Machine Learning for Infection and Disease


Machine Learning for Infection and Disease (MLID) group aims to develop novel computational methods to facilitate our understanding of Infection Biology and Disease Biology. To achieve these aims we combine expertise in Deep Learning, Generative AI, Computer Vision, Biomedical Image Analysis, Simulations and Systems Biology. We develop novel algorithms, tasks and benchmarks to power the new generation of Biomedical AI and Biomedical Discovery.


  • Dr. Artur Yakimovich
  • Jacob Anter
  • Jonathan Berthold
  • Trina De
  • Dr. Pierre Joubert
  • Rui Li
  • Gabriel della Maggiora
  • Ashkan Mokarian
  • Adrian Urbański
  • Maria Wyrzykowska
  • Marina Neira Rey

Research Topics

  • Generative AI for super-resolution microscopy

    Understanding Infection Biology and Disease Biology requires visualising and measuring the invisible microscopic world from molecules through cells to organisms, as well as understanding the occurring interactions on a molecular and spatiotemporal level. Therefore the primary research focus of the group is to develop novel biomedical algorithms to be used in BioMedical Image Analysis and Systems Biology. Algorithms we work on for BioMedical Image Analysis involve deep learning from label-efficient discriminative models to generative AI for deep-learning powered microscopy. The latter notably includes tasks like deep learning-based super-resolution microscopy, denoising and phase retrieval. To address open problems in Systems Biology we involve machine learning in omics. To tie the two fields together we develop simulations for spatiotemporal understanding of virus infection spread.

  • Deep Learning for high-content screening microscopy

    The discovery of novel drugs and the unravelling of complex pathways in Molecular Biology can be addressed by Big Data obtained in a large image-based screening fashion. Fully automated (“robotised”) high-content microscopy allows for the simultaneous acquisition of multiple parameters from biological samples, such as cells or tissues, at a high-throughput scale. Big Data obtained from high-content microscopy allows for the extraction of detailed quantitative information on various cellular features, including morphology, protein expression, and subcellular localization, facilitating comprehensive analysis of cellular phenotypes. These phenotypes can be quantified using discriminative Deep Learning models allowing for accurate semantic and instance segmentation. The unintended result of automation is the introduction of a higher-than-usual number of sample preparation artefacts. We explore how such artefacts can be addressed by generative AI, for example using generative inpainting.

  • Deep Learning for label-free microscopy on research and clinical data

    Combined with Deep Learning, microscopy can serve as a powerful tool to support clinical decisions. One such example is using brightfield light microscopy to facilitate diagnostics of chronic urinary tract infections (UTIs). UTIs represent the predominant bacterial infection in humans and pose risks of recurrence or progression to severe infections like sepsis. Women face elevated susceptibility to UTIs and are prone to developing complex infections. Consequently, UTIs impose significant health and economic burdens, evident in escalating antibiotic prescriptions and hospitalizations associated with urinary infections. Conventional diagnostics like urine dipstick tests often fail in chronic cases. Discriminative Deep Learning objects can facilitate the segmentation of individual objects in patients' urine, enabling mophocytometric analysis. This, in turn, allows clinicians to search for novel biomarkers of chronic UTIs.

  • Interpretable and Label-efficient Deep Learning for Microscopy

    Using Deep Learning models to analyse microscopy data comes at a cost. Firstly, it requires the annotation of large datasets which can be cumbersome for tasks like object detection and segmentation. Secondly, once trained, deep learning models are difficult to interpret. To address the former we proposed leveraging weak labeling techniques. In this approach image segmentation labels can be generated algorithmically. The benefit of such an approach is that it allows for the generation of approximate labels for a large amount of data automatically. A model pre-trained on such approximate labels can be further fine-tuned on very precise labels, lowering the burden of the annotation efforts. To address the lack of interpretability of deep learning models, we have recently proposed a tandem segmentation classification connected through classifier saliency. In such a setting segmentation models can help make classification more interpretable by computing the overlap between them.

  • Machine Learning for Systems Biology of host-pathogen interactions

    Systems biology is an interdisciplinary approach that employs computational models to study complex biological systems. It integrates data from multiple levels of organization to gain a comprehensive understanding of biological phenomena. Machine learning can help unravel complex interactions in Systems Biology. For example, host-pathogen interactions refer to the dynamic relationships between a host organism and a pathogenic microorganism. They encompass the complex interplay of molecular, cellular, and immunological processes that determine the outcome of infection. Leveraging the described interactomes, as well as Representation Learning one can encode information from unstructured data like images and sequences to help uncover mechanisms and find novel targets to fight infections.


Has no events.

Related Publications