Yazan Haddad, CTO at MendelFOLD, and Assistant Professor at Mendel University in Brno presents a knowledge-driven approach to protein 3D structure prediction as an alternative to current AI-based methods like AlphaFold.
The talk highlights challenges such as high computational cost, limited interpretability, and difficulties with certain protein types. The proposed approach leverages genetic and biophysical principles, including amino acid residue pairing and constraint-based simulations, to enable more interpretable and flexible structure prediction.
As deep learning approaches such as AlphaFold continue to redefine structural biology, their limitations highlight persistent challenges, including high computational demands, dependence on homologous structural data, limited interpretability, difficulty with orphan, membrane, multistate and disordered proteins, and insufficient sensitivity to mutation-driven structural changes.
By contrast, we take a knowledge-driven approach to protein 3D structure and function prediction based on genetic and biophysical principles of amino acid residue pairing (the proteomic code), rotameric/geometric shapes, and relay information system (RIS) mechanisms which represent interpretable fundamentals for protein 3D structure and function prediction methodologies accessing constraint-driven simulation protocols. Further knowledge gained by our approach and the broad scope of protein families accessible to our approach should also benefit future AI strategies.
Predavanje bo v angleščini, in sicer na Fakulteti za računalištvo in in informatiko v Ljubljani.