AI in Structural Bioinformatics

This talk covers AI methods in structural bioinformatics, with a focus
on modeling protein flexibility and disorder. It introduces DreamFold,
an AI "world model" that learns folding pathways in latent space,
replacing a slower classical sampling approach (discard-and-restart).
It presents machine-learning improvements to Kohn-Sham
Hamiltonian estimation for faster DFT calculations on larger
molecules. It also shows that ensembles of sequence-based deep
learning models outperform individual predictors and 3D docking for
drug screening. Finally, it addresses protein aggregation in disease
(e.g., AD) via liquid-liquid phase separation (LLPS), using LLM
embeddings and graph neural networks to predict LLPS-prone regions,
intrinsically disordered regions (IDRs), and the effects of specific
mutations.