Enabling Scalable Data Analysis for Large Computa9onal Structural Biology Datasets on Distributed Memory Systems

Size: px

Start display at page:

Download "Enabling Scalable Data Analysis for Large Computa9onal Structural Biology Datasets on Distributed Memory Systems"

Peter Skinner
5 years ago
Views:

1 Enabling Scalable Data Analysis for Large Computa9onal Structural Biology Datasets on Distributed Memory Systems Michela Taufer Global Compu9ng Laboratory Computer and Informa9on Sciences University of Delaware

In- situ Analysis The perfect in- situ analysis algorithm: Avoids the need for moving data Uses a limited amount of memory Executes sufficiently fast Dataset

2 In- situ Analysis The perfect in- situ analysis algorithm: Avoids the need for moving data Uses a limited amount of memory Executes sufficiently fast Dataset NleNle - PDB ID 2F4K from Vijay Pande group Can we perform in- situ analysis on trajectories generated in protein folding, predic9on, or refinement simula9ons? 1

3 Start from unfolded conforma9ons of a protein with correct chemical bonds but random torsion angles Search for a conforma9on close to the observed na9ve (folded) conforma9on Protein Folding Process Reynaud, E. (21) Protein Misfolding and Degenera<ve Diseases. Nature Educa<on 3(9):28 2

4 Start from unfolded conforma9ons of a protein with correct chemical bonds but random torsion angles Search for a conforma9on close to the observed na9ve (folded) conforma9on Protein Folding Process Reynaud, E. (21) Protein Misfolding and Degenera<ve Diseases. Nature Educa<on 3(9):28 3

Scien9fic Problem Cluster folding trajectories into recurrent paverns based on geometrical varia9ons in 9me of the folding protein frames Node 1 Node 2 Trajectory segment 1

5 Scien9fic Problem Cluster folding trajectories into recurrent paverns based on geometrical varia9ons in 9me of the folding protein frames Node 1 Node 2 Trajectory segment 1 Trajectory segment RMSD RMSD RMSD RMSD frame / conformations Frame (snapshot) 2 frame / conformations Frame (snapshot) 4

6 Scien9fic Problem Cluster folding trajectories into recurrent paverns based on geometrical varia9ons in 9me of the folding protein conforma9ons Intra- trajectory analysis - > iden9fy meta- stable and transi9on stages within trajectory Meta- stable 1 8 RMSD RMSD 6 4 RMSD RMSD Frame (snapshot) 2 Transi9on Frame (snapshot) 5

7 Limits of Current Prac9ce Centralized clustering analysis [Phillips et al. 213] Wait for a job to end before to move its data (trajectory segment) to a centralize server Does not scale when the dataset is large May end up was9ng compu9ng resources e.g., by trying to further fold proteins that are already in a stable condi9on 6

8 From Protein Conforma9ons to Metadata From a single protein conforma9on to a 3D point capturing the atom distances within the frame distance matrix (DM) atoms three leading eigenvalues atoms (x,y,z) Conforma9on (Data from F@H) mul9- dimensional scaling metadata B. Zhang, T. Estrada, P. Cicoc, and M. Taufer. Enabling In- situ Data Analysis for Large Protein- folding Trajectory Datasets. In IPDPS, 214 7

9 From Metadata to Scien9fic Knowledge Given a set of confirma9ons and their 3D points in a segment of the trajectory: 5 3 largest eigenvalues as (x,y,z)

10 From Metadata to Scien9fic Knowledge Given a set of confirma9ons and their 3D points in a segment of the trajectory Par99on the 3D points into 2 clusters using fuzzy c- means (with c = 2)

11 From Metadata to Scien9fic Knowledge Given a set of confirma9ons and their 3D points in a segment of the trajectory: Par99on the 3D points into 2 clusters using fuzzy c- means (with c = 2) Test the equality of the 2 clusters using Welch s t- test with p- value <

12 From Metadata to Scien9fic Knowledge Given a set of confirma9ons and their 3D points in a segment of the trajectory: Par99on the 3D points into 2 clusters using fuzzy c- means (with c = 2) Test the equality of the 2 clusters using Welch s t- test with p- value <.1 Itera9ve par99on on the cluster with more points

13 From Metadata to Scien9fic Knowledge Given a set of confirma9ons and their 3D points in a segment of the trajectory: Par99on the 3D points into 2 clusters using fuzzy c- means (with c = 2) Test the equality of the 2 clusters using Welch s t- test with p- value <.1 Itera9ve par99on on the cluster with more points Finish when the 2 clusters are equal Stage 1 Stage 2 Stage 3 Representative

14 From Metadata to Scien9fic Knowledge Intra- trajectory analysis of an ensemble of - conforma9ons containing one meta- stable stage followed by a transi9on stage and another meta- stable stage Our clustering iden9fies two stable stages and one transi9on stage 13

15 From Metadata to Scien9fic Knowledge Intra- trajectory analysis of an ensemble of conforma9ons containing one meta- stable stage followed by a transi9on stage and another meta- stable stage 5 3 RMSD Stage 1 Stage 2 Stage 3 Representative Our clustering iden9fies two stable stages and one transi9on stage Protein conformations RMSDs of protein conforma9ons iden9fy the same three stages 14

16 From Metadata to Scien9fic Knowledge Intra- trajectory analysis of an ensemble of - conforma9ons containing one meta- stable stage only Our clustering iden9fies one single stage 15

17 From Metadata to Scien9fic Knowledge Intra- trajectory analysis of an ensemble of - conforma9ons containing one meta- stable stage only Stage 1 Representative Our clustering iden9fies one single stage RMSD Protein conformations RMSDs of protein conforma9ons iden9fy one single stage 16

18 Time (sec) Performance Compare our approach with tradi9onal clustering method proposed by Phillips et al s B Folding trajectories of villin headpiece subdomain (HP- 35 NleNle) Parallel MATLAB on 256 Gordon compute cores Time(sec) Memory usage (MB) Data movement (KB) 1116s B 1 Memory per core (MB) MB 1MB Data moved per analysis (KB) KB 539MB - - Our method Phillips' method Our method Phillips' method Our method Phillips' method Our method performs orders of magnitude bever in terms of 9me, memory usage and data movement 17

19 Lessons Learned and Future Direc9ons The distributed analysis of structural biology data is feasible and scalable Our approach transforms data proper9es into metadata concurrently and extracts scien9fic insights from metadata with small data movement Our approach makes the in- situ analysis feasible by: By avoiding the need for moving data, using a limited amount of memory, and execu9ng sufficiently fast Can we extend our approach to larger scales, i.e., larger datasets and more diverse datasets? 18

20 Global Compu9ng Lab: Boyu Zhang Travis Johnston Acknowledgments Collaborators: Trilce Estrada (UNM) Pietro Cicoc (SDSC) Roger Armen (TJU) Special thanks to: D.E. Shaw group Vijay Pande group (Stanford) volunteers volunteer Sponsors: 19

Enabling in-situ data analysis for large protein-folding trajectory datasets

Enabling in-situ data analysis for large protein-folding trajectory datasets Boyu Zhang, Trilce Estrada, Pietro Cicotti, Michela Taufer University of Delaware {bzhang, taufer}@udel.edu University of New