OpenFold - Lessons and Insights From Rebuilding and Retraining AlphaFold2

Explore the groundbreaking OpenFold project in this 48-minute lecture by Mohammed AlQuraishi from Harvard Medical School. Delve into the lessons learned and insights gained from rebuilding and retraining AlphaFold2, a revolutionary tool in structural biology. Discover how OpenFold addresses limitations in the original implementation, including the lack of training code and data for new tasks, optimization for commercial hardware, and understanding of training data influence on accuracy. Gain valuable knowledge about the relationships between data size, diversity, and prediction accuracy, as well as insights into the protein folding learning process. Examine topics such as complex prediction, modularity, outliers, inverse characteristics, convergence characteristics, fine-tuning, and multiscale learning. Understand how the model generalizes across protein fault families and the interplay between local and global aspects of protein structure prediction.

Intro
The basic problem
A larger set of problems
Reinventing the space
Summary
Calibration
Questions
Motivations
Complex prediction
Applications
Protein Complex Prediction
Modularity
Outliers
Inverse characteristics
Decals
Convergence characteristics
Fine tuning
Selfassessment
Learning secondary structure elements
Multiscale learning
Learning spatial dimensions
Learning PCA projections
Learning PCA projections independently
How well does a model generalize
Protein fault families
Local vs global aspects