Foldit and Games for Scientific Discovery

Jeff Flaen Center for Game Science Department of Computer Science University of Washington

hp://marketplace.veer.com/stock-photo/Teenage-Boy-Playing-With-Game-1783548 Human Power Computational Power Proteins: Foldit Proteins Proteins

MQIFVKTLTGKTILEVEPSDTIE…

Sequence

5 Proteins

MQIFVKTLTGKTILEVEPSDTIE…

Sequence 3D Structure

6 Proteins

MQIFVKTLTGKTILEVEPSDTIE…

MGKYDKQIDLSTVDLKKLRVKEL…

KPVSLSYRCPCRFFESHVARANV… … …

Sequence 3D Structure

7 Proteins

MQIFVKTLTGKTILEVEPSDTIE…

MGKYDKQIDLSTVDLKKLRVKEL…

KPVSLSYRCPCRFFESHVARANV… … …

Sequence 3D Structure EASY HARD 8 Proteins

MKTLYDLPIVLRNLPEDLVLEKP… ?

SMSDTAERVKKIVVEHLGVDADK… ?

GPLGSDQYIVVNGAPVIPSAKVP… ? … …

Sequence 3D Structure EASY HARD 9 Proteins

MKTLYDLPIVLRNLPEDLVLEKP… ?

SMSDTAERVKKIVVEHLGVDADK… ?

GPLGSDQYIVVNGAPVIPSAKVP… ? … …

Sequence 3D Structure EASY HARD 10 Proteins

MKTLYDLPIVLRNLPEDLVLEKP… ?

SMSDTAERVKKIVVEHLGVDADK… ?

GPLGSDQYIVVNGAPVIPSAKVP… ? … …

Sequence 3D Structure EASY HARD 11 Distributed compung

Rosea@home Folding@home (most powerful distributed compung cluster in the world!) Why a game?

Computaonally Difficult

13 Why a game?

Computaonally Difficult

Human Skills

14 Why a game?

Computaonally Difficult

Human Skills

Impact

15 Foldit

16 Foldit Architecture

Open problems

17 Foldit Architecture

Puzzles Open problems Database Database Web Web

Servers

18 Foldit Architecture

Puzzles

Open problems Puzzles Database Database Web Web

Servers Game clients

19 Foldit Architecture

Puzzles

Open problems Puzzles Database Database Web Web

Solutions

Servers Game clients

20 Foldit Architecture

Puzzles

Open problems Puzzles Database Database Web Web

Solutions

Aggregate solutions Servers Game clients Solution analysis

21 Foldit Architecture

Puzzles

Open problems Puzzles Database Database Web Web

Solutions

Aggregate solutions Servers Game clients Solution analysis

22 Foldit Architecture

Game Updates Puzzles

Open problems Puzzles Database Database Web Web

Solutions

Aggregate solutions Servers Game clients Solution analysis Foldit Architecture

Game Updates Puzzles

Open problems Puzzles Database Database Web Web

Solutions

Aggregate Player solutions Servers Data Game clients Solution analysis Foldit Scoring

25 Foldit Scoring

26 Visualizaons

27 Visualizaons

Game Visualizaon 28 Visualizaons

Clashes

29 Visualizaons

Clashes

Voids 30 Visualizaons

Clashes Hydrophobics

Voids 31 Visualizaons

Clashes Hydrophobics

Voids Hydrogen Bonds 32 Visualizaons

Clashes, hydrogen bonds, sidechains OFF Clashes, hydrogen bonds, sidechains ON

33 Visualizaons refinement

34 Visualizaons refinement

Older Version

35 Visualizaons refinement

Older Version Newer Version

36 Introductory Levels

37 What have we found so far? Very broad appeal Example demographic data Gender Occupation Country

Student Computer USA UK Male Female Labor Unemployed Germany France Technical Science Poland Sweden Busn./Finan./Legal Art Netherlands Australia Education Retired Russian Federation Portugal New Zealand Argentina Other Very broad appeal

Prior knowledge of biochemistry

None

High school / Basic

One undergraduate course

Majored in biology or similar

PhD in chemistry or organic chemistry PhD in biochemistry 30 years ago Professionally involved Protein Structure Predicon

MQIFVKTLTGKTILEVEPSDTIE… ?

41 Players can solve structures

• Fix incorrectly folded proteins • Move through worse scores to find beer ones • Outperform computaonal methods Players can solve structures

SCORE

TIME Players can solve structures

SCORE

TIME Players can solve structures

SCORE

TIME Players can solve structures

SCORE

TIME Players can solve structures

SCORE

TIME Players can solve structures

SCORE

TIME Comparison to Algorithm

Player soluons Top player path Algorithm soluons Beer energy

Closer to nave Comparison to Algorithm

Player soluons Top player path Algorithm soluons

Swapped Strands Comparison to Algorithm

Player soluons Top player path Algorithm soluons Experimental Structure Solved

• Mason-Pfizer Monkey Virus Retroviral Protease (MPMV PR) • Plays a role in AIDS in monkeys • Experimentalists worked on for ~15 years • Computaonal methods failed to solve • Gave to players for 3 weeks Experimental Structure Solved Starng structure Nave Quality

53 Time Algorithm Discovery

MQIFVKTLTGKTILEVEPSDTIE… ?

54 Players develop strategies Players develop strategies Top Cookbook Recipes Usage

Colors are different players Recipes Algorithm Comparison Player Algorithm (Blue Fuse v1.1)

Set FA_REP to 0.05

Shake

Set FA_REP to 1

Wiggle

Set FA_REP to 0.07

Shake

Set FA_REP to 1

Wiggle

Accept Best Energy

Set FA_REP to 0.3

Wiggle Adjust repulsive force

Set FA_REP to 1 Discrete opmizaon Wiggle Accept Best Energy Connuous opmizaon Algorithm Comparison Player Scienst Algorithm Algorithm (Blue Fuse v1.1) (Fast Relax)

Set FA_REP to 0.05 Set FA_REP to 0.02

Shake Repack/Minimize

Set FA_REP to 1 Set FA_REP to 0.25

Wiggle Repack/Minimize

x5-15 Set FA_REP to 0.07 Set FA_REP to 0.55

Shake Repack/Minimize

Set FA_REP to 1 Set FA_REP to 1

Wiggle Repack/Minimize

Accept Best Energy

Set FA_REP to 0.3 Accept Best Energy

Wiggle Adjust repulsive force

Set FA_REP to 1 Discrete opmizaon Wiggle Accept Best Energy Connuous opmizaon Algorithm Comparison Player Scienst Algorithm Algorithm (Blue Fuse v1.1) (Fast Relax) • Independent discovery Set FA_REP to 0.05 Set FA_REP to 0.02 of sciensts’ algorithmic Shake Repack/Minimize

Set FA_REP to 1 Set FA_REP to 0.25 techniques

Wiggle Repack/Minimize

x5-15 Set FA_REP to 0.07 Set FA_REP to 0.55

Shake Repack/Minimize

Set FA_REP to 1 Set FA_REP to 1

Wiggle Repack/Minimize

Accept Best Energy

Set FA_REP to 0.3 Accept Best Energy

Wiggle Adjust repulsive force

Set FA_REP to 1 Discrete opmizaon Wiggle Accept Best Energy Connuous opmizaon Algorithm Comparison Player Scienst Algorithm Algorithm (Blue Fuse v1.1) (Fast Relax) • Independent discovery Set FA_REP to 0.05 Set FA_REP to 0.02 of sciensts’ algorithmic Shake Repack/Minimize

Set FA_REP to 1 Set FA_REP to 0.25 techniques

Wiggle Repack/Minimize

x5-15 Set FA_REP to 0.07 Set FA_REP to 0.55 • Improved performance Shake Repack/Minimize compared to published Set FA_REP to 1 Set FA_REP to 1

Wiggle Repack/Minimize algorithms

Accept Best Energy

Set FA_REP to 0.3 Accept Best Energy

Wiggle Adjust repulsive force

Set FA_REP to 1 Discrete opmizaon Wiggle Accept Best Energy Connuous opmizaon Protein Design

MQIFVKTLTGKTILEVEPSDTIE… ?

62 Protein Design Experimental Validaon

64 Early Player Design Protein Design Protein Design

Starng scaffold

67 Protein Design

Starng scaffold Player/scienst design

68 What’s next? Experimental Density Integraon

70 Tool for Sciensts

71 More Intuive Interacons Educaon

73 Nanotechnology Nanotechnology Humans Computers Thank you!

jfl[email protected]

Foldit: hp://fold.it

77

Digizing the World