Lectures and seminars Biologically Informed Neural Networks
We are happy to invite you to two seminars on the theme of Biologically Informed Neural Networks, Friday March 13 at 10am in room Ragnar Granit. The talks will highlight how integrating prior biological knowledge with modern optimization and neural network approaches enables more interpretable, predictive, and scalable models of disease and omics data. Join us for an engaging discussion!
Talk 1, 10am: Erik Hartman
Lund University
Title: "Modelling biological systems using machine learning"
Biological modelling is shifting from simple, expert-designed equations to data-driven machine learning models capable of capturing complex biological processes. In this talk, I will present biologically informed neural networks (BINNs), which integrate prior biological knowledge into neural networks to improve their interpretability in classifying disease, and show how these ideas extend to modelling proteolysis using degradation graphs. Together, these approaches illustrate how combining biological structure with modern machine learning can advance our understanding of complex systems and disease.
Talk 2, 11am: Pablo Rodriguez Mier
Heidelberg University
Title: “Learning, Predicting, and Interpreting Omics Data with Biologically Informed Models”
High-throughput omics assays capture molecular states at scale, yet building robust, interpretable, and actionable models from high-dimensional measurements remains difficult. Key problems include limited numbers of observed conditions and perturbations, technical confounding and batch effects, difficulty integrating heterogeneous modalities, and the challenge of separating causal effects of interventions from correlations driven by shared regulation or confounding. Incorporating prior biological knowledge can overcome these limitations by constraining the hypothesis space, providing contextual structure, and encoding mechanistic assumptions that are not directly identifiable from data alone.
This talk brings together three lines of work on using optimization to connect biological knowledge with omics data and predictive modeling. First, we present optimization-based methods for inferring interpretable, context-specific biological networks from omics measurements by solving constrained problems over prior-knowledge graphs using CORNETO and NetworkCommons. Second, we show how the same optimization viewpoint can be used to discover biologically inspired neural architectures in two stages: interaction graphs define structured hypothesis spaces, and search/selection over these spaces yields neural architectures whose connectivity mirrors biological networks. Third, we describe how to build end-to-end differentiable, biologically informed models by incorporating biological knowledge through differentiable optimization layers that enforce constraints in a principled and general way, and how these models can be scaled to large networks by learning fast approximations to repeated constrained solves across many contexts. We close with applications in perturbation biology, discussing current challenges in predicting intervention responses and how biological knowledge can guide the development of the next generation of models.
Host
Avlant Nilsson, Department of Cell and Molecular Biology, Karolinska Institutet