Summary
Protein folding neural networks can be combined with VAEs for conformational modeling (1). Authors used RosettaFold in conjunction with the VAE to sample cryptic pockets in K-Ras.
Details
- Used RMSD loss (RMSD is a poor training objective for structure prediction)
- Resulting models were more amenable to ligand docking (High-accuracy computational models might not always be effective for ligand docking)
1.
Mansoor S, Baek M, Park H, Lee GR, Baker D. Protein Ensemble Generation Through Variational Autoencoder Latent Space Sampling. Journal of Chemical Theory and Computation. 2024;20(7):2689–95. Available from: https://doi.org/10.1021/acs.jctc.3c01057