NIAID contract progress October 6 - November 17, 2022

Tom Goddard
November 17, 2022
Group meeting

• NIAID contract progress
  • Eric - NIH 3D pipeline
  • Greg - small molecule CIF
  • Zach - DICOM
  • Elaine - documentation for contract features
  • Tom - machine learning structure prediction for cryoEM
• CZI Imaging Institute electron tomography workshop
• ESMFold protein structure prediction from Meta
• AlphaFold prediction service for longer sequences

CZI Imaging Institute electron tomography workshop

• Workshop to discuss what the new CZI Imaging Institute should do.
• About 50 participants, held at Mountain View 5-star hotel.
• Organized by Dave Agard, Bridget Carragher and Clint Potter, directors of the institute.
• Agenda and participants
• Lots of interesting participants, got to talk with 20 of them.
  • Utz Ermel - ChimeraX ArtiaX developer, Goethe Univ Frankfurt
  • Ellen Zhong - cryoDRGN developer, Princeton
  • Alister Burt - Napari plugin developer, MRC-LMB
  • Nick Sofroniew - Napari project leader, CZI
  • Michael Schmid - ChimeraX VR expert with Wah Chiu, Stanford
  • Wah Chiu - director of SLAC cryoEM
  • Many others...
• Lots of machine learning methods presented or discussed: cryoDRGN, cell segmentation, IsoNet missing wedge filling.
• My talk Visualizing Tomograms focused on need for database of cell segmentations.
• Napari
  • Napari has about 250 plugins for 2D and 3D microscopy image analysis.
  • CZI gave $2 million in grants to develop plugins.
  • Two Napari talks made strong pitch to develop future cryoET visualization using Napari plugins.
  • Napari distributed as PyPi and Conda packages for developers and desktop app for users.

ESMFold protein structure prediction from Meta

• Meta released 600 million protein structure predictions, called ESM Metagenomic Atlas, on November 1, 2022.
• ESM stands for "Evolutionary Scale Modeling".
• Uses machine learning "language model" to predict structures 10 times faster than Alphafold, but less accurate.
• Does not use deep multiple sequence alignment as input, just single sequence.
• Added ChimeraX command "esmfold predict" to predict structure from sequence using Meta's server. Example.
• Prediction server limits sequence length to 400 amino acids.
• Prediction made in under 30 seconds.
• Working on ChimeraX web service for Atlas sequence search (fast kmer search and blast).

AlphaFold prediction service for longer sequences

• Can we use new lab computer with 4 Nvidia A40 GPUs for ChimeraX AlphaFold prediction?
• Main advantage over Google Colab is ability to predict sequences longer than 1000, up to ~4000, using more GPU memory than available on Colab.
• Problem is long sequences are slow to predict, e.g. 28 hours for 3200 sequence length on Wynton.
• With only 4 GPUs few long sequence jobs (5-10) could run per day.
• Instead I think we should allow ChimeraX predictions to run on paid cloud GPU services: AWS, Lambda Cloud, Microsoft Azure, Google Cloud, ... many others.
• Researchers would have ChimeraX use their GPU virtual machine. Cost ~$2-3 / hour.