Lectures

To kick things off, each participant is expected to provide a 2-minute presentation of their research interests on Sunday evening, starting at 8 p.m. (right after dinner).

The morning sessions (Monday to Friday, starting at 9 a.m.) will be devoted to the following lectures:

Monday: B. Keller and C. Clementi, Learning models of complex dynamics from simulation data

Tuesday: T. Lelièvre and G. Stoltz, Constructing collective variables using Machine Learning and free energy biased simulations

Wednesday: S. Ovchinnikov, Unified framework for understanding and evaluating generative sequence models

Thursday: M. Weigt, Wandering through sequence space : data-driven landscapes and protein evolution

Friday: S. Grudinin and E. Laine, Machine learning in the post CASP14 era : from protein structure to protein interactions

 


Monday: Learning models of complex dynamics from simulation data

 

Lecturers:

Bettina Keller and  Cecilia Clementi (Freie Universität Berlin)

Topics:

  • Introduction: protein dynamics in complex energy landscapes and kinetic models
  • Dimensionality reduction using PCA and TICA
  • Density-based clustering
  • Variational and core-set Markov models
  • Learning models of complex dynamics from simulation data

top

Tuesday: Constructing collective variables using Machine Learning and free energy biased simulations

Lecturers:

Tony Lelièvre and Gabriel Stoltz (Ecole des Ponts and Inria Paris)

Topics:

Collective variables are needed in many atomistic simulations to compute free energies, bias the dynamics, or construct effective dynamics on some coarse grained degrees of freedom of the system. These variables are often chosen using chemical intuition or a prior knowledge on the system. Tools from machine learning offer the promise of automating and abstracting these ad hoc choices. Our presentation is constructed around the following points:

  • what is a "good"collective variable?
    - free energies associated with collective variables
    - collective variable as numerical tools
  • constructing collective variables beyond chemical intuition
    - a review of various machine learning approaches
    - a focus on methods based on autoencoders

top

Wednesday: Unified framework for understanding and evaluating generative sequence models

Lecturer:

Sergey Ovchinnikov (Harvard University)

Topics:

  • Introduction to protein sequence models (PSSM, MRF/Potts)
  • Relationship to Neural Networks (AE, VAE)
  • Relationship to Language Models (Transformers, BERT, MSA_transformer)
  • Using sequence models for protein structure prediction (Alphafold, TrRosetta)

top

Thursday: Wandering through sequence space : data-driven landscapes and protein evolution

 

Lecturer:

Martin Weigt (Sorbonne Université)

Topics:

With the rapid increase of biological sequence databases, data-driven modeling approaches gain broad interest in computational biology. Here we will discuss recent efforts in statistical and machine learning, which aim at extracting biological information from sequence data. We will in particular concentrate on mostly unsupervised methods (i.e. using abundant unlabeled sequence data)

  • generative modeling : use models to reproduce the statistics of natural sequences, with applications to protein design;
  • predictive modeling : using unsupervised models to predict biological function, like effects of mutations or protein-protein interactions
  • interpretable modeling : design minimal models, whose structure and parameters are interpretable in terms of biomolecular structure and function
  • evolutionary modeling : use models to understand how proteins explore sequence space by balancing mutation and selection

 Friday: Machine learning in the post CASP14 era : from protein structure to protein interactions



Lecturers:

Sergei Grudinin (LJK-CNRS, Grenoble) and Elodie Laine (Sorbonne Université)

 Topics:

  • Representations of proteins and other molecules
  • Geometric learning (graphs, point clouds, Voronoi tessellations and 3D densities)
  • The puzzles of 3D learning : equivariance and space symmetry
  • Towards end-to-end learning : handcrafted versus learned features
  • From protein structure to protein communities, sociability and interactions

top

Online user: 2 Privacy
Loading...