Drug Discovery is an essential endeavor to tackle threats to human health. Notwithstanding, the development and subsequent market penetration of new pharmaceuticals is a critical yet time-consuming and expensive process. To address shortcomings in this process, new approaches have been explored to combine both experimental and computational routes. In particular, as an in-silico approach, virtual screening is proposed as an alternative to identify active molecules towards therapeutic biological targets. Recent advances in artificial intelligence (AI) have provided more effective search algorithms while reducing the time it takes to do virtual screening and implementing new therapeutic candidates.

In this research line, we employ different state-of-the-art techniques of deep learning (specifically Natural Language Processing (NLP), Artificial Vision, and Graph Networks) to model relationships between molecules, proteins, and the variables involved in the task of predicting the affinities among ligands and targets.

Smart Pooling is an Artificial Intelligence (AI) system that increases the number of COVID-19 samples that can be tested with the same amount of resources.

Massive molecular testing for COVID-19 has become fundamental to moderate the spread of the disease. While pooling samples can enhance the efficiency of testing, current pooling strategies are only useful at very early stages of the epidemic.

We proposed the use of clinical and sociodemographic information of patients, to train an AI model to predict the probability of a sample being positive for COVID-19. With this information, we proposed a novel pooled testing protocol that triples the efficiency in COVID-19 testing when 6% of the population is infected and could also improve efficiency for prevalences up to 50%. Smart Pooling is an innovative application of Machine Learning that augments the efficiency of testing regardless of the prevalence of COVID-19 or the selected pooling strategy.

In November 2020, with the generous support of The Rockefeller Foundation, Universidad de Los Andes conducted a research study to assess the possible use of Smart Pooling to increase COVID-19 testing capacities in Bogotá, Colombia. Study results revealed that the adoption of Smart Pooling had the potential to increase testing efficiency by 27% over a four-month period in Bogotá. Through this study, the research team was able to identify bottlenecks in COVID-19 testing processes, which were then shared with key stakeholders along with recommendations to improve testing efficiency. Discussions have also begun with the Instituto Nacional de Salud and the Pan American Health Organization to demonstrate Smart Pooling’s potential use for the current COVID-19 pandemic and future pandemic response initiatives.

congenital heart defects are changes in the function and shape of the heart that are developed during cardiac embryogenesis and are evident at birth. During prenatal ultrasounds, the OB-GYN doctor should completely review the fetal heart to determine if there is any malformation in it. In this way, when having suspicion of a heart defect, it is expected to be treated before delivery or at birth time to increase the baby’s life chance. However, congenital heart defects are not detected efficiently, which is why they cause 24% of deaths in the first year of infants’ life.

For this reason, we propose an Artificial Intelligence algorithm that accompanies doctors in the diagnosis of congenital heart defects. This project is developed within the framework of the program “DELFOS: Diagnosis, training, logistics, education, organization and monitoring of patients with congenital heart defects” and is in charge of integrating the data obtained in the stage of congenital heart defects follow-up and training of screening centers.

Biomedical images are useful for diagnosis, treatment, and follow-up of patients with diverse pathologies and conditions. AI-based methods are tools to analyze these images. However, this task requires the intervention of specialized medical personal, and the interpretation of the images is dependent on the perception and expertise of the specialist analyzing each image. Medical image analysis requires considerable processing times and significant use of computational resources. Due to the complexity of the data, the anatomical variations, different modalities, three-dimensionality, among other challenges of this data.

We focus on the segmentation task, i.e., to find and differentiate a structure of interest within an image. This structure could be an organ or a pathology and could be in multiple locations and have different morphologies and sizes. Thus, we are developing a general segmentation method that has an excellent performance in various tasks and is independent of the possible variations and modalities found in data. We are implementing our methodology using the data provided by the Medical Segmentation Decathlon. That is a challenge with ten different categories for different organs and pathologies.