1. D7.4 A REFMAC server for EM and NMR

2. D7.7: Quality analysis workflow for predicted complexes
Joint project:
EMBL-EBI
Marc F. Lensink (University of Science and Technology of Lille)

3. Background & rationale
Protein–protein interactions and protein assemblies play a crucial role in all cellular processes.
A small fraction of protein complexes have been solved experimentally. Computational procedures to generate models of macromolecular assemblies is important to supplement experimental methods.
Validation of experimentally determined structures: coordinated by the wwPDB structure validation task forces (X-ray, NMR, EM, hybrid/integrative methods). These are available to worldwide user.
Validation of predicted complexes is established by the CAPRI (Critical Assessment of Predicted Interactions) community, however it is not available to users.
This deliverable (D7.7) aims to make this analysis protocol available to users worldwide.

4. CAPRI + evaluation criteria
The Critical Assessment of Predicted Interactions (CAPRI) is a community-wide initiative to organize worldwide experiment for macromolecular complex prediction. Established in 2001.
>130 complexes have been predicted by groups worldwide
Since its inception, the CAPRI committee has organized six evaluation meetings. During these evaluation meetings, discussions within the community have led to established standards for the parameters and criteria used to evaluate the quality of the predicted complexes
Assess geometric (L-rms, I-rms) and biological (f(nat)) properties of the models

5. Web server Run workflow on EBI farm
Input 1: Structures of Unbound proteins
Input 2: Structure of Complex
Input 3: Predicted
models
Run workflow on EBI farm

6. Landing page + expert interface

7. Result User provide e-mail – unique link to result
Currently, result shows a table showing classification of each model
Will allow download of superposition files, etc

8. D7.8 Report on prototypes using Big Data approaches
Selected projects:
Implementation of Convolutional Neural Network for structural biology maps
Clustering PDB entries
Extracting mentions of residues from literature

9. Machine learning for cryoEM
First prototype for "Big Data" technologies in structural biology
Other examples:
Steve Ludtke CNNs for cryoET
SuRVoS for segmentation
Several attempts at particle picking
Use case:
Analyse 3D maps to learn and recognise features at different scales:
e.g. missing components, domains, side chains
Protein vs nucleic acid vs solvent
EM: effect of map sharpening
xtal: effect of phase quality

10. CNN for maps Network Data Classify as protein or noise
EMD-2984 betagal 2.2Å
48*48 pixel 2D slices
Fitted model 5a1a to annotate as protein or not
Compare deposited map with blurred map
33,148 and 18,750 slices respectively
Network
Classify as protein or noise
Implemented with Keras
Tried standard VGG16 model, but reverted to 5-layer model
No. images
Protein
Non-protein
Loss
Accuracy
Model trained and tested on images from blurred map
33148
15696
17452
0.099
0.978
Model trained on blurred map, tested on images from unblurred map
18750
4736
14014
11.842
0.252
Model trained and tested on images from unblurred map
0.176
0.938
Model trained on unblurred map, tested on images from blurred map
1.631
0.587
deposited
blurred

11. Similarity and clustering in PDB
The problem
New protein/chain, find similar ones in PDB
Several methods to assess similarity, e.g.
GESAMT: E. Krissinel, Enhanced fold recognition using efficient short fragment clustering, J Mol Biochem., 2012, 1: 76-85
RCC: R.Corral-Corral et al., Machine Learnable Fold Space Representation Based on Residue Cluster Classes,  Comp. biol. and chem., 2015, 59: 1-7.
Too many (140k entries/500k chains) to compare with all
Representative clean set (65k chains)
Thorough statistics (space coverage, similarity measure correlation, pairwise similarity distribution, …)
Two proposed approaches
Dissimilarity search, dimensionality reduction and clustering

12. Similarity and clustering in PDB
“Dissimilarity search”
If the new structure t is dissimilar to some r, anything similar to r needn’t be considered
Precompute similarity of all entries with some fixed representative set N
Still feasible -- dne once, extensible with new entries
On query with t, choose random samples of N
Based on precomputed similarities, use them to identify candidates
Compute similarity of t with all candidates, choose the best one

13. Similarity and clustering in PDB
Dimensionality reduction and clustering
Compute residue cluster classes – 26 integer descriptors
Embed in real space, perform linear and non-linear dimensionality reductions
Apply multiple clustering techniques
Evaluate quality of resulting clusters
Framework to plug in different algorithms in all steps being developed

14. Similarity and clustering in PDB
Plans
Develop prototypes
Cross-validate results
Use independent similarity assessment techniques
Publish!
Integrate with PDB

15. Annotations for EuropePMC

16. Natural Language Processing
Annotations will also be imported to PDBe
Dashboard at
Uses spaCy
Software pyresid available via pip
IUCR considering use
Work done by Rob Firth STFC, Francisco Talo EBI
Rob will attend OpenMinTed conference
Challenges: sentence boundary recognition
Possible continuation by future grant application
Ingest from Proteopedia?

17. D7.9 Report on existing metadata standards, and proposals for new vocabularies
Provenance: D6.4 should add PROV-O to Virtual Folder
Projects: translate ARIA metadata to CERIF
Workflows: CSIC EOSC Pilot on cryoEM workflows
NUTS metadata in Repository, ? NMR-STAR
mmCIF…

18. UU Contribution to data standards
Utrecht member of the integrative modelling task force of PBD
Contributing to expending the mmCIF dictionary
Visit to Sali’s lab at USSF and RCSB in Rutgers in January
First HADDOCK integrative model deposited in PDB dev:

19. mmCIF DipCheck now accepts mmCIF input
ARP/WARP accepts mmCIF for ligands but not proteins (EMBL-Hamburg)
The PDB-REDO databank now stores mmCIF files
The PDB_REDO service reads and writes mmCIF
Version 8 of PDB-REDO will use mmCIF internally (NKI).
3DBionotes reads mmCIF

20. D7.5 A HADDOCK server for EM