Over 80% of pathogenic proteins are considered 'undruggable' because traditional inhibition methods are ineffective for them.
This biotechnology shifts the paradigm from inhibiting symptoms to eliminating the cause of disease by degrading pathogenic proteins.
Graphs provide a native, non-Euclidean way to represent complex molecular structures and interactions, which is essential for effective machine learning models.
The process involves predicting protein-ligand interactions, generating molecular conformations, and using Bayesian optimization to find optimal candidates.
Bayesian optimization and active learning are used to efficiently search the high-dimensional space of molecular rotations and translations to find the best interaction.
Generative models create entirely new linker molecules de novo, optimizing for properties like low toxicity and minimal energy during generation.
A custom cloud architecture using AWS spot instances and persistent storage is necessary to handle the immense computational cost of geometric deep learning.
Success requires a combination of high-quality data, advanced algorithms like geometric deep learning, cloud computing power, and an interdisciplinary team of domain experts.
The discussion covers the challenges of communication in interdisciplinary teams, the rationale for a multi-cloud strategy, and specific technical questions about the models.
46:51
59:39Julian Joseph
44:11Alexandra Waldherr
59:38Erik Bamberg
50:37Svetlin Penkov
56:55Ekaterina Sirazitdinova
26:55Jan Zawadzki
32:54Adrian Spataru & Bohdan Andrusyak
Jobs that call for the skills explored in this talk.
Proclinical Limited
Deutsches Medizinrechenzentrum GmbH
Mindpeak
CERN European Organization for Nuclear Research
CERN European Organization for Nuclear Research
Avelios Medical