The main focus of the workshop will be a hands-on practical project that allow participants to work in a team environment (5-6 participants/team) on a particular problem which can be tackled within a 2-day time frame, using ML methods applied to biomedical data.
Participants will be assigned to one of the following projects:
Project 1: Mathematical modeling of single-cell/multi-omics data of inflammatory diseases
Fortunately, the availability of high-throughput data sets, such as single-cell mRNA-sequencing (scRNA-seq), is increasing. Unfortunately, accessible approaches to mathematically model these data are scarce, limiting scientific progress. In this course you'll learn how to complement powerful machine learning (ML) pipelines for scRNA-seq analysis (such as Seurat, scanpy) with mathematical modelling to understand the pathomechanisms of obesity and inflammatory bowel disease. The concepts taught here can be transferred beyond scRNA-seq data from inflammatory diseases, thereby distinguishing it from current ML approaches.
Project 2: Drug response prediction based on gene expression data
Personalized medicine has emerged as a promising approach to revolutionize healthcare by tailoring treatment to individual patients. The rapid development of molecular profiling technologies has increased interest in machine learning (ML) techniques that could be used to predict drug responses in patients in a personalized manner. In this project, participants will learn to implement and evaluate an ML pipeline for drug response prediction. Specifically, we will cover three key steps: (i) feature reduction, (ii) training ML models, and (iii) testing and evaluating these models. The dataset used in this project consists of cell line gene expression profiles and their corresponding drug response data.
Project 3: Clinical diagnostics from Microscopy image analysis
In this project we will start from microscopy images of human tissue biopsies and use deep learning-based image analysis to extract features. With these features we will classify the patients into different groups. We will use data from different sources (clinics) to highlight and ideally counter the lack of robustness to data variations commonly observed in deep neural networks.
Project 4: Using Large Language Models to Build Neural Networks and Deep Learning Applications
This hands-on workshop will teach attendees how to leverage large language models and AI-assisted programming to build neural networks and deep learning applications. Participants will apply practical techniques for utilizing pre-trained large language models (LLM) like ChatGPT or Claude to kickstart model development, customizing and fine-tuning language models, and integrating them with other deep learning components. The workshop will also cover automating the coding process through the use LLMs and deploying AI-powered applications in production
Project 5: Representation learning for multimodal images
Images acquired by different sensors capture complimentary information and are often characterized by very different appearances. Generally, these so-called multimodal images have to be acquired in different machines and image alignment is required prior to image fusion. This alignment can be very challenging, in particular in the case of different types of microscopy images. In this workshop, attendees will learn to draw upon deep learning and representation learning techniques originating from self-supervised learning such as contrastive learning which allows to learn shared, abstract representations between such multimodal images, that further down the line can be used for image alignment or image retrieval across modalities. |
