Attendees

Selina, Zach, Scooter, Greg, Eric, Elaine, Tom Ferrin

June 6, 2024

Agenda

• journal club (Elaine): ​EncoderMap embedding residue interaction networks from trajectories

Discussion Notes

• Wynton upgrades
  • UCSD got money for GPUs but other UCs can't use them
    • they're hosted at SDSC which has painful security requirements
  • Wynton is getting two new GPU nodes that the RBVI will get access to
    • 8 x L40 48GB GPUs to be added to preexisting A40s
    • Officially NRNB nodes but RBVI is a member of NRNB
    • UCSD and Toronto people may also use them
  • Plato is also getting a memory upgrade (2 nodes go up to 1TB)
• 1.8 release
  • Nearly ready, Eric needs to track down tricky GC bug
• Tom gave LookSee demo with his wife, was successful
  • New intern, in the middle of finals, may be here next Thursday
• Elaine presentation
  • Embedding Residue Interaction Networks
    • No training required, uses autoencoder
    • Handles many time points efficiently
  • Motivation
    • Hard to get meaning from MD trajectories, many variables
    • Typically need expert knowledge to know which variables to pay attention to
    • RINs capture structural changes on many scales
    • Prior analyses used averages or only a few snapshots, lost temporal information
  • Implementation
    • Edge drawn b/t non-sequence-adjacent residues with any atoms within 6 angstroms
    • residue closeness centrality --> N-dimensional feature vector for each time point
  • Small Protein Example: Trp-Cage (20 residues)
    • Super long simulation (208 microseconds, >1M frames)
  • Supp data: Trp-Cage PCA and UMAP
    • Author's note: compared to EncoderMap, PCA more global, UMAP more local
  • Multidomain Example: FAT10 (165 residues)
    • FAT10 has two ubiquitin-like domains wiht a flexible linker and tails
    • 50 simulations, 50ns
    • Authors hypothesize that different closed conformations offer different interaction sites for FAT10's many binding partners
    • Map resolves functionally different conformations of a complex system over time, yet retains an interpretable global organization
  • Alternative Featurizations
    • Adjacency matrix colored by density or R_g
    • Degree centrality colored by density or R_g
    • Backbone dihedral angles colored by density or R_g
  • Discussion
    • Protein graph can be defined at different resolutions/with different criteria
    • RINs cannot distinguish fine local changes, best for systems that undergo strong changes

Action Items