Introduction to Bioinformatics 10' Machine Learning for Protein Structure Prediction

1
Introduction to Bioinformatics10. Machine
Learning for Protein Structure Prediction 2

• Course 341
• Department of Computing
• Imperial College, London

2
Coursework

• Link from web page now works
• Use gap penalty of -1
• An alignment must have
• Both sequences the same length after adding gaps

3
Proteins

• Macromolecules called heteropolymers
• Complex conformation of amino acids (residues)
• Below around 40 residues theyre called peptides
• Conformation Fold 3D Structure
• Dictates the function of the protein
• Which dictates how cells behave, etc.
• 40-50 residues tends to be the lower limit
• On proteins having a biological function
• See LESK textbook

4
Why Study Proteins?

• Major biological macromolecule responsible for
  the machinery and scaffold of cellular life
• Prof. Michael Sternberg
• 3D Structure provides insight into
• Function
• Evolution
• Experimental design
• Systematic design of drugs

5
Amino Acids

• Amino acids form the basic unit of all proteins
• A single amino acid always has
• An amino group NH2
• A carboxyl group COOH
• A hydrogen H (sometimes left off diagrams)
• A chemical group or side chain -"R".
• These are all joined to a central carbon atom
• The alpha carbon

6
Amino Acid Picture
7
The Primary Structure of Proteins

• Fixed main chain backbone
• Variable side chain
• One of 20 different amino acid side chains
• Chirality (left and right handedness)

8
Hydrophobic Interactions

• Atomic charges dictate how folds occur
• Groups of C-H atoms have little charge
• Called hydrophobic or non-polar
• Hydrophobic groups pack together
• To avoid contact with solvent (aqueous solution)
• To minimise energy
• Hydrophobic and hydrophilic regions are
• Main driving force behind the folding process

9
Protein Folding in anAqueous Environment
H20
H20
H20
H20
H20
H20
H20
H20
H20
Folded chain Many atoms shielded from water
contact
Unfolded chain
10
Secondary Structures

• Energetics of the side chain
• Stop certain fold conformations
• Protein structures are not random
• Some larger (secondary) structures are observable
• Two common structures
• Alpha-helix
• Beta-sheet
• Connected by turns and loops
• E.g., Beta-turns

11
Alpha Helices
12
Beta Sheets (front and side)
13
Protein Folds

• Tertiary structure
• 3D structure of an entire amino acid chain
• Sequential arrangements of chain sections
• Particularly alpha-helices and beta-sheets/strands
  
• Quaternary structure
• Proteins with more than one chain
• Arrangement of these chains

14
Example of Quaternary Structure
15
Structural Classes of Folds

• Different classes, including
• Alpha/Alpha
• Mainly packing of alpha helices
• Beta/Beta
• Mainly one or more beta sheets
• Alpha/Beta
• Roughly alternate alpha helices and beta sheets
• AlphaBeta
• Mixed alpha helices and beta
• Coil
• Mainly small proteins (fewer than 50 residues)

16
Examples of Common Classes
17
Example Proteins (OspC protein)
18
Triose Phosphate Isomerase

• Tim (Beta/Alpha)8 Barrel

19
Hemerythrin

• 4-Alpha-up-down bundle

20
Homologous Genes

• Similar sequences produce very similar structures
• Secondary structures are well preserved

21
Back to Machine Learning

• Remember the task
• Automatically learn methods for predicting the
  structure of proteins

Structure Predictor
Learning process for the chosen representation
Machine readable format
22
Predicting Secondary Structurefor Each Residue

• G A G D G A N A A A

Structure Predictor
a a a a coil coil ß ß ß ß
Alpha Helix
Beta Sheet
Further Processing
23
The CASP Competition

• Determining actual structure of a protein is hard
• Its a good idea to predict the structure
• Many, varied approaches to structure prediction
• Which one is the best?
• Comparative Assessment of Protein Structure
  Prediction
• Blind trial to evaluate the different approaches
• Chemists do not publish the structures theyve
  found
• CASP - manual intervention is allowed
• CAFASP2 - completely automatic prediction only
• Around 60 targets are used for the evaluation

24
Continuous Server Evaluation

• Every time a new structure is determined
• Every method on a server is assessed
• Regular updates on status are provided
• Two well known services
• Livebench (Poland)
• Eva (Columbia, USA)
• Roughly 100 structures per annum are tested

25
Current State of the Art

• Secondary structure prediction
• Usually in the form of 3-state prediction (Q3)
• Alpha, Beta, Coil
• Majority class predictor gets about 40
• Prediction has gone from 60
• Using rule-based methods
• To about 80 using machine learning methods
• Neural networks, SVMs are particularly good
• Inductive Logic Programming also very useful
• Learns explainable rules which add to the science