recognition and analysis.

Drinking behavior is an important parameter in laboratory animal experiments. Traditionally, the amount of fluid consumed is recorded manually by the experimenter, who weighs the bottle at predetermined timepoints to calculate the intake. This approach is labor-intensive and involves human intervention, which can disturb the animals and affect their natural drinking patterns. Automated drinkometer systems have been developed to enable continuous and automated recording of fluid consumption. One advantage of these devices is that they allow researchers not only to measure total fluid intake but also to analyze the microstructure of drinking behavior (e.g., number of licks, bouts, and inter-bout intervals). However, these systems typically require dedicated cages equipped with proprietary drinkometers, making them highly expensive and often unsuitable for monitoring fluid consumption under standard home cage conditions. Here, we describe the development of a new cost-effective microcontroller-based device for continuous monitoring of liquid consumption and analysis of licking microstructure, which can be integrated with existing cages. The system is based on capacitive sensing technology for the assessment of liquid consumption. A customized bottle handler and capacitive sensor have been designed, which can be mounted on a standard 250 mL sized bottle for data acquisition. The device has been tested both in mice and in rats. The results demonstrated that for both animal species the system reproducibly and reliably recorded fluid consumption, also allowing for licking analysis. The data are acquired autonomously through custom devices and directly transferred to a computer for storage and analysis.

The pollution of plastic materials represents one of the most important environmental challenges, due to the rapid and uncontrolled increase in their production, consumption, and use. Therefore, the recycling and valorization of post-consumer plastics are strategic solutions to mitigate this phenomenon. These processes are subject to rigorous international regulations that require recycled polymers to exhibit physical properties and quality consistent with those of their virgin counterparts. To achieve this goal, a thorough and accurate classification of plastic materials before and after recycling is essential. Infrared (IR) spectroscopy has emerged as a fundamental technique for this purpose, enabling non-destructive identification of polymers through their characteristic vibrational features. However, the high spectral similarity between different plastics often leads to misclassification. To overcome this limitation, in this paper, we integrate IR spectroscopy with advanced Machine Learning (ML) algorithms. We develop and validate a rapid, automated ML-based classifier for the identification of four prevalent plastic types: HDPE-B, HDPE-P, LDPE, and PP. The classification models were trained on a dedicated IR spectral database generated from original experimental measurements. Using the open-source Quasar platform, we performed a comparative analysis between traditional ML algorithms and standard Convolutional Neural Network (CNN) architecture. The study shows that both CNN and classical algorithms, in particular Random Forest (RF), achieve optimal performance in the classification task, with accuracies of 1.000 and 0.998, respectively. Our results demonstrate that the integration of deep learning architectures with IR data significantly enhances the accuracy and reliability of plastic recognition, providing a robust tool to support the industrial transition toward high-quality recycled polymers.