Identification of common protein motifs through the application of machine learning

Scott Hollingsworth (Department of Biochemistry & Biophysics, Oregon State University)

Mentor: Dr. P. Andrew Karplus (Department Of Biochemistry & Biophysics, OSU)In Collaboration With: Dr. Weng-Keen Wong (Department Of Computer Science, OSU)

Dr. Donald Berkholz (Department of Biochemistry and Molecular Biology, Mayo Clinic)Dr. Dale Tronrud (Department of Biochemistry & Biophysics, OSU)

Each protein has an individual structure

Structure flows from function

Understand structure, understand function

Ptr Tox A

Phi & Psi (φ, ψ) Phi and psi describe the

conformation of the planar peptide (amino acid) in regards to other peptides

One amino acid – two angles

Ramachandran PlotVoet, Voet & Pratt Biochemistry

(Upcoming 4th Edition)

φ

ψ

Use of Protein Geometry Database (PGD) to identify linear group existence (i.e. α-helix, β-sheet, π-helix…) Simple repeating structures Methods: manual searches Hollingsworth et al. 2009. “On the

occurrence of linear groups in proteins.” Protein Sci. 18:1321-25

α-Helix

310 Helix

Linear groups are only part of the picture Not all common protein motifs are repeating structures Many have changing conformations

Goal of this research: Identify all common motifs in proteins

Too complex for manual searches Enter machine learning

Form of artificial intelligence

Can identify clusters within a dataset Cluster – significant grouping of data points

Visual example…

Topographical map of OregonData value: Elevation

Highest points (Individual peaks)

Mt. Hood(11,239 Feet)

Mt. Jefferson(10,497 Feet)

Three Sisters(10,358-10,047 Feet)

Topographical map of OregonData value: Elevation

Highest points (Individual peaks)

Topographical map of OregonData value: Elevation

Mountain ranges (Broad patterns)

C A

S C

A D

E S

C O

A S

T

R A

N G

E

S I S K I Y O U S( K A L A M A T H )

B L U E M T S

W A L L O W A S

S T

E E

N S

S T R A W B E R R I E S

O C H O C O

M A H O G A N Y M T S

J A C K A S S M T S

H A R T M T N

T U A L A T I N H I L L S

T R O U T C R E E KM T S

P A U L I N AM T S

Similar approach with our data2-Dimensional Example

φ

ψ

Similar approach with our data2-Dimensional Example

α-helix

β

PII

αL

φψ

Complications…

Our Data: 4-dimensional dataset 4D to 2D distance conversions

What has and hasn’t been observed? No definitive source Abundance / Peak Heights

Machine learning programs can identify both previously documented and unknown common motifs and their abundances

1) Create and prep datasets with resolution of at least 1.2Å or higher, 1.75Å or higher

2) Run cuevas

3) Analyze identified clusters Automated process using Python

to remove bias

4) Analyze context of motifs

2D-visual example of cuevas clustering

Goal: Definitive list of the most common protein motifs In order of abundance

“Everest” Method Locate “highest” peak

first▪ Bad pun : “Mt. Alpha-rest”

Locate second highest peak

Locate third…….

Identifying motifs Search for peaks while

looking for ranges

Results: Definitive list of common

protein motifs in order of abundance

The list…

Points Per ResidueCircle r=10 Degree2 φi ψi φi+1 ψi+1 i i+1 Cluster Size Motif Name New Motif

5644 18.07 -63.4 -42 -64 -40.6 α α 1 α-helix / 310-helix247 0.7909 -125.5 132.4 -118 130.2 β β 1 β-strand173 0.5540 -69.9 157.4 -61 -36.3 PII α 1 PII- Helix N-Cap / Capping Box147 0.4707 -65.5 -21.4 -90.3 1.5 α δ 1 Type I Turn#

125 0.4003 -70.4 153.6 -60.4 143 PII PII 1 PII

117 0.3747 -57.2 131 82.4 -0.6 PII δL 1 Type II Turn88 0.2818 -88.3 -2 -64.7 136.9 δ PII 1 Type I Turn Cap55 0.1761 -88.1 1.3 87.9 5.7 δ δL 1 Schellman Motif51 0.1633 -91.8 -1.9 -58.4 -42.5 δ α 1 Reverse Type I Turn X43 0.1377 93.5 -0.1 -71.7 146 δL PII 1 Reverse Type II Turn X40 0.1281 -133.9 164.3 -62.2 -34.1 β α 1 βα Turn36 0.1153 -82.4 -26.8 -146.3 152.1 δ β 2 Classic Beta Bulge‡

35 0.1121 54.9 38.3 84.5 0.8 αL δL 1 Type I` Turn34 0.1089 -122.3 119.6 52.7 41 β αL 1 β → αL X31 0.0993 -136.1 70.4 -65 -19 ζ α 1 ζ → αP

†

31 0.0993 65.3 28.3 -67.2 140.8 αL PII 1 G1 Beta Bulge30 0.0961 82.6 5.6 -103.1 137.5 δL β 1 δL → β X29 0.0929 56.7 -133.5 -73.7 -10.7 PII` δ 3 Type II` Turn24 0.0769 78 0.5 -67.5 -43.1 δL α 1 δL → α X20 0.0640 -78.3 116 -89.1 -31.1 PII δ 1 Type VIa1 Turn (S)20 0.0640 -96.6 0.9 -133.8 156.3 δ β 1 Classic Beta Bulge (S)20 0.0640 50.5 49.9 -61.2 148.3 αL PII 1 Wide Beta Bulge (S)19 0.0608 -69.9 -32.3 -129.8 73.1 α ζ 2 α → ζ†

17 0.0544 -129.1 80.8 -70.3 141.9 ζ PII 1 ζ → PII X15 0.0480 53.7 48 -118.9 126.6 αL β 1 αL → β (S) X14 0.0448 -87.6 61 -140.3 149.5 γ` β 1 γ` Turn11 0.0352 76.3 -169.3 -61.4 138.3 PII` PII 2 PII` → PII X10 0.0320 78.8 171.1 -69.3 -29.6 PII` α 1 PII` → α (S) X

9 0.0288 -138.5 165.7 57.7 -137.8 β PII` 2 β → PII` X9 0.0288 92.8 165.9 -62.5 -35.7 ε α 1 ε → α X8 0.0256 -107.6 16.8 80 -177 δ PII` 1 Reverse Type II` Turn X8 0.0256 84.6 8.1 -143 169.3 δL β 1 δL → β X7 0.0224 -85.8 71.8 -83.1 163.5 γ` PII 3 γ` → PII X6 0.0192 -102.4 -9 92.6 163.3 δ ε 4 δ → ε X6 0.0192 -77.9 -8.6 86.7 174.2 δ ε 1 δ → ε (S) X6 0.0192 83.8 -166.3 -121.9 132.1 PII` β 1 PII` → β X6 0.0192 57.1 44.5 -152.5 158.8 αL β 1 αL → β X6 0.0192 -128.3 98.7 56.7 -133.3 ζ PII` 1 ζ → PII` X

Motif “shapes” Each motif analyzed by

plotting of each motif range

Understand the shape of the cluster/motif

Results: New insight into each

motif’s structure Context Comparisons

Type II Vs. Type II`

Hairpin turns 180° Turn Two Residues

Defined as mirror images of each other

Distributions show differences between the two structures

Nearly four years in the making…

φ

ψ

The results go on… Motif analysis

▪ Viral forming of “Pangea”

Range and peak method sections▪ Adapting cuevas for our data▪ Python automation

▪ Identification of 310 Helix & Type I Turn

6D, 8D, 10D and 12D clustering▪ Full helix caps, loops, halfturns…

For full story, a manuscript for publication is being prepared: Hollingsworth et al. “The protein parts list: motif identification

through the application of machine learning.”(Unpublished)

Cuevas was successful in identifying both documented and undocumented motifs Previously described: Linear groups, helix caps, β-turns (&

reverses), β-bulges, α-turns, loops, helix bends, π-structures… Numerous new motifs Successful from 4D through 20D

Results form the “Protein Parts” List Comprehensive list of all common protein motifs found in proteins

• Dr. P. Andrew Karplus• Dr. Weng-Keen Wong• Dr. Donald Berkholz• Dr. Dale Tronrud• Dr. Kevin Ahern• Howard Hughes Medical

Institute