Registration
You are invited to join this webinar, part of the ongoing activities in the ELIXIR 3D-BioInfo Community. This will include two presentations from leading researchers in the field of Intrinsically Disordered Proteins.
Zsuzsanna Dosztanyi
Integrating Biophysics and Deep Learning for Interpretable IDP Prediction
Intrinsically disordered proteins (IDPs) lack stable three-dimensional structure, yet play essential roles in cellular regulation, signaling, and disease. Their conformational heterogeneity and context-dependent behavior pose significant challenges for traditional experimental and structure-based computational approaches. While recent advances in deep learning have revolutionized protein structure prediction, standard energy functions still fail to capture the dynamic and thermodynamic properties that govern IDP function.
To overcome these limitations, we developed a hybrid framework that integrates simple biophysical models with deep learning. Our method builds on IUPred, which estimates residue-level interaction energies directly from amino acid sequence using pairwise statistics derived from known structures. We extend this framework by incorporating deep learning and expanding the training dataset with AlphaFold models and large-scale variant effect scores, enabling improved identification of disordered regions and their functional sites.As a next step, we focus on refining the energy model itself. By reframing the original approach to derive pairwise energies from known structures, we developed a graph neural network that learns effective interaction energies directly from protein structures. This approach not only discriminates native structures from decoy conformations but also provides zero-shot predictions of ensemble-level properties , such as the radius of gyration, disordered proteins. Importantly, by grounding the model in biophysical principles, we enhance interpretability and open the way toward mechanistic insights into IDP behaviour.
Michele Vendruscolo
Expediting Drug Discovery for Undruggable Targets Using AI
Protein-ligand interactions play central roles in biological processes and are of key importance in drug design. Deep learning approaches hold the promise of becoming cost-effective alternatives to high-throughput experimental methods for ligand identification. This is particularly the case for disordered proteins, where there are few experimental options for accurate binding measurements that can be scaled up to cover large chemical libraries. To predict the binding affinity between disordered proteins and small molecules, I will describe Ligand-Transformer, a deep learning framework based on the AlphaFold intermediate structural representations. To show the applicability of this approach, we explored the binding of small molecules to the Alzheimer’s Aβ peptide, identifying compounds that delayed its aggregation. Overall, the results that we obtained illustrate the potential of Ligand-Transformer in accurately predicting the interactions of small molecules with disordered proteins, thus uncovering molecular mechanisms and facilitating the initial steps in drug discovery for these otherwise largely undruggable targets.
This is part of a continuing series building on the success of previous years, recordings of which are available here.
