Trainer: Gabor Erdos
Machine learning methods have become exceptionally powerful tools for exploring and understanding biological systems in ways that were not possible before. In this tutorial, we will take a deep and intuitive dive into how classical, “old-school” machine learning techniques operate—including linear and logistic regression, decision trees, and random forests—before gradually moving toward neural networks and modern deep learning approaches.

Rather than treating these models as black boxes, we will build several machine learning models step by step, essentially “by hand,” to develop a clear understanding of their underlying principles. Along the way, we will discuss where each method excels, the types of problems it is best suited for, and the limitations and trade-offs that come with it. The tutorial is designed to be fully accessible and does not require any prior knowledge of machine learning or programming; for participants without coding experience, all necessary code will be provided in advance so the focus can remain on concepts and intuition.

Trainer: Nicola Bordin
The emergence of accurate protein structure models, large-scale sequence databases, and protein language models is dramatically accelerating research in structural bioinformatics. This deluge of data has given rise to two main areas of research: classification and generation of novel, synthetic data.
We will cover, and run a hands-on practical, on how to classify globular domains in the AlphaFold Database, visualise outputs from various predictors, and assess multiple metrics for confident domain-boundary assignments. This protocol was used to identify and classify domains that were subsequently used to train a protein family language model, ProFam.
We will then use ProFam to generate synthetic sequences conditioned on a protein of your choice, producing sequences with similar characteristics that do not exist in nature. Finally, we will model these sequences and analyse their relationships to existing proteins using tools such as FoldTree and Foldseek.

Trainer: Giulio Tesei
Molecular dynamics simulations of coarse-grained models offer an efficient and accurate approach for generating conformational ensembles of non-globular proteins (NGPs). This tutorial introduces CALVADOS, a physics-based, residue-level model with amino acid–specific parameters optimized using experimental SAXS and NMR data. CALVADOS simulations enable the generation of conformational ensembles of NGPs at the proteome scale, and such datasets have been used to train machine-learning models that predict conformational properties from sequence.
We will run a hands-on practical using Jupyter Notebooks on Google Colab, in which participants will simulate an NGP with CALVADOS, and integrate the resulting ensemble with experimental SAXS data through Bayesian/Maximum-Entropy reweighting. Participants will then explore how proteome-scale simulation data can be used to train a machine-learning model that predicts the compaction of intrinsically disordered proteins based on selected sequence features.

Trainer: Giacomo Janson
Deep generative models are emerging as a tool to characterize the structural ensembles of non-globular proteins (NGPs). They offer a powerful and computationally efficient complement to traditional biophysical and simulation-based approaches. In this training session, we will introduce the concepts behind deep generative models of biomolecular structures, with an emphasis on intrinsically disordered proteins (IDPs) and dynamical proteins. Participants will learn what it means for a model to be “generative,” how existing techniques (e.g., diffusion models) capture structural variability, and how they are used to model the conformational landscapes of NGPs.
The hands-on component will guide participants through practical applications of both first-generation and state-of-the-art models for ensemble generation. All exercises will be carried out in Jupyter notebooks, allowing participants to run the models, visualize outputs, and experiment with the data. In the first part, attendees will work with idpGAN, a first-generation method for coarse-grained IDPs. We will examine how to generate structural ensembles, analyze their properties and validate them against classical simulation-based approaches.
The second hands-on part will focus on the current-generation models for atomistic ensembles, highlighting their improved resolution and environmental conditioning. Participants will use the aSAM model and compare its ensembles with the ones from other recent models (e.g.: BioEmu and AlphaFold3), learning how to interpret outputs, evaluate ensemble quality, and identify common pitfalls in methods for ML-driven structural modeling of protein ensembles.
By the end of the session, participants will have both a conceptual understanding of generative modeling for biomolecular structures and practical experience applying this approach to real protein systems. The training aims to provide participants with the skills needed to critically deploy generative models in their own studies of protein disorder and dynamics.